CN119672961B - Traffic data prediction method and device - Google Patents

Abstract

The application discloses a traffic data prediction method, which comprises the steps of utilizing a trained prediction model for traffic data prediction, acquiring predicted traffic data of a traffic position to be predicted at a future time based on historical traffic data of the traffic position to be predicted at a historical time, wherein the prediction model comprises a time sequence sub-model for acquiring time sequence characteristics of the historical traffic data to obtain a first prediction result, a static space sub-model for acquiring static space characteristics of the historical traffic data to obtain a second prediction result, a multi-model layer formed by at least two dynamic space sub-models for acquiring dynamic space characteristics of the historical traffic data to obtain a third prediction result, and a routing layer for selecting the prediction result output by each sub-model in the multi-model layer. The application realizes the prediction combination of at least two of time sequence characteristics, static space characteristics and dynamic space characteristics, and improves the adaptability of traffic data prediction in various application scenes.

Description

Traffic data prediction method and device

Technical Field

The present invention relates to the field of intelligent transportation, and in particular, to a traffic data prediction method.

Background Art

Accurate traffic data prediction in the intelligent transportation system is a prerequisite for accurate analysis of traffic problems. Traffic data includes but is not limited to traffic flow, queue length, traffic speed and other traffic network indicators at the set traffic location, for example, the flow, queue length, and traffic speed at a certain intersection. Traffic data prediction involves a variety of different scenarios, including but not limited to highways and urban roads. In addition to normal road connections, there are also some spatially isolated roads in the road network, such as some roads that are widely distributed on the edge of the city and are not connected to other roads; there are also roads with complex spatiotemporal characteristics, such as highway ramps in cities, which have special road structures and are easily affected by specific events such as holidays.

Related traffic data prediction methods usually use a single modeling method to better adapt to traffic data prediction in a certain scenario, but it is difficult to generalize to other scenarios and achieve better results.

Summary of the invention

The present invention provides a traffic data prediction method to achieve the generalization capability of traffic data prediction in different scenarios.

The first aspect of the present application provides a traffic data prediction method, the method comprising:

Using the trained prediction model for traffic data prediction, based on the historical traffic data of the traffic location to be predicted at the historical time, the predicted traffic data of the traffic location to be predicted at the future time is obtained,

in,

The prediction model includes: a multi-model layer consisting of at least two of a temporal sub-model for obtaining temporal characteristics of historical traffic data to obtain a first prediction result, a static spatial sub-model for obtaining static spatial characteristics of historical traffic data to obtain a second prediction result, and a dynamic spatial sub-model for obtaining dynamic spatial characteristics of historical traffic data to obtain a third prediction result, and a routing layer for selecting the prediction results output by each sub-model in the multi-model layer.

As a possible implementation, the prediction model is trained in the following manner:

The historical traffic data is divided into time series segments of set sample time length, and the time series segments are used as sample data and input into the prediction model to be trained.

The prediction loss function value of the prediction model is calculated according to the number of sub-models, the classification prediction loss function value of each sub-model, and the sample output probability of each sub-model, wherein the classification prediction loss function value includes: a first type of prediction loss function value for abandoning a poor sub-model, and a second type of prediction loss function value for selecting a better sub-model.

Back propagation is performed based on the predicted loss function value of the prediction model,

Repeat the training until it is completed.

As a possible implementation, the prediction loss function value of the prediction model is calculated as follows:

For any sub-model, calculate the product of the classification prediction loss function value of the sub-model and the logarithm of the sample output probability of the sub-model to obtain the prediction loss function value of the sub-model.

Calculate the average of the prediction loss function values of all sub-models to obtain the prediction loss function value of the prediction model;

The first type of prediction loss function value is determined as follows:

For any sub-model,

When the difference between the predicted sample value output by the sub-model and the true value is less than the set first difference threshold, and the sample output probability of the sub-model is greater than the set sample output probability threshold, the first type of prediction loss function value of the sub-model is set to 1.

When the difference between the predicted sample value output by the sub-model and the true value is greater than the first difference threshold, and the sample output probability of the sub-model is not greater than the sample output probability threshold, the first type of prediction loss function value of the sub-model is set to the reciprocal of the number of sub-models minus 1.

Otherwise, the first-class prediction loss function value of this sub-model is set to 0;

The second type of prediction loss function value is determined as follows:

For any sub-model,

When the difference between the predicted sample value output by the sub-model and the true value is less than the second difference threshold, and the sample output probability of the sub-model is greater than the sample output probability threshold, the second type of prediction loss function value of the sub-model is set to 1.

When the difference between the predicted sample value output by the sub-model and the true value is greater than the second difference threshold, and the sample output probability of the sub-model is not greater than the sample output probability threshold, the second type prediction loss function value of the sub-model is set to the reciprocal of the number of sub-models minus 1.

Otherwise, the second-category prediction loss function value of this sub-model is set to 0.

in,

The second difference threshold is the difference between 1 and the first difference threshold. The difference threshold is the quantile threshold used to characterize the equal-division numerical points within the probability distribution range.

The output probability threshold is determined according to the number of types of prediction results of the required output sub-model.

As a possible implementation, the method of using the trained prediction model for traffic data prediction to obtain predicted traffic data for the traffic location to be predicted at a future time based on historical traffic data of the traffic location to be predicted at a historical time includes:

According to the set historical time step, the historical traffic data at the traffic location to be predicted is extracted to obtain the historical traffic data at each historical time point of the set historical time step.

In chronological order, the historical traffic data at each historical time point is sequentially input into the trained prediction model, so that: each sub-model in the multi-model layer of the prediction model obtains the prediction results of each sub-model from the currently input historical sequence data, and the routing layer selects the prediction results of each sub-model whose output probability is greater than the set output probability threshold according to the output probability of each sub-model.

in,

The output probability of the sub-model is used to represent: the proportion of the similarity between the prediction results output by the sub-model and the pattern features of the current historical sequence data in the sum of the similarities between the prediction results output by each sub-model and the pattern features of the current historical sequence data,

Pattern features are used to characterize the correlation between the current historical sequence data and the current time series pattern.

The current time series pattern is used to characterize each traffic data pattern with a high probability of recurrence that is searched through the time series.

As a possible implementation, the prediction result is prediction sequence data composed of predicted traffic data at each future time point, wherein the prediction time step is set between two adjacent future time points;

As a possible implementation, the output probability is determined in the following manner:

For any sub-model,

According to the linear weight and linear offset, the multi-model layer prediction sequence data used to characterize the prediction results of the multi-model layer for the current historical sequence data is calculated.

According to the multi-model layer prediction sequence data and the current time series mode, the correlation between the multi-model layer prediction sequence data and each traffic data mode in the current time series mode is calculated, and each traffic data mode in the current time series mode is weighted by the calculated correlation to obtain the pattern characteristics of the current historical sequence data.

According to the pattern features of the current historical sequence data and the prediction results output by the sub-model, the similarity between the prediction results output by the sub-model and the pattern features of the current historical sequence data is calculated.

According to the pattern characteristics of the current historical sequence data and the prediction results output by each sub-model, the similarity between the prediction results output by each sub-model and the pattern characteristics of the current historical sequence data is calculated to obtain the similarity of each sub-model, and the similarity of each sub-model is accumulated to obtain the sum of each similarity.

The ratio of the similarity between the prediction result output by the sub-model and the pattern features of the current historical sequence data to the sum of the similarities is calculated to obtain the output probability of the sub-model.

As a possible implementation, the calculation of the multi-model layer prediction sequence data for characterizing the prediction result of the multi-model layer for the current historical sequence data according to the linear weight and the linear offset includes:

The product of the linear weight and the current historical sequence data is added with the linear offset to obtain the multi-model layer prediction sequence data of the current historical sequence data;

As a possible implementation, the calculation of the correlation between the multi-model layer prediction sequence data and each traffic data pattern in the current time series pattern according to the multi-model layer prediction sequence data and the current time series pattern includes:

For any traffic data mode in the current time series mode,

Calculate an exponential function value with a natural constant as the base and a product of the multi-model layer prediction sequence data and the transpose of the row vector used to characterize the traffic data pattern as the exponent to obtain the exponential function value of the traffic data pattern, wherein the multi-model layer prediction sequence data is a row vector, the transpose of the row vector of the traffic data pattern is a column vector, the current time series pattern is a time series pattern matrix, each row in the time series pattern matrix corresponds to a traffic data pattern, and each column corresponds to a time series,

Calculate the exponential function value with the natural constant as the base and the product result of the transposition of the row vector of each traffic data mode and the multi-model layer prediction sequence data as the exponent, obtain the exponential function value of each traffic data mode, and accumulate the exponential function values of each traffic data mode,

The ratio of the exponential function value of the traffic data pattern to the accumulated exponential function values of each traffic data pattern is calculated to obtain the correlation between the multi-model layer prediction sequence data and the traffic data pattern.

As a possible implementation, weighting each traffic data mode in the current time series mode by using each calculated correlation degree includes:

For each traffic data pattern relevance, the product of the traffic data pattern relevance and the row vector of the traffic data pattern is calculated to obtain a weighted row vector of the traffic data pattern.

Add the weighted row vectors of each traffic data pattern to obtain a pattern feature vector;

As a possible implementation manner, the selecting the prediction results of each sub-model whose output probability is greater than a set output probability threshold further includes:

According to the output probability of each selected sub-model, the fusion weight is determined.

The selected prediction results are weighted and summed using the fusion weights to obtain the predicted traffic data.

As a possible implementation, the time series sub-model is a time series large model, which is used to model the traffic time series characteristics of traffic data at any traffic location at the next time point, which are only related to the traffic time series characteristics of the traffic data at the traffic location at the previous time point.

As a possible implementation, the static space sub-model is a graph neural network model, which is used to model the static spatial characteristics of traffic data at any traffic location at the next time point, which are related to the static spatial characteristics of traffic data at the traffic location at the previous time point, and the static spatial characteristics of traffic data at a static traffic location statically adjacent to the traffic location.

As a possible implementation, the dynamic spatial sub-model is a graph neural network model and an attention mechanism model, which is used to model that the spatial characteristics of traffic data at any traffic location at the next time point are related to the spatial characteristics of all traffic locations at the previous time point, and the spatial feature correlation is obtained through the attention mechanism.

A second aspect of the present application provides a prediction model structure for traffic data prediction, the prediction model structure comprising:

a multi-model layer consisting of at least two of a temporal sub-model for obtaining temporal features of historical traffic data to obtain a first prediction result, a static spatial sub-model for obtaining static spatial features of historical traffic data to obtain a second prediction result, and a dynamic spatial sub-model for obtaining dynamic spatial features of historical traffic data to obtain a third prediction result, and

A routing layer used to select the prediction results output by each sub-model in a multi-model layer.

A third aspect of the present application provides a traffic data prediction device, the device comprising:

The prediction module uses the trained prediction model for traffic data prediction to obtain the predicted traffic data of the traffic location to be predicted at a future time based on the historical traffic data of the traffic location to be predicted at a historical time.

in,

The prediction model includes: a multi-model layer consisting of at least two of a temporal sub-model for obtaining temporal characteristics of historical traffic data to obtain a first prediction result, a static spatial sub-model for obtaining static spatial characteristics of historical traffic data to obtain a second prediction result, and a dynamic spatial sub-model for obtaining dynamic spatial characteristics of historical traffic data to obtain a third prediction result, and a routing layer for selecting the prediction results output by each sub-model in the multi-model layer.

A traffic data prediction method provided by the present application realizes the fusion of prediction results of multiple sub-models on historical traffic data through multiple categories of sub-models and routing layers in the multi-model layer of the prediction model, avoids dependence on the prediction results of a single model, realizes the prediction combination of at least two of the time series characteristics, static spatial characteristics, and dynamic spatial characteristics, and improves the adaptability of traffic data prediction in various application scenarios; further, the time series large model used by the time series sub-model in the multi-category sub-model is conducive to combining the time series large model with the spatial small model, thereby enhancing the overall time series generalization ability of the prediction model.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a traffic data prediction method according to an embodiment of the present application.

FIG2 is a flow chart of a traffic data prediction method according to an embodiment of the present application.

FIG3 is a schematic diagram of modeling of a sequential sub-model in this embodiment.

FIG4 is a schematic diagram of modeling a static space sub-model in this embodiment.

FIG5 is a schematic diagram of modeling a static space sub-model in this embodiment.

FIG. 6 is a schematic diagram of a gated network layer obtaining control data according to this embodiment.

FIG. 7 is a schematic diagram of a prediction model trained in this embodiment performing prediction based on input historical traffic data.

FIG8 is a schematic diagram of a traffic data prediction device according to an embodiment of the present application.

FIG. 9 is another schematic diagram of the traffic data prediction device according to an embodiment of the present application.

DETAILED DESCRIPTION

In order to make the objectives, technical means and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings.

The embodiment of the present application utilizes a trained prediction model having a multi-model layer composed of multiple sub-models to obtain predicted traffic data based on historical traffic data.

Referring to FIG. 1 , FIG. 1 is a flow chart of a traffic data prediction method according to an embodiment of the present application. The method comprises:

Using the trained prediction model for traffic data prediction, based on the historical traffic data of the traffic location to be predicted at the historical time, the predicted traffic data of the traffic location to be predicted at the future time is obtained,

in,

The prediction model includes: a multi-model layer consisting of at least two of a temporal sub-model for obtaining temporal characteristics of historical traffic data to obtain a first prediction result, a static spatial sub-model for obtaining static spatial characteristics of historical traffic data to obtain a second prediction result, and a dynamic spatial sub-model for obtaining dynamic spatial characteristics of historical traffic data to obtain a third prediction result, and a routing layer for selecting the prediction results output by each sub-model in the multi-model layer.

As an example, the first prediction result represents the predicted traffic data for the traffic location to be predicted at various future time points, the second prediction result represents the traffic data for the traffic location to be predicted at various future time points predicted based on the traffic data at various fixed traffic locations related to the traffic location to be predicted in the road network and the historical traffic data at the traffic location to be predicted, and the third prediction result represents the predicted traffic data for the traffic location to be predicted at various future time points predicted based on the historical traffic data at all traffic locations in the road network.

As an example, according to the set historical time step, the historical traffic data at the traffic location to be predicted is extracted to obtain the historical traffic data at each historical time point of the set historical time step.

In chronological order, the historical traffic data at each historical time point is sequentially input into the trained prediction model, so that: each sub-model in the multi-model layer of the prediction model obtains the prediction results of each sub-model from the currently input historical sequence data, and the routing layer selects the prediction results of each sub-model whose output probability is greater than the set output probability threshold according to the output probability of each sub-model.

Among them, the prediction result is the predicted sequence traffic data with a set prediction time length. The total number of future time points included in the predicted sequence traffic data may be the same as or different from the total number of historical time points included in the currently input historical sequence data. The time between two adjacent future time points is the set prediction time step.

For example, in ascending order of time, the historical traffic data at each historical time point is input into the trained prediction model in sequence. In this way, the first future time point in the prediction sequence data is the next time point adjacent to the last historical time point in the currently input historical sequence data, and the difference between the first future time point and the last historical time point is the set prediction time step.

The historical time step can be determined according to the time interval for collecting historical traffic data. It should be understood that the historical time step can be a fixed value or a dynamically changing value, and this application does not limit this. The prediction time step can be the same as the historical time step or different, and this application does not limit this.

The output probability of the sub-model is used to represent: the proportion of the similarity between the prediction results output by the sub-model and the pattern features of the current historical sequence data in the sum of the similarities between the prediction results output by each sub-model and the pattern features of the current historical sequence data,

Pattern features are used to characterize the correlation between the current historical sequence data and the current time series pattern.

The current time series mode is used to characterize each traffic data mode with a high probability of recurrence obtained through the time series.

It should be understood that the traffic location to be predicted can be one traffic location to be predicted, for example, any intersection in the road network, or it can be multiple traffic locations to be predicted, for example, multiple intersections at a certain distance in the road network. The multiple traffic locations to be predicted can have different forms, including but not limited to intersections, T-junctions, one-way intersections, tidal intersections, multi-directional intersections, designated locations, etc., and the present application does not impose any restrictions on this.

For multiple traffic locations to be predicted, the historical traffic data at each of the traffic locations to be predicted can be input into the prediction model in parallel, or can be input into the prediction model in series, or can be input into the prediction model alternately in parallel and in series, and this application does not limit this. As a possible implementation method, when the historical traffic data at each of the traffic locations to be predicted are input in parallel in the same time sequence, the input historical traffic data of each route can have the same set time length, so as to improve the accuracy of the prediction and reduce the complexity of the prediction process.

The traffic data prediction method provided in the embodiment of the present application, through a multi-model layer composed of various types of sub-models with independent prediction capabilities, enables the prediction model to have multiple expert model functions, and through the routing layer, the prediction results of each sub-model can be selected and integrated, which is equivalent to realizing a hybrid expert model, which is beneficial to improving the generalization ability of the prediction model.

To facilitate understanding of the embodiments of the present application, the following description uses a prediction model using a multi-model layer consisting of a timing sub-model, a static space sub-model, and a dynamic space sub-model as an example. It should be understood that the embodiments of the present application are not limited to this, and any two of the sub-models may also be applicable.

See Figure 2, which is a schematic diagram of the prediction model of this embodiment. The prediction model includes: a first fully connected layer, a multi-model layer, a second fully connected layer, and a routing layer connected in sequence, wherein the routing layer includes a gating network layer and a routing selection layer, the input end of the gating network layer is respectively connected to the output end of the multi-model layer and the input end of the first fully connected layer, and the output end of the gating network layer is connected to the selection input end of the routing selection layer.

The first fully connected layer is used to combine and transform the traffic characteristic data of each historical time point at each traffic location.

The multi-model layer is used for performing spatiotemporal prediction based on the traffic feature data output from the first fully connected layer. The multi-model layer includes a time series sub-model, a static space sub-model, and a dynamic space sub-model, wherein:

The time series sub-model is used to realize the traffic data prediction of each traffic location at each future time.

The static space sub-model is used to realize traffic data prediction at static traffic locations.

The dynamic spatial sub-model is used to realize traffic data prediction at dynamic traffic locations;

Static traffic positions refer to static traffic positions with fixed attributes, such as road intersections and intersections, etc. Dynamic traffic positions refer to dynamic traffic positions with non-fixed attributes, such as road intersections and intersections with temporary control or restrictions.

The above-mentioned multi-model layer includes multiple types of sub-models. Each type of sub-model acts as an expert model, responsible for processing a specific part of the input data or a subset of the tasks. The performance and efficiency of the model are improved through division of labor and cooperation among expert models, so as to be suitable for processing complex and diverse traffic data.

In view of the fact that there are many related works on large time series models, but a single large time series model ignores spatial information, and in complex large-scale transportation networks, spatial information plays a very important role in tasks such as traffic time series prediction and completion; at the same time, there are also various spatial modeling methods, such as those based on graph neural networks, attention mechanisms, etc., but the time series part used in these spatial modeling methods is often a small time series model when modeling. Since the small time series model is relatively poor in time series modeling ability and generalization ability compared with the large time series model, that is to say, the large time series model and the small space model actually have their own advantages and disadvantages. In this embodiment, a variety of spatial modeling methods are combined with the time series modeling ability of the large time series model, and the multi-model includes a time series sub-model, a static space sub-model, and a dynamic space sub-model.

As an example, the time series sub-model adopts the time series big model, which is a deep learning model for processing and analyzing time series data to capture long-term dependencies and dynamic changes in time series, which is conducive to predicting future trends, anomaly detection, etc. The time series sub-model is used to extract the time series characteristics of traffic data, without considering the relationship between each traffic location, to model the traffic time series characteristics of traffic data at any traffic location at the next time point is only related to the traffic time series characteristics of the traffic data at the previous time point at the traffic location, that is, based on the historical time series characteristics of the traffic location at the previous time point to predict the future time series characteristics of the traffic location at the next time point. It can be expressed as:

Wherein, _Hi ^(t) represents the traffic time sequence mapping of the traffic data at the traffic position i at the next time t, _Hi ^(t-1) represents the traffic time sequence mapping of the traffic data at the traffic position i at the previous time t-1. The traffic time sequence mapping is used to characterize the time sequence characteristic information of the traffic data. The traffic data includes but is not limited to the traffic flow information, speed information, the length of the queue to be passed and other road network indicator information.

See Figure 3, which is a schematic diagram of the modeling of the time series sub-model in this embodiment. In the figure, circles of different colors represent different traffic locations, and different planes represent different time points.

The static space submodel is used to extract the static spatial features of traffic data at each traffic location. As an example, the static space submodel uses the GCN model to extract the static correlation between each traffic location, that is, to model the static spatial features of traffic data at any traffic location at the next time point, which is related to the static spatial features of traffic data at the traffic location at the previous time point, and the static spatial features of traffic data at static traffic locations that are statically adjacent to the traffic location. In other words, based on the historical static spatial features of the traffic location at the previous time point and the static spatial features of each static traffic location, the future static spatial features of the location at the next time point are predicted. It can be expressed as:

Among them, _Si ^(t) represents the static spatial characteristics of traffic data at traffic location i at the next time t,

S _i ^(t-1) represents the static spatial characteristics of the traffic data at the traffic position i at the previous time t-1, g() represents the static spatial feature mapping, and θ _j is the static spatial characteristics of the traffic data at the static traffic position j that is statically adjacent to the traffic position i. For example, the four sections at the intersection are fixedly distributed in space, and the entrance and exit of the highway are fixed in space.

See Figure 4, which is a schematic diagram of modeling the static space sub-model of this embodiment. In the figure, circles of different colors represent different traffic positions, different planes represent different time points, black line segments represent the spatial position relationship of the positions represented by the circles at both ends of the line segment, and red line segments represent the relationship between the spatial positions of the positions represented by the circles at both ends of the line segment at different time points.

The dynamic spatial submodel is used to extract the dynamic spatial features of traffic data at each traffic location. As an example, the dynamic spatial submodel is an attention mechanism model to extract the dynamic correlation between traffic locations, that is, the spatial features of traffic data at any traffic location in the modeled road network at the next time point are related to the spatial features of all traffic locations in the road network at the previous time point. The correlation is obtained through the attention mechanism, that is, the future dynamic spatial features of traffic at the next time point at the traffic location are predicted based on the historical dynamic spatial features of traffic at all traffic locations at the previous time point. It can be expressed mathematically as:

Where D _i ^(t) represents the traffic dynamic spatial characteristics of the traffic data at the traffic location i at the next time t,

_DJ ^(t-1) represents the traffic dynamic spatial characteristics of the traffic data at all traffic locations J at the previous time t-1, and Attention() represents the attention mechanism operation.

See Figure 5, which is a schematic diagram of modeling the static space sub-model of this embodiment. In the figure, circles of different colors represent different traffic positions, different planes represent different time points, black line segments represent the spatial position relationship of the positions represented by the circles at both ends of the line segment, and red line segments represent the relationship between the spatial positions of the positions represented by the circles at both ends of the line segment at different time points.

In the above sub-model, the time interval between time t and time t-1 may be a set prediction time step.

The second fully connected layer is used to combine and transform the prediction results from the sub-models.

The routing layer is used to select the prediction results of each sub-model to output the predicted traffic data. As an example, the gated network layer is used to calculate the probabilities of different sub-models, and the routing selection layer selects each sub-model according to the probability vector and fuses the outputs of each sub-model to achieve a mixed output of multiple models.

The routing layer includes a gating network layer and a routing selection layer. The prediction results output by each sub-model in the multi-model layer are respectively input into the gating network layer, the control data output by the gating network layer is input into the routing selection layer, and the traffic feature data of each sub-model output by the second fully connected layer is input into the routing selection layer as the selected data.

See FIG6 , which is a schematic diagram of the gated network layer obtaining control data in this embodiment. By learning the current time series pattern used to characterize each traffic data pattern with a high probability of recurrence searched through the time series, a time series pattern matrix used to characterize the traffic data pattern is obtained, each row of the matrix represents a traffic data pattern, a row vector is used to characterize the pattern feature vector of a traffic data pattern, and each column characterizes a time series information, that is, at each time t, any element in the matrix represents the traffic data pattern of the row where the element is located and the characteristic value of the time series of the column where the element is located. The traffic data pattern can be a characteristic behavior pattern, for example, traffic peak, flat peak, low peak, congestion and other patterns.

The current input sequence traffic data input to the prediction model is input to a linear weight network to obtain the predicted sequence traffic data of the multi-model layer, which can be expressed mathematically as follows:

Among them, _Xi ^(t) is the current input sequence traffic data at traffic location i at time t, that is, the input data input to the prediction model, _Wq is the weight of the linear weight network, _bq is the offset of the linear weight network, and _Qi ^(t) is the multi-model layer prediction sequence traffic data at traffic location i at time t output by the linear weight network.

According to the multi-model layer prediction sequence traffic data and the row vector of each traffic data pattern in the time series pattern matrix, the correlation between the multi-model layer prediction sequence traffic data and each traffic data pattern is obtained. Specifically, for any traffic data pattern in the current time series pattern:

An exponential function value with a natural constant as the base and a product of the multi-model layer prediction sequence data and the transpose of the row vector used to characterize the traffic data pattern as the exponent is calculated to obtain the exponential function value of the traffic data pattern, wherein the multi-model layer prediction sequence data is a row vector, and the transpose of the row vector of the traffic data pattern is a column vector,

Calculate the exponential function value with the natural constant as the base and the product result of the transposition of the row vector of each traffic data mode and the multi-model layer prediction sequence data as the exponent, obtain the exponential function value of each traffic data mode, and accumulate the exponential function values of each traffic data mode,

The ratio of the exponential function value of the traffic data pattern to the accumulated exponential function values of each traffic data pattern is calculated to obtain the correlation between the multi-model layer prediction sequence data and the traffic data pattern.

Mathematically expressed as:

Among them, _αm is the correlation between the multi-model layer predicted sequence traffic data and the traffic data pattern m, M[m] ^T is the transpose of the row vector M[m] corresponding to the traffic data pattern m in the time series pattern matrix, which is a column vector, _Qi ^(t) is a row vector, and M is the total number of traffic data patterns, that is, the total number of rows in the time series pattern matrix.

Perform weighted summation on each traffic data pattern to obtain the pattern feature used to characterize the traffic data. Specifically, for each traffic data pattern relevance, calculate the product of the traffic data pattern relevance and the row vector of the traffic data pattern to obtain the weighted row vector of the traffic data pattern, and add the weighted row vectors of all traffic data patterns to obtain the pattern feature vector. It can be expressed as:

Wherein, O _i ^(t) is the pattern feature vector of the traffic data at the traffic position i at time t, which is a row vector, and α _m M[m] is the weighted row vector of the traffic data pattern m.

The correlation between the pattern feature vector of the traffic data and the prediction results of each sub-model in the multi-model layer is calculated to obtain the sub-model similarity used to characterize the similarity between the prediction results of each sub-model and the pattern feature vector, which is expressed as follows:

Among them, _re represents the similarity between sub-model e and the pattern feature vector, _ze represents the prediction result of sub-model e, which can be sequence data and can be expressed in vector form, and R( ) represents the similarity function, which can be cosine similarity, Euclidean distance, etc.

Based on the similarity of each sub-model, the output probability of each sub-model is calculated, and the calculated output probability of each sub-model is input into the routing selection layer as the control data output by the gated network layer. Specifically, the similarities of each sub-model are accumulated to obtain the sum of each similarity; for any sub-model, the ratio of the similarity of the sub-model to the sum of each similarity is calculated to obtain the output probability of the sub-model, which is expressed as follows:

Wherein, p _e is the output probability of sub-model e. In this embodiment, the value of e is 3, that is, the total number of sub-models included.

It should be understood that the prediction results of the above-mentioned sub-models can be the traffic data output by the hidden layer of the sub-model, or the traffic data output by the output layer of each sub-model in the multi-model layer, and this application does not impose any restrictions on this.

The routing selection layer selects the prediction results output by the sub-models whose output probability is greater than the set output probability threshold at each time t as the predicted traffic data according to the output probability of each sub-model. For example, the prediction results output by the k sub-models with the highest output probability are selected as the predicted traffic data.

In this embodiment, the prediction model is trained in the following manner: the historical traffic data is divided into time series segments of a set sample time length, and the time series segments are used as sample data and input into the prediction model to be trained. In order to find a suitable route, two types of classification prediction loss function values are used in back propagation: one type is the first type of prediction loss function value used to abandon the poor model, and the other type is the second type of prediction loss function value used to select the better model, so as to select the output results of k sub-models with relatively better prediction results.

As an example, the loss function of the prediction model is calculated as follows: the prediction loss function value of the prediction model is calculated according to the classification prediction loss function value of each sub-model, the number of sub-models, and the sample output probability of each sub-model. Specifically,

For any sub-model, the product of the classification prediction loss function value of the sub-model and the logarithm of the sample output probability of the sub-model is calculated to obtain the prediction loss function value of the sub-model, where the logarithm can be a logarithm with base 2.

Calculate the average of the prediction loss function values of all sub-models to get the prediction loss function value of the prediction model. The mathematical expression is:

Wherein, L(P′) represents the loss function value of the prediction model when the sample data within the set sample time step, i.e., the sample time interval P′, is input, E is the total number of sub-models, p _e ′ is the sample output probability of sub-model e, and l _e is the classification prediction loss function value of sub-model e.

When the first type of prediction loss function value is used, a quantile threshold qth used to characterize the numerical points in the probability distribution range is set to be greater than 0.5. In this way, the accuracy of the sample prediction value of any sub-model and the output probability of the sub-model will have the following four situations.

(1) The difference between the predicted value of the sub-model sample and the true value of the sub-model is less than qth, and the output probability of the sub-model sample is among the top k.

(2) The difference between the predicted value of the sub-model sample and the true value of the sub-model is less than or equal to qth, and the output probability of the sub-model sample is not in the top k.

(3) The difference between the predicted value of the sub-model sample and the true value of the sub-model is greater than or equal to qth, and the output probability of the sub-model sample is among the top k.

(4) The difference between the predicted value of the sub-model sample and the true value of the sub-model is greater than qth, and the probability of the output sample of the sub-model is not in the top k.

Among them, cases (2) and (3) are the prediction loss function values of the poor model, so their prediction loss function values le are set to 0.

Based on this, the first-class prediction loss function value of any sub-model can be determined as follows:

When the difference between the predicted sample value output by the sub-model and the true value is less than the set first difference threshold, and the sample output probability of the sub-model is greater than the set sample output probability threshold, the first type of prediction loss function value of the sub-model is set to 1.

When the difference between the predicted sample value output by the sub-model and the true value is greater than the first difference threshold, and the sample output probability of the sub-model is not greater than the sample output probability threshold, the first type of prediction loss function value of the sub-model is set to the reciprocal of the number of sub-models minus 1.

Otherwise, the first-class prediction loss function value of this sub-model is set to 0;

Mathematically expressed as:

Among them, l _e is the loss function value of sub-model e, Represents the sample prediction value of sub-model e The difference between the true value y and the model, is the first difference threshold, which is taken as the quantile threshold. top k(P′) represents the k highest sample output probabilities when the sample data within the sample time step, i.e., the sample time interval P′, is input into the prediction model. p _e ′ is the sample output probability of sub-model e.

When the second type of prediction loss function value is used, the accuracy of the sample prediction value of any sub-model and the probability of the sub-model will have the following four situations.

(1) The difference between the sample prediction value of the sub-model and the true value of the sub-model is less than 1-qth, and the sample output probability of the sub-model is among the top k.

(2) The difference between the sample prediction value of the sub-model and the true value of the sub-model is less than or equal to 1-qth, and the sample output probability of the sub-model is not in the top k.

(3) The difference between the sample prediction value of the sub-model and the true value of the sub-model is greater than or equal to 1-qth, and the sample output probability of the sub-model is among the top k.

(4) The difference between the sample prediction value of the sub-model and the true value of the sub-model is greater than 1-qth, and the sample output probability of the sub-model is not in the top k.

Among them, cases (1) and (4) are the prediction loss function values of the better models, so their prediction loss function values le are not set to 0.

Based on this, the second type prediction loss function value of any sub-model can be determined in this way: for any sub-model,

When the difference between the predicted sample value output by the sub-model and the true value is less than the second difference threshold, and the sample output probability of the sub-model is greater than the sample output probability threshold, the second type of prediction loss function value of the sub-model is set to 1.

When the difference between the predicted sample value output by the sub-model and the true value is greater than the second difference threshold, and the sample output probability of the sub-model is not greater than the sample output probability threshold, the second type prediction loss function value of the sub-model is set to the reciprocal of the number of sub-models minus 1.

Otherwise, the second type of prediction loss function value of the sub-model is set to 0. The mathematical formula is:

Among them, 1-qth is the second difference threshold.

Refer to FIG. 7 , which is a schematic diagram of the trained prediction model of this embodiment performing prediction based on input historical traffic data.

In this embodiment, the number of traffic locations to be predicted includes N. According to the set historical time step P, the historical traffic data at each traffic location to be predicted is extracted to obtain the historical traffic data at each historical time point of the set historical time step; in the ascending order of time, the historical traffic data at each historical time point of each traffic location to be predicted is sequentially input into the trained prediction model to input the historical sequence traffic data of each traffic location to be predicted. For example, the current input historical traffic data _XP in FIG7 includes the historical traffic data of N traffic locations to be predicted at each time point tP-1, tP, ...t. The traffic data may be one-dimensional, such as flow, or may be multi-dimensional, such as flow and rate.

In the prediction model:

The historical sequence traffic data at each traffic location to be predicted is input into the gating network layer in the routing layer and simultaneously input into the first fully connected layer.

The historical sequence traffic data at each traffic location to be predicted are combined by the first fully connected layer and then input into each sub-model in the multi-model layer at the same time.

Each sub-model outputs the prediction results of each traffic location to be predicted predicted by the sub-model at its prediction time step S, wherein the prediction results predict the sequence traffic data, and the prediction time step S at each traffic location to be predicted can be the same or different, and this embodiment does not impose any restrictions on this.

The prediction results output by each sub-model are respectively input to the routing selection layer in the routing layer through the second fully connected layer, and are respectively input to the gating network layer in the routing layer.

The gated network layer calculates the output probability of each sub-model based on the input prediction results and input historical traffic data, and inputs it to the routing selection layer in the routing layer.

The routing selection layer selects the prediction results output by the sub-model whose output probability is greater than the set output probability threshold from the prediction results output by the sub-models according to the output probability of each sub-model.

Furthermore, the prediction model can also fuse the selected prediction results according to the output probability of the selected sub-model to obtain predicted traffic data. For example, the prediction results in FIG7 include predicted traffic data of N traffic locations to be predicted at each time point t+1, t+S+1, ... It should be understood that the fusion of the selected prediction results can be performed in the routing selection layer or outside the routing layer, and the present application does not limit this.

To facilitate understanding of this embodiment, the following takes the prediction of traffic data of a road network of 200 intersections as an example.

Assuming that the road network traffic collection time interval is 5 minutes, and the prediction task is to predict the entire road network traffic data for the next hour based on the entire road network traffic data for the past hour, the collection time interval can be used as the historical time step and the prediction time step. The trained prediction model can be used to predict the traffic data for the next 12 prediction time steps based on the traffic data for 12 historical time steps.

The prediction model can be trained as follows:

Obtain the historical traffic data of the road network for the past several days, for example, 7 days, and divide the historical traffic data of each intersection into time series segments of 2 hours in a sliding window manner as the training sample data of the prediction model. As shown in Figure 7, the number of historical time steps P of the historical sample data is set to 12, the number of prediction time steps S is set to 12, the number of intersections N=200, and the traffic data dimension C=1.

The input sample data, whose dimension is [12, 200, 1], is first transformed through the first fully connected (FCN) layer. The output data of the first FCN layer is sent to each sub-model in the multi-model to obtain the output z ₁ , z ₂ , and z ₃ respectively.

The routing selection layer is based on the input sample data and the time series pattern matrix M. According to formulas (4), (5), and (6), the pattern feature vector O can be obtained. The similarity between the pattern feature vector O and the output z1 _~3 of each sub-model can be calculated based on formula (7) to obtain _re . According to _re , the sample output probability _p'e of each sub-model can be obtained.

Based on the sample prediction value y output by each sub-model and the sample output probability _p'e of the sub-model, the prediction loss function value can be obtained based on the loss function of the prediction model, and back propagation is performed to optimize the network parameters.

For example: select 2 sub-models, that is, k=2, and the sample output probabilities _p'e of the three sub-models are p ₁ =0.5, p ₂ =0.4, p ₃ =0.1, respectively. The differences between the sample prediction values and the actual values of the three sub-models are:

qth is set to 0.7;

If we abandon the loss function value of the worse model, we have:

Submodel 1 is case (1), le = 1;

Submodel 2 is case (3), le = 0;

Submodel 3 is case (2), le = 0;

Therefore, sub-models 2 and 3 are worse, and the prediction loss function value of the prediction model is as follows:

If a better model loss function value is selected, then:

Submodel 1 is case (1), le = 1;

Submodel 2 is case (3), le = 0;

Submodel 3 is case (4), le = 1/2;

Therefore, sub-models 1 and 3 are better choices, and the prediction loss function value of the prediction model is as follows:

According to the two prediction loss function values of the prediction model calculated by the two types of prediction loss functions adopted, the model parameters of the prediction model are adjusted until the training is completed to obtain the trained prediction model.

Based on the trained prediction model, the input data is processed: time series and space are modeled through three sub-models; at each moment, the routing layer is used to select the prediction results output by each sub-model to obtain the best k models to achieve the prediction of the entire road network traffic.

Specifically, at time t, the traffic flow data of intersection A is predicted, and the predicted traffic flows output by each sub-model are v1, v2, and v3, respectively. The output probabilities of each sub-model are 0.6, 0.37, and 0.03, respectively. Then, the predicted traffic flow at intersection A at time t is:

This embodiment achieves better prediction result selection through a prediction model composed of multiple sub-models, which is conducive to dealing with traffic data prediction in different scenarios. Specifically, prediction is performed through a time series sub-model, which increases the time series generalization ability of the prediction model. The static space sub-model is conducive to obtaining regular traffic characteristics, and the dynamic space sub-model is conducive to obtaining irregular traffic characteristics. The selection of prediction results of each sub-model by the routing layer not only realizes the fusion of the prediction results of each sub-model, but also avoids relying on the prediction results of a single model. The time series large model used by the time series sub-model can capture the long-term dependencies and dynamic changes in the time series, which is conducive to combining the time series large model with the space small model, and enhances the overall time series generalization ability of the prediction model. During the training process, the prediction loss function value of the prediction model is calculated by the classification prediction function value during back propagation, so as to realize learning in a way of selecting the better model and avoiding the worse model.

See FIG8 , which is a schematic diagram of a traffic data prediction device according to an embodiment of the present application. The device includes:

The prediction module is used to obtain predicted traffic data of the traffic location to be predicted at a future time based on the historical traffic data of the traffic location to be predicted at a historical time by using the trained prediction model for traffic data prediction.

in,

The prediction model includes: a multi-model layer consisting of at least two of a temporal sub-model for obtaining temporal characteristics of historical traffic data to obtain a first prediction result, a static spatial sub-model for obtaining static spatial characteristics of historical traffic data to obtain a second prediction result, and a dynamic spatial sub-model for obtaining dynamic spatial characteristics of historical traffic data to obtain a third prediction result, and a routing layer for selecting the prediction results output by each sub-model in the multi-model layer.

Referring to Figure 9, Figure 9 is another schematic diagram of a traffic data prediction device or an electronic device for traffic data prediction according to an embodiment of the present application. The device includes: a memory and a processor, the memory stores a computer program, and the processor is configured to execute the computer program to implement the steps of the traffic data prediction method according to an embodiment of the present application.

The memory may include a random access memory (RAM) or a non-volatile memory (NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

An embodiment of the present invention further provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the steps of the traffic data prediction method described in this embodiment are implemented.

As for the apparatus/network-side device/storage medium embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

In this article, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims (10)

1. A traffic data prediction method, the method comprising:

obtaining predicted traffic data of the traffic location to be predicted at a future time based on historical traffic data of the traffic location to be predicted at a historical time by using the trained prediction model for traffic data prediction,

Wherein,

The prediction model comprises a time sequence sub-model for acquiring time sequence characteristics of historical traffic data to obtain a first prediction result, a static space sub-model for acquiring static space characteristics of the historical traffic data to obtain a second prediction result, a multi-model layer formed by at least two dynamic space sub-models for acquiring dynamic space characteristics of the historical traffic data to obtain a third prediction result, and a routing layer for selecting the prediction result output by each sub-model in the multi-model layer;

The obtaining, by using the trained prediction model for traffic data prediction, predicted traffic data of the traffic location to be predicted at a future time based on historical traffic data of the traffic location to be predicted at a historical time, includes:

each sub-model in the multi-model layer obtains the prediction result of each sub-model from the current input historical sequence data,

The routing layer selects the prediction result of each sub-model with the output probability larger than the set output probability threshold according to the output probability of each sub-model,

Wherein,

The output probability of the sub-model is used for representing the similarity between the prediction result output by the sub-model and the mode characteristics of the current historical sequence data, the proportion occupied in the sum of the similarities between the prediction result output by each sub-model and the mode characteristics of the current historical sequence data,

The pattern feature is used to characterize the correlation of the current historical sequence data with the current timing pattern,

The current time sequence mode is used for representing each traffic data mode which is searched out through the time sequence and has high repeated occurrence probability.

2. The traffic data prediction method according to claim 1, wherein the prediction model is trained in the following manner:

Dividing the historical traffic data into time sequence segments with set sample time length, taking the time sequence segments as sample data, inputting the sample data into a prediction model to be trained,

Calculating a predictive loss function value of the predictive model based on the number of sub-models, the classified predictive loss function value of each sub-model, and the sample output probability of each sub-model, wherein the classified predictive loss function value includes a first type predictive loss function value for discarding a worse sub-model, and a second type predictive loss function value for selecting a better sub-model,

Back-propagation is performed based on the predictive loss function value of the predictive model,

And (5) training is repeated until the training is finished.

3. The traffic data prediction method according to claim 2, wherein the prediction loss function value of the prediction model is calculated as follows:

for any sub-model, calculating the product between the classified predicted loss function value of the sub-model and the logarithm of the sample output probability of the sub-model to obtain the predicted loss function value of the sub-model,

Calculating the average value of the predicted loss function values of all the sub-models to obtain the predicted loss function value of the predicted model;

The first type of predictive loss function value is determined as follows:

For any one of the sub-models,

In the case where the difference between the predicted sample value and the true value outputted by the sub-model is smaller than a set first difference threshold value, and the sample output probability of the sub-model is larger than the set sample output probability threshold value, the first-type predictive loss function value of the sub-model is set to 1,

In the case where the difference between the predicted sample value and the true value output by the sub-model is greater than a first difference threshold and the sample output probability of the sub-model is not greater than the sample output probability threshold, the first type predictive loss function value of the sub-model is set to the inverse of the number of sub-models minus 1,

Otherwise, the first type predictive loss function value of the sub-model is set to 0;

the second type of predictive loss function value is determined as follows:

For any one of the sub-models,

In case that the difference between the predicted sample value and the true value outputted by the sub-model is smaller than a second difference threshold value, and the sample output probability of the sub-model is larger than the sample output probability threshold value, the second type prediction loss function value of the sub-model is set to 1,

In the case where the difference between the predicted sample value and the true value output by the sub-model is greater than a second difference threshold and the sample output probability of the sub-model is not greater than the sample output probability threshold, the second type predictive loss function value of the sub-model is set to the inverse of the number of sub-models minus 1,

Otherwise, the second type predictive loss function value of the sub-model is set to 0,

Wherein,

The second difference threshold is the difference between 1 and the first difference threshold, the difference threshold is a fractional threshold used for representing the equal number points in the probability distribution range,

The output probability threshold is determined according to the number of types of output submodel predictors required.

4. A traffic data prediction method according to any one of claims 1 to 3, wherein the currently inputted history sequence data is inputted to the trained prediction model in such a manner that the history traffic data at the traffic location to be predicted is extracted according to the set history time step, so as to obtain the history traffic data at each history time point of the set history time step,

And sequentially inputting the historical traffic data at each historical time point into the trained prediction model according to the time sequence.

5. The traffic data prediction method according to claim 1, wherein the prediction result is prediction sequence data composed of predicted traffic data of each future time point, and wherein a set prediction time step is set between two adjacent future time points;

The output probability is determined as follows:

For any one of the sub-models,

Calculating multi-model layer predicted sequence data for characterizing the prediction result of the multi-model layer on the current historical sequence data according to the linear weight and the linear offset,

Calculating the correlation between the multi-model layer predicted sequence data and each traffic data mode in the current time sequence mode according to the multi-model layer predicted sequence data and the current time sequence mode, weighting each traffic data mode in the current time sequence mode according to each calculated correlation to obtain the mode characteristics of the current historical sequence data,

Calculating the similarity between the prediction result output by the sub-model and the mode characteristic of the current historical sequence data according to the mode characteristic of the current historical sequence data and the prediction result output by the sub-model,

Calculating the similarity between the prediction result output by each sub-model and the mode characteristic of the current historical sequence data according to the mode characteristic of the current historical sequence data and the prediction result output by each sub-model to obtain the similarity of each sub-model, accumulating the similarity of each sub-model to obtain the sum of the similarities,

And calculating the ratio of the similarity between the prediction result output by the sub-model and the mode characteristics of the current historical sequence data to the sum of the similarities to obtain the output probability of the sub-model.

6. The traffic data prediction method according to claim 5, wherein the calculating the multi-model layer prediction sequence data for characterizing the prediction result of the multi-model layer on the current history sequence data according to the linear weight and the linear offset comprises:

Accumulating the linear offset according to the product result of the linear weight and the current historical sequence data to obtain multi-model layer predicted sequence data of the current historical sequence data;

According to the multi-model layer prediction sequence data and the current time sequence mode, calculating the correlation between the multi-model layer prediction sequence data and each traffic data mode in the current time sequence mode comprises the following steps:

for any of the current timing modes,

Calculating an exponential function value based on a natural constant and taking the product of multi-model layer prediction sequence data and a transpose of a row vector used for representing the traffic data mode as an exponent to obtain the exponent function value of the traffic data mode, wherein the multi-model layer prediction sequence data is a row vector, the transpose of the row vector of the traffic data mode is a column vector, the current time sequence mode is a time sequence mode matrix, each row in the time sequence mode matrix corresponds to one traffic data mode, each column corresponds to one time sequence,

Calculating an exponential function value based on a natural constant and taking the product of the multi-model layer prediction sequence data and the transposed of the row vectors of each traffic data mode as an index to obtain an exponential function value of each traffic data mode, accumulating the exponential function values of each traffic data mode,

And calculating the ratio of the index function value of the traffic data mode to the accumulated index function value of each traffic data mode to obtain the correlation degree between the multi-model layer prediction sequence data and the traffic data mode.

7. The traffic data prediction method according to claim 6, wherein the weighting each traffic data pattern in the current time series pattern by each calculated correlation includes:

For the relevance of each traffic data mode, calculating the product of the relevance of the traffic data mode and the row vector of the traffic data mode to obtain the weighted row vector of the traffic data mode,

Adding the weighted row vectors of each traffic data mode to obtain a mode feature vector;

the prediction result of each sub-model with the selected output probability larger than the set output probability threshold value further comprises:

determining a fusion weight according to the output probability of each selected submodel,

And carrying out weighted summation on the selected prediction results by using the fusion weight to obtain the predicted traffic data.

8. The traffic data prediction method according to claim 6, wherein the time sequence sub-model is a time sequence large model for modeling that traffic time sequence characteristics of traffic data at any traffic location at a next time point are related only to traffic time sequence characteristics of traffic data at the traffic location at a previous time point,

The static space sub-model is a graph neural network model which is used for modeling that the traffic static space characteristics of traffic data at any traffic position at the next time point are related to the traffic static space characteristics of traffic data at the traffic position at the last time point and the traffic static space characteristics of traffic data at the static adjacent traffic position at the traffic position,

The dynamic space sub-model is a graph neural network model and is an attention mechanism model, the attention mechanism model is used for modeling the spatial characteristics of traffic data at any traffic position at the next time point to be correlated with the spatial characteristics of all traffic positions at the last time point, and the spatial characteristics are acquired through an attention mechanism.

9. A predictive model structure for traffic data prediction, the predictive model structure comprising:

A time sequence sub-model for acquiring time sequence characteristics of historical traffic data to obtain a first prediction result, a static space sub-model for acquiring static space characteristics of the historical traffic data to obtain a second prediction result, a multi-model layer formed by at least two of dynamic space sub-models for acquiring dynamic space characteristics of the historical traffic data to obtain a third prediction result, and

A routing layer for selecting the prediction results output by each sub-model in the multi-model layer;

Wherein,

Each sub-model in the multi-model layer obtains the prediction result of each sub-model from the current input historical sequence data,

The routing layer selects the prediction result of each sub-model with the output probability larger than the set output probability threshold according to the output probability of each sub-model,

Wherein,

The output probability of the sub-model is used for representing the similarity between the prediction result output by the sub-model and the mode characteristics of the current historical sequence data, the proportion occupied in the sum of the similarities between the prediction result output by each sub-model and the mode characteristics of the current historical sequence data,

The pattern feature is used to characterize the correlation of the current historical sequence data with the current timing pattern,

The current time sequence mode is used for representing each traffic data mode which is searched out through the time sequence and has high repeated occurrence probability.

10. A traffic data prediction apparatus, the apparatus comprising:

a prediction module for obtaining predicted traffic data of the traffic location to be predicted at a future time based on historical traffic data of the traffic location to be predicted at a historical time by using the trained prediction model for traffic data prediction,

Wherein,

The prediction model comprises a time sequence sub-model for acquiring time sequence characteristics of historical traffic data to obtain a first prediction result, a static space sub-model for acquiring static space characteristics of the historical traffic data to obtain a second prediction result, a multi-model layer formed by at least two dynamic space sub-models for acquiring dynamic space characteristics of the historical traffic data to obtain a third prediction result, and a routing layer for selecting the prediction result output by each sub-model in the multi-model layer;

The obtaining, by using the trained prediction model for traffic data prediction, predicted traffic data of the traffic location to be predicted at a future time based on historical traffic data of the traffic location to be predicted at a historical time, includes:

Each sub-model in the multi-model layer respectively acquires the prediction result of each sub-model from the current input historical sequence data, the routing layer selects the prediction result of each sub-model with the output probability larger than the set output probability threshold value according to the output probability of each sub-model,

Wherein,

The output probability of the sub-model is used for representing the similarity between the prediction result output by the sub-model and the mode characteristics of the current historical sequence data, the proportion occupied in the sum of the similarities between the prediction result output by each sub-model and the mode characteristics of the current historical sequence data,

The pattern feature is used to characterize the correlation of the current historical sequence data with the current timing pattern,

The current time sequence mode is used for representing each traffic data mode which is searched out through the time sequence and has high repeated occurrence probability.