CN118536009A - Power data model construction method and system based on generative artificial intelligence - Google Patents

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a method and a system for constructing an electric power data model based on generated artificial intelligence. The method comprises the following steps: collecting power load data; determining the abnormal probability of each data point in each period, and clustering to obtain a plurality of clusters; judging whether the cluster is an abnormal cluster according to the abnormal probability of the data points in the cluster; calculating a suitable degree coefficient; determining a target input length through the minimum value of the appropriate degree coefficient; and processing the power load data through a power data processing model to replace the abnormal data. According to the invention, the input data length of the data which is most suitable for the processing of the model is obtained through the fluctuation degree of the data and the length of the abnormal data which is identified through the abnormal algorithm, so that the model complements the input data (namely, the abnormal data is replaced by the generated data), and the obtained output data is more accurate.

Description

Power data model construction method and system based on generative artificial intelligence

Technical Field

The present invention relates to the field of artificial intelligence technology. More specifically, the present invention relates to a method and system for constructing a power data model based on generative artificial intelligence.

Background Art

Generative AI is the use of AI technology to generate content, such as natural language generation, image generation, audio generation, etc. In terms of natural language generation, generative AI can be used for text summarization, machine translation, dialogue systems, story generation, etc. The core idea of generative AI is to train a model to learn the distribution of data, and then use the learned model to generate new data. For example, in natural language generation, a model can be trained to learn the structure and grammatical rules of sentences, and then the model can be used to generate new sentences. In image generation, a model can be trained to learn the features and styles of images, and then the model can be used to generate new images.

In the power system, the accuracy and completeness of load data are crucial to the control and maintenance of the circuit system. Therefore, existing technologies can fill in missing or abnormal data through generative artificial intelligence. In practical applications, generative artificial intelligence can learn the existing data distribution and patterns, and then use these patterns to fill in missing data or repair abnormal data. For example, a generative adversarial network (GAN) can be used to generate new data that conforms to the real data distribution, thereby filling in missing or abnormal data points. Another common method is to use a sequence-to-sequence (seq2seq) model or a similar recurrent neural network (RNN) model to predict the value of missing or abnormal data points based on the existing data sequence.

Existing technologies usually identify data points with a large degree of outliers in charge data and replace them through generative artificial intelligence. For example, application document CN117150402A proposes a power data anomaly detection method and model based on a generative adversarial network. The invention processes power time series data through an anomaly detection model of a generator and a discriminator, and continuously optimizes the data generation ability and anomaly identification ability of the generator and the discriminator through adversarial training, thereby improving the anomaly detection ability of the model.

However, when inputting data into the model, even if the identified abnormal data is relatively accurate, the accuracy of the replaced charge data may still be low due to the large proportion of abnormal data in the input data.

Summary of the invention

The present invention provides a method and system for constructing an electric power data model based on generative artificial intelligence, aiming to solve the problem in the related art that there is a large amount of abnormal data in the input data, resulting in low accuracy of the replaced charge data.

In a first aspect, the present invention provides a method for constructing a power data model based on generative artificial intelligence, the method comprising: collecting power load data of power-consuming equipment in multiple consecutive time periods; for any time period, determining the abnormal probability of each data point, and clustering according to the abnormal probability to obtain a predetermined number of clusters; for any cluster, judging whether the cluster is an abnormal cluster according to the average value of _the abnormal probability of the data points in the cluster, wherein the data points in a cluster are continuous in time series; calculating a fitness coefficient Px : Px = exp( Ax + _{Bx )} , wherein Ax _is related to the length x of the input data and the data length of each abnormal cluster, Bx is related to the input length _x , the value _of the data point and its local entropy, and exp() is an exponential function with a natural constant e as the base; calculating the minimum value of the fitness coefficient Px by an optimization algorithm to determine the corresponding input length _x , recorded as the target input length xA ; processing the power load data using a pre-trained power data processing model to obtain power load data that replaces _the abnormal data; wherein the data length input to the power data processing model is the target input length xA _. , the abnormal data are data points belonging to abnormal clusters in the power load data.

The beneficial effects of the present invention are:

The inventors found that when replacing abnormal data in power load data through the power data model, the length of data input to the power data processing model will affect the accuracy of the data obtained after replacing the abnormal data. For example, if the length of data input to the power data processing model is short, then the less abnormal data will have a greater impact on the accuracy of the data obtained by model processing; if the length of data input to the power data processing model is long, the strength of smoothing the abnormal data may be small. Therefore, the inventors proposed to obtain the input data length that is most suitable for the data processed by the model through the fluctuation degree of the data itself and the length of the abnormal data identified by the abnormal algorithm, so that the model can complete the input data (that is, replace the abnormal data with the model-generated data), and the output data obtained is more accurate.

In one embodiment, the formula for calculating A _x is: , where ui _is the data length of all data in the i -th abnormal cluster in a period, M is the total number of abnormal clusters in a period, and uT _is the data length of a predetermined size.

In one embodiment, the formula for calculating B _x includes: for any time period, calculating the variance σ ² _j of the power load value corresponding to the j -th data point in the power load data, and normalizing the variance of the j -th data point to obtain the weight coefficient α _j of the j -th data point in each time period; obtaining the local entropy S _j of the j -th data point in each time period; for any number segment, extracting multiple groups of calculation data sets with a data length of the length x of the model input data in the time sequence of collecting data points in the time period; for any calculation data set, calculating the sum of the products of the weight coefficients corresponding to each data point in the calculation data set and the local entropy to obtain the fluctuation degree value of the calculation data set; the maximum value of the fluctuation degree values in all calculation data sets is B _x .

In one embodiment, for all time periods, counting whether the j -th data point in each time period is abnormal, so as to determine the probability that the j - th data point is abnormal data, includes: obtaining the number n _j of abnormal data points in each time period; obtaining the probability that the j -th data point is abnormal data: p _j = n _j / N , wherein p _j is the probability that the j -th data point in each time period is abnormal data, and N is the number of time periods.

Beneficial effects are as follows: the inventors have found that the power load data of each time period, for example, the power load data collected throughout the day, has similar changes in amplitude with the power load data of other time periods, and abnormal data will appear at the same time point in a time period. For example, from 13:00 to 14:00 every day, the power load data collected has a greater probability of being abnormal. Therefore, the present invention collects multiple time periods, and the proportion of the probability of the data collected at the same time point in each time period being abnormal is calculated as the probability of a failure in each time period at that time point. Based on this, by comparing the data of multiple time periods, it is possible to more accurately identify abnormal situations at specific time points, avoid the limitations of data in a single time period, and improve the accuracy and reliability of abnormal probability.

In one embodiment, judging whether the cluster is an abnormal cluster according to the average value of the abnormal probabilities of the data points in the cluster includes: for any cluster, calculating the average value of the probability that each data point in the cluster is abnormal data; in response to the average value being greater than a threshold, determining that the cluster is an abnormal cluster; in response to the average value being greater than a threshold, determining that the cluster is not an abnormal cluster.

The beneficial effect is: the average value of all data points in the cluster is taken as the probability that the cluster is an abnormal cluster. By comprehensively considering the abnormality of the entire cluster, the abnormality of the cluster is more comprehensively evaluated. Furthermore, by using the average value as the judgment indicator, the influence of a single abnormal data point on the abnormal judgment of the cluster can be reduced, thereby reducing the risk of misjudgment and improving the accuracy of judgment.

In one embodiment, the processing of the power load data using a pre-trained power data processing model includes: constructing a window with a length equal to the target input length x _A ; for any abnormal cluster: in response to the length of the window being greater than the length of the abnormal cluster, causing the window to cover the abnormal cluster for processing; in response to the length of the window being less than the length of the abnormal cluster, causing the window to traverse the abnormal cluster, and during the first traversal, causing a portion of the window to cover the head end of the abnormal cluster; during the traversal process, sliding the window.

The beneficial effect is that for abnormal clusters with smaller data length, the window can directly cover them, so that they include normal data and abnormal data, and the abnormal data is replaced by the power data processing model. By making the window cover the non-abnormal data (normal data), it is ensured that the replaced data can better reflect the actual situation, making the data output by the model more accurate. For abnormal clusters with larger data length, the window cannot be completely covered. At this time, the traversal starts from the end where the window partially covers the abnormal cluster, and the window is slid during the traversal process. Through localized processing, the data in the window better retains the original characteristics, thereby maintaining the integrity and continuity of the data.

In one embodiment, the anomaly detection algorithm is a random forest anomaly detection algorithm.

The beneficial effect is that the power load data of the present invention is single-dimensional data (current data, power data, etc.) and there is a time series relationship between each data point, which is suitable for the isolation forest anomaly detection algorithm. The reason is that the time series relationship between data points can be considered when constructing a random tree, so that the time series information can be used to more effectively judge the abnormal points.

In one embodiment, the optimization algorithm is a PSO optimization algorithm.

In a second aspect, the present invention provides a power data model construction system based on generative artificial intelligence, which includes a processor and a memory, the memory stores a computer program, and the processor executes the computer program to implement a power data model construction method based on generative artificial intelligence as described in any one of the above invention contents.

BRIEF DESCRIPTION OF THE DRAWINGS

By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

FIG1 is a flowchart schematically illustrating the steps of a method for constructing a power data model based on generative artificial intelligence according to an embodiment of the present invention;

FIG2 is a flow chart schematically illustrating step S4 according to an embodiment of the present invention;

FIG3 is a block diagram schematically showing a structure of a power data model building system based on generative artificial intelligence according to an embodiment of the present invention.

DETAILED DESCRIPTION

The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

FIG1 is a flowchart schematically illustrating the steps of a method for constructing a power data model based on generative artificial intelligence according to this embodiment.

As shown in FIG1 , the method for constructing a power data model based on generative artificial intelligence includes steps S1 to S4.

Step S1: Collecting power load data of electrical equipment in multiple consecutive time periods.

The power load data is single-dimensional data, which is current data, voltage data or frequency data. The power load data corresponds to a power load value, which is the current value, voltage value or frequency of the power equipment. The length of a time period is equal to a working cycle of the power equipment.

In one embodiment, the working hours of industrial electrical equipment are from 10:00 to 16:00 every day, so one period is 6 hours (from 10:00 to 16:00 a day), and the power load data in multiple consecutive periods are the current values of the industrial electrical equipment collected from 10:00 to 16:00 for multiple consecutive days. The power load data can also be the power size of the industrial electrical equipment.

In another embodiment, a household electrical appliance, such as a refrigerator, works from 0:00 to 24:00 every day (working all day), and a period is 24 hours (10:00 to 16:00 a day). The power load data in multiple consecutive periods are the current values of the household electrical appliances collected continuously for multiple consecutive days. The power load data can also be the power size of the household electrical appliances.

Step S2: for any time period, determine the abnormal probability of each data point, and cluster according to the abnormal probability to obtain a predetermined number of clusters; for any cluster, determine whether the cluster is an abnormal cluster according to the average value of the abnormal probability of the data points in the cluster.

In one embodiment, the algorithm for clustering data points is the DBSCAN clustering algorithm. It should be noted that when the DBSCAN clustering algorithm is used for classification in this embodiment, the category radius is 0.05, the minimum number of neighbors is 5, and the time values corresponding to all time points in a single cycle and the values of the abnormal probability of each data point are normalized during classification to eliminate dimensional interference. In addition, the data points in the clusters obtained by the DBSCAN clustering algorithm are continuous in time series. For example, a cluster includes the 10th to 20th data points collected in a time period.

In one embodiment, the process of determining the abnormal probability of each data point is as follows: obtaining the _{number nj} of abnormal data points in each time period, wherein the random forest algorithm _{is used to determine whether the jth data point is abnormal. Obtaining the probability that the jth data point is abnormal data: pj = nj} / N , wherein pj is _the probability that the jth data point in each time period is abnormal data, _and N is the number of time periods.

Furthermore, the process of judging whether a cluster is an abnormal cluster according to the average value of the abnormal probability of the data points in the cluster is as follows: for any cluster, the average value of the probability that each data point in the cluster is abnormal data is calculated; in response to the average value being greater than a threshold, the cluster is determined to be an abnormal cluster; in response to the average value being greater than a threshold, the cluster is determined not to be an abnormal cluster. In this embodiment, the threshold is 0.3.

Step S3: Calculate the fitness coefficient P _x : P _x =exp( A _x + B _x ); calculate the minimum value of the fitness coefficient P _x through an optimization algorithm to determine the corresponding input length x, recorded as the target input length x _A .

Among them, A _x is related to the length x of the input data and the data length of each abnormal cluster, B _x is related to the input length x , the value of the data point and its local entropy, and exp() is an exponential function with the natural constant e as the base. The fitness coefficient P _x is positively correlated with A _x , and the fitness coefficient P _x is positively correlated with B _x .

In one embodiment, the formula for calculating A _x is: , ui is the data length of _all data in the ith abnormal cluster in a period, M is the total number of abnormal clusters in a period, and uT is the data length of a predetermined size. Among them, _{Ax is} positively correlated with the square of the data length of all data in each abnormal cluster. When Ax _is smaller, it means that the proportion of data in the abnormal cluster in the input data is lower, and the accuracy of the output data obtained by supplementing the input data is higher.

Furthermore, the process of calculating B _x is as follows: the variance σ ² _j of the power load value corresponding to the j -th data point in all time periods of the power load data is calculated, and the variance of the j -th data point is normalized to obtain the weight coefficient α _j of the j -th data point in each time period; the local entropy S _j of the j -th data point in each time period is obtained; for any number of segments, multiple groups of calculation data sets with a data length of the length x of the model input data are sequentially extracted in the time sequence of the data points collected in the time period; for any calculation data set, the sum of the product of the weight coefficient corresponding to each data point in the calculation data set and the local entropy is calculated to obtain the fluctuation degree value of the calculation data set; the maximum value of the fluctuation degree values in all calculation data sets is B _x . Among them, the smaller B _x is, the smaller the fluctuation of the input data input into the model in the data collected in a time period (the product of the variance corresponding to the data point and the local entropy is small), and the accuracy of the output data output by the model is high.

The local entropy S _j of the jth data point in each time period is obtained by selecting power load data with a time length of t with the jth data point as the center and obtaining the local entropy S _j of the j data points through an entropy calculation method (such as Shannon entropy).

In summary, when A _x is smaller, P _x is smaller, and the accuracy of the model output data is higher; when B _x is smaller, P _x is smaller, and the accuracy of the model output data is higher. When A _x is larger, P _x is larger, and the accuracy of the model output data is lower; when B _x is larger, P _x is larger, and the accuracy of the model output data is lower.

Step S4: Process the power load data using a pre-trained power data processing model to obtain power load data that replaces the abnormal data.

In one embodiment, the power data processing model expands the input data through a VAE network so that the input data length is the same as the output data length. In another embodiment, the input data is expanded through a GAN network.

The data length input to the power data processing model is the target input length x _A , and the abnormal data is the data points belonging to the abnormal cluster in the power load data.

As shown in FIG. 2 , step S4 includes step S401 to step S403 .

Step S401: construct a window with a length equal to the target input length x _A.

Step S402: For any abnormal cluster: in response to the length of the window being greater than the length of the abnormal cluster, the window is made to cover the abnormal cluster for processing.

Step S403: for any abnormal cluster: in response to the length of the window being less than the length of the abnormal cluster, the window is traversed through the abnormal cluster, and during the first traversal, a portion of the window covers the head end of the abnormal cluster; during the traversal, the window is slid, and a portion of the sliding window covers the abnormal cluster that has not been traversed.

In one embodiment, it is necessary to replace the data points belonging to the abnormal clusters within a time period (delete them and generate new data through the data points of the non-abnormal clusters). When the data length of the data in one of the abnormal clusters is less than the window length, align the center of the window with the center of the data in the abnormal cluster, and input the data in the window into the model as input data. When the data length of the data in one of the abnormal clusters is greater than the window length, a part of the window covers the head end of the abnormal cluster, and after the data in the window is input into the model as input data, the output data of the model is replaced with the original data in the window, and the window is slid so that a part of the window covers the unprocessed data belonging to the abnormal cluster until the window traverses the entire abnormal cluster.

The length of the window covering the first segment of the abnormal cluster is less than or equal to half of the total length of the window, so that the data in the abnormal cluster accounts for a smaller proportion and the model output data is more accurate.

The present invention also provides a system for constructing a power data model based on generative artificial intelligence. As shown in FIG3 , the system includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a method and system for constructing a power data model based on generative artificial intelligence according to the first aspect of the present invention are implemented.

Those skilled in the art will understand that the structure shown in FIG. 3 is merely a block diagram of a partial structure related to the solution of the present invention, and does not constitute a limitation on the computer device of the present invention. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

The system also includes other components familiar to those skilled in the art, such as a communication bus and a communication interface, whose configuration and functions are known in the art and thus will not be described in detail here.

In the present invention, the aforementioned memory may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, apparatus or device. For example, a computer-readable storage medium may be any appropriate magnetic storage medium or magneto-optical storage medium, such as a resistive random access memory RRAM (Resistive Random Access Memory), a dynamic random access memory DRAM (Dynamic Random Access Memory), a static random access memory SRAM (Static Random-Access Memory), an enhanced dynamic random access memory EDRAM (Enhanced Dynamic Random Access Memory), a high-bandwidth memory HBM (High-Bandwidth Memory), a hybrid memory cube HMC (Hybrid Memory Cube), etc., or any other medium that can be used to store the required information and can be accessed by an application, a module, or both. Any such computer storage medium may be part of a device or accessible or connectable to a device. Any application or module described in the present invention may be implemented using computer-readable/executable instructions that may be stored or otherwise maintained by such a computer-readable medium.

In the description of this specification, "plurality" or "several" means at least two, such as two, three or more, etc., unless otherwise clearly and specifically defined.

The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent application. It should be pointed out that for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims (9)

1. The power data model construction method based on the generated artificial intelligence is characterized by comprising the following steps of:

Collecting power load data of electric equipment in a plurality of continuous time periods;

For any period, determining the abnormal probability of each data point, and clustering according to the abnormal probability to obtain a preset number of clusters; judging whether any cluster is an abnormal cluster according to the average value of the abnormal probability of the data points in the cluster, wherein the data points in one cluster are continuous in time sequence;

Calculating a fitness coefficient P _x：P_x=exp(A_x+B_x, wherein A _x relates to the length x of input data and the data length of each abnormal cluster, B _x relates to the input length x, the value of a data point and the local entropy thereof, and exp () is an exponential function with a natural constant e; calculating the minimum value of the suitability coefficient P _x through an optimization algorithm to determine the corresponding input length x, and recording the corresponding input length x as a target input length x _A;

Processing the power load data by using a pre-trained power data processing model to obtain power load data for replacing abnormal data; the data length input into the power data processing model is the target input length x _A, and the abnormal data is data points belonging to an abnormal cluster in power load data.

2. The method for constructing a power data model based on generated artificial intelligence according to claim 1, wherein the formula for calculating a _x is: Where u _i is the data length of all data in the ith abnormal cluster in one period, M is the total number of abnormal clusters in one period, and u _T is the data length of a predetermined size.

3. The method for constructing a power data model based on generative artificial intelligence as claimed in claim 1, wherein the formula for calculating B _x comprises:

Calculating the variance sigma ² _j of the power load value corresponding to the jth data point in all the time periods in the power load data, and normalizing the variance of the jth data point to obtain a weight coefficient alpha _j of the jth data point in each time period; obtaining local entropy S _j of the j-th data point of each period;

for any number of segments, sequentially extracting a plurality of groups of calculation data sets with the data length being the length x of the model input data in the time sequence of data point collection in the period;

For any calculation data set, calculating the sum of products of the weight coefficients corresponding to all data points in the calculation data set and the local entropy to obtain a fluctuation degree value of the calculation data set;

the maximum value of the fluctuation degree values in all the calculation data sets is B _x.

4. The method for constructing a power data model based on generated artificial intelligence according to claim 1, wherein counting whether the j-th data point of each period is abnormal for all periods of time, thereby determining the probability that the j-th data point is abnormal data comprises:

obtaining the number n _j of data points of which the j-th data point of each period is abnormal;

Obtaining the probability that the jth data point is abnormal data: p _j=n_j/N, where p _j is the probability that the jth data point of each period is abnormal data, and N is the number of periods.

5. The method for constructing a power data model based on a generative artificial intelligence according to claim 1, wherein the determining whether the cluster is an abnormal cluster according to an average value of abnormal probabilities of data points in the cluster comprises:

For any cluster, calculating the average value of the probability that each data point in the cluster is abnormal data;

Determining that the cluster is an abnormal cluster in response to the average value being greater than a threshold value;

And in response to the average value being greater than a threshold, determining that the cluster is not an outlier cluster.

6. The method for constructing a power data model based on generated artificial intelligence according to claim 1, wherein the processing the power load data by using a pre-trained power data processing model comprises:

Constructing a window with the length of the target input length x _A;

for any outlier cluster:

In response to the length of the window being greater than the length of the abnormal cluster, covering the abnormal cluster with the window for processing; in response to the length of the window being less than the length of the abnormal cluster, traversing the window through the abnormal cluster for the first time, and enabling a part of the window to cover the head end of the abnormal cluster; and sliding the window in the traversal process, wherein a part of the window after sliding covers the abnormal cluster which is not traversed.

7. The power data model construction method based on the generated artificial intelligence according to claim 1, wherein the anomaly detection algorithm is a random forest anomaly detection algorithm.

8. The power data model construction method based on the generated artificial intelligence according to claim 1, wherein the optimization algorithm is a PSO optimization algorithm.

9. A power data model construction system based on generative artificial intelligence, comprising a processor and a memory, the memory storing a computer program, characterized in that the processor executes the computer program to implement the power data model construction method based on generative artificial intelligence as claimed in any one of claims 1-8.