DE102024101854A1 - Method for continuously providing measurement data of a measured variable of a power transformer - Google Patents

Abstract

Method (100) and system (41) for the continuous provision of measurement data (34, 40) relating to a measured variable of a power transformer (10). A higher accuracy of substitute calculated model measurement data (40) of a sensor (11, 38) in parallel with corresponding real measurement data (34) of a sensor (11) is achieved by an optimized selection (140) of a preference data model (54) from a plurality of potentially suitable data models (50) in conjunction with automatic condition monitoring of a sensor signal (34). In the event of a faulty or failed physical sensor (11), higher-level transformer control systems (42) can thus be provided with more precise and reliable model measurement values (40) of the power transformer (10). At the same time, the number of necessary physical sensors (11) on the power transformer (10) and thus costs and complexity can be reduced. Furthermore, the presented solution enables the implementation of additional virtual sensors (38) and additional measured variables without additional physical sensor hardware.

Description

The invention relates to power transformers for energy networks and their monitoring and operation. In particular, the invention relates to a method and system for continuously providing measurement data relating to a measured variable of a power transformer.

Due to changing topologies and increasing complexity, future energy grids must be reliably controllable and manageable. For example, these assets should be able to tolerate short-term peak loads. At the same time, the assets must be operated in a safe and cost-effective mode. One option here may be to operate power transformers in an overload situation for a limited period of time. However, this requires reliable data that can more accurately determine the effects of such temporary overload situations on the remaining service life of the transformer, its safety, and other properties. Under such controlled conditions, overload situations could be temporarily enabled in an economically and technically safe manner.

A higher-level control system can monitor certain measured variables of a transformer in order to optimize its function, safety and economic efficiency. US2017011612A1 For example, it describes a system for monitoring and controlling transformer operating parameters and calculating service life, including fault diagnosis. Such a control or monitoring system requires measurement data from a wide variety of measured variables as input variables. Examples include physical variables such as temperatures at various points within the transformer, as well as ambient temperatures, oil temperature, gas composition, degree of polymerization of insulation materials, electrical load, field strengths, and other electrical variables.

These measured variables are predominantly recorded using sensors located at suitable locations on the power transformer or in its surroundings. To ensure stable operation of the control and monitoring system, continuous availability of measurement data is necessary. At the same time, sensor failures or measurement errors, for example, due to aging or environmental influences, are not uncommon. Therefore, greater robustness and lower susceptibility to errors would be desirable in the event that sensors fail or provide incorrect measurement data.

Furthermore, additional measured variables can provide additional information about the condition of a transformer beyond the sensor-measured data. However, additional hardware sensors on the transformer increase complexity and costs. Therefore, it is also desirable to obtain highly meaningful information from the smallest possible number of sensors on the transformer. In summary, it should be possible to reliably provide measurement data to a higher-level control system even in failure or fault situations. Although methods for mathematically simulating sensors are known, the accuracy of the simulated and calculated substitute measured values is often insufficient.

Embodiments of the invention can provide measurement data for specific measured variables significantly more reliably and continuously, even in error situations or sensor failure situations. At the same time, the accuracy of the measurement data generated alternatively via simulation models is to be increased compared to known solutions. This object is achieved by the subject matter of the independent claims. Further embodiments of the invention emerge from the dependent claims and the following description.

The invention described below is based on the following solution concepts: A first fundamental approach to the solution is to use a mathematical model or digital twin of the sensor to provide a substitute measured value in the event of a sensor failure. However, a multitude of different potentially suitable data models with very different properties and parameters exist.

A method is proposed for the continuous provision of measurement data relating to a measurand of a power transformer. A measurand here is understood to mean various technically conceivable physical, chemical, electrical, or mechanical measurands that are suitable for describing a state, an operating characteristic of a power transformer. Examples of this are temperature, chemical composition and other chemical properties such as degree of polymerization, electrical measurands such as power, currents, voltages, digital quantities, field strengths, radiation values, vibrations, or oscillations. The measurement data can be understood to mean concrete quantitative characteristics, i.e. measured values in digital or analog form relating to this measurand. These do not necessarily have to be in digital form as data or data sets, but can also be in the form of analog signals or recordings.

A control unit of the power transformer, which can be configured for monitoring, parameterization, optimization, and configuration of a power transformer, requests measurement data for a measured variable. In other words, the control unit requires specific measurement data for a measured variable in order to effectively monitor and control the transformer, detect fault conditions, or generate suggested actions. A model database provides a variety of data models. These data models are suitable for calculating model measurement data for a measured variable based on reference measurement data for reference measured variables.

This means that these data models are capable of calculating substitute model measurement data from other measured variables or other input data. These data models therefore contain the causal, functional, or statistical relationships between various input variables that influence the measured variable. These influencing variables and dependencies are generally complex. The data models are defined by their basic structure and have additional parameters that can be used to adapt a model to specific needs and levels of accuracy. In one example, the reference measurement data was generated at the power transformer in question. In another example, the reference measurement data was generated at a second power transformer, for example, as part of a transformer fleet.

In other words, the data models can use not only data from the same power transformer, but also measurement data from ideally similar or identical power transformers. This effectively allows data from one asset in a fleet of transformers to be used in models of other transformers without requiring additional data collection effort. According to one example, reference measurement data can be model measurement data calculated using a data model or simulation. According to another example, the reference measurement data is sensor-recorded measurement data.

The data models are designed to store the interdependencies of the individual influencing variables among themselves and with regard to the result and method of calculating the model measurement data. Since the data models differ in their structure and properties, it is important to select a preferred or best data model for the respective application. Therefore, the next step involves selecting a preferred model from the multitude of provided data models using a simulation unit. A simulation unit can, for example, be a computing unit configured to generate and store a digital twin of a sensor or a mathematical representation of a sensor and its physical behavior.

The selection of the preference model is based on the smallest possible deviation of the calculated model measurement data of the preference model from the corresponding real measurement data of this measured variable. Real measurement data can be understood, for example, as measurement data recorded by sensors on the power transformer in question or generated by similar transformers in the fleet related to a specific point in time or period. The real measurement data used for training are preferably clean, correct, or usable signals in terms of signal technology in order to achieve optimal training results. In another example, real measurement data can be historical real measurement data from the same or other power transformers. For example, the selection step can be a comparison between real, actual real-time measurement data from a functioning sensor and, in parallel, model measurement data from a data model.

During the selection process, the simulation unit compares the calculated model measurement data with the actual real sensor measurements, i.e., the real measurement data for this measurand. Based on this, the model is selected from the multitude of available data models with the smallest deviation between the corresponding real measurement data and the model measurement data. In other words, the model error is lowest for the preferred model.

The deviation or model error can relate to various criteria, such as amplitudes, frequencies, envelopes, signal shapes, signal quality, noise components, and the like. In the next step, the preference model is trained using training data by the simulation unit. Training can, on the one hand, be understood as training using training data sets, for example in connection with artificial neural networks. This term also includes adjustments to parameters in the relevant data models, such as linear regression. In general, the term training refers to adjusting the parameters of a data model in order to achieve the best possible approximation of the model measurement data to the real measurement data and in order to achieve a significantly higher accuracy of substitute generated model measurement data.

The training data includes reference measurement data of reference metrics and/or real measurement data of the metric. According to one example, the training data are pre-qualified and validated data sets from reference measurement data or real measurement data, which may include real-time measurement values as well as historical measurement data. The preference model is trained in such a way that parameters of the preference model are adjusted in such a way that any deviation of the model measurement data calculated by the preference model from the real measurement data of the training data is minimized. In other words, the model measurement values calculated by the preference model and the real measurement values become increasingly similar over time until only a small model error remains. For further information, please refer to the explanations on 6 As a result, the accuracy of the model measurement data is significantly improved.

A simplified example will illustrate the correlation of measurement data of a measurand with reference measurement values. For example, a physical gas sensor measures a gas concentration using an electrical resistance value. This resistance value, in turn, depends on a temperature value. If the primary temperature sensor for this temperature value fails or malfunctions, a suitable data model can be used to utilize a temperature value at another location in the power transformer, and the missing temperature value can be derived for the correct calculation of a gas concentration.

According to one example, a maximum model error is defined, which serves as a threshold for ending training. Thus, once training of a preferred data model reaches a certain level of accuracy, training is terminated. Input data for such a model can come from various sources. Real measurement values can be read online while the power transformer is in operation. Alternatively, the measurement data can be collected offline on-site at the transformer. Furthermore, fleet information, i.e. historical or real-time real measurement values from transformers at other locations, can be used as a source of reference measurement data or real measurement data. Using the procedure described here, virtual sensor values, which do not physically exist as hardware sensors but are based only on calculations, can also be used as a data source.

In the next step, a monitoring unit calculates a deviation between the model measurement data of the measurand calculated by the preference model and the real measurement data of the measurand. The monitoring unit continuously compares the values from the simulation, i.e. from the preference data model, with the corresponding incoming actual sensor values. In this case, the trained model serves as a reference and benchmark for a properly functioning sensor, since the training was based on verified real measurement data or reference measurement data. In other words, the prediction error is used as a measure of anomalies, i.e., to determine whether the sensor is functioning properly or whether there is a fault in the sensor or the transmission media. Here, too, a deviation can relate to various criteria, such as signal shape, amplitude, temporal patterns of values, noise components, and others.

In a further step, the model measurement data of the measured variable is transmitted to the control unit of the transformer if the deviation reaches or exceeds a defined threshold. Alternatively, the real measurement data is transmitted to the control unit of the power transformer if the calculated deviation falls below the threshold. If, for example, the monitoring unit detects a faulty sensor or a noisy sensor signal due to a high deviation equal to or above the threshold, the model measurement value calculated using the preference model is transmitted instead of the faulty or failed sensor signal. This ensures that, despite a sensor error or a failed sensor, measurement data of sufficient accuracy and quality is still transmitted to a higher-level control unit. This ensures that higher-level systems continue to function and can trigger appropriate actions. At the same time, in conjunction with the selection of a preferred data model, the accuracy of the model measurement value can be increased.

In one embodiment, the preference model is retrained by the simulation unit when the deviation between real measurement data of the measured variable and model measurement data reaches or exceeds a defined threshold. If the monitoring unit detects an anomaly, i.e., a deviation between real measurement values and model measurement values above the defined threshold, one of the causes may be an unfavorable parameterization of the data model. The retraining is performed in such a way that the deviation between model measurement data of the model of the measured variable and real measurement data of the measured variable is reduced. Retraining can be carried out in In one example, the modeling process can be terminated when the deviation falls below the threshold. This improves model accuracy and reduces model errors. In principle, the method presented here can be universally applied to various types of actual or virtual sensor values.

In one embodiment, the retraining of the preference model and the adjustment of the preference model parameters are repeated at defined intervals. According to one example, a repetition occurs every 12 hours or 24 hours. In other words, the data model is continuously adapted to the current real measurement data of the last 12 or 24 hours, thus further improving accuracy. According to another embodiment, the provision of the data models comprises an IEC 60076-7 model, an ETF model (Estimate Transfer Function), a model based on linear regression, and/or an IEEE Std C57-91-2011 Annex G model. Depending on the application and context, modeling the complex relationships between a measured variable and reference measured variables may require different models, each with its own characteristics. Examples of possible models are described in more detail in the exemplary embodiments.

In one embodiment, the preference model is an IEC 60076-7 model, and the simulation unit calculates the parameters of the preference model by solving the differential equation of the preference model. The parameters are determined by minimizing the model error, the penalty term (short-term change), and the penalty term (initial value change). The determination of the parameters is based on measured data of the ambient temperature, maximum winding temperature, top oil temperature, and electrical load. A particular advantage can be seen in the fact that many of the necessary parameters of the data model are not determined manually but calculated automatically. This requires significantly less interaction with customers and faster productive availability of the model.

In one embodiment, the measured variable is a hotspot temperature of a power transformer, and the reference measured variables are an average oil temperature, a top oil temperature, an electrical load of the power transformer, and/or an ambient temperature. This describes a use case where, for example, a physical sensor for measuring the hotspot temperature is faulty or has failed. Taking into account the physical and mathematical relationships between the hotspot temperature to be determined and the specified temperatures and the electrical load, the solution presented here can transmit model measurement data to the higher-level control unit despite a faulty or failed sensor, thus still enabling the correct functioning of the overall system.

In one embodiment, the measured variable is a degree of polymerization of the paper insulation of a power transformer. In another embodiment, the measured variable relates to a gas analysis (DGA) of a power transformer. Thus, various types of measured variables can be calculated using the data models or the preference model. In one embodiment, before the preference model is selected, a plurality of data models are pre-trained based on training data. The training data comprises reference measurement data of reference measured variables and/or real measurement data of the measured variable. The preference model is trained in such a way that parameters of the preference model are adjusted such that a deviation of the model measurement data from the real measurement data of the training data is reduced.

The underlying idea is that, for a qualified selection of a suitable data model, a large number of possible models must be considered. On the other hand, the subsequent behavior of a model in a specific context with respect to a specific measured variable should be represented as realistically as possible. For this purpose, it is useful to pre-train the models. This is intended to find a compromise between training effort and accuracy. The training phase continues, as already described above, in the step of training the preference model with training data by the simulation unit.

In one embodiment, the deviation of the calculated measurement data of the preference model from the corresponding real measurement data of this measurement variable refers to a noise component in the measurement data. Noisy sensor signals can be a reason why measurement data from physical sensors are unusable. In this case, a deviation should be understood as, for example, calculating a signal-to-noise ratio and comparing it with a threshold value. In another embodiment, the deviation of the calculated measurement data of the preference model from the corresponding real measurement data of this measurement variable describes an amplitude, measurement data pattern and/or signal shape of the measurement data. In other words, the aim is to detect whether, for example, the sensor is defective, implausible or incorrect measurement data is present, the signal is distorted, its amplitude is too weak, too strong, or is superimposed with a disturbance. Suitable parameters can be used for this purpose. which are then compared with a set threshold value to decide whether real measurement data or model measurement data are transmitted to the control unit.

In one embodiment, the real measurement data is sensor-captured measurement data. This means that this measurement data can originate from one of the physical sensors arranged on the power transformer. In contrast, according to another embodiment, the real measurement data is measurement data obtained offline using laboratory technology or laboratory analysis. In contrast to sensors attached to the power transformer that generate measurement data on-site, there are laboratory-based options for determining measurement data, for example, when analyzing oil or paper.

In a further aspect of the invention, a system is proposed that comprises a control unit for monitoring a power transformer, a monitoring unit for calculating a deviation between model measurement data and real measurement data, and a model database for storing data models. The simulation unit is configured to calculate model measurement values based on data models. The system is configured and implemented such that it can execute the method steps described above.

It should be understood that features of the method as described above and below may also be features of the system and vice versa.

• 1 zeigt einen Leistungstransformator mit verschiedenen physischen Sensoren und Messgrößen.
• 2 zeigt das Grundprinzip der ersatzweisen Berechnung von Modellmessdaten.
• 3 zeigt ein erfindungsgemäßes Verfahren zum kontinuierlichen Bereitstellen von Messdaten einer Messgröße eines Leistungstransformators.
• 4 zeigt schematisch ein System zur Bereitstellung von Messdaten mit einer Simulationseinheit, einer Überwachungseinheit, einem physischen Sensor und einer Steuereinheit gemäß der Erfindung.
• 5 zeigt ein Anwendungsbeispiel eines Systems gemäß der Erfindung mit einer optimierten Modellauswahl und einem Zustandsautomaten mit Überwachungseinheit.
• 6 zeigt ein Diagramm des zeitlichen Verlaufs von Realmessdaten und Modellmessdaten während eines Trainings eines Datenmodells.

In the following, exemplary embodiments of the invention are described in detail with reference to the accompanying figures. Neither the description nor the figures should be construed as limiting the invention.

• 1 shows a power transformer with various physical sensors and measured variables.
• 2 shows the basic principle of the substitute calculation of model measurement data.
• 3 shows a method according to the invention for continuously providing measurement data of a measured variable of a power transformer.
• 4 shows schematically a system for providing measurement data with a simulation unit, a monitoring unit, a physical sensor and a control unit according to the invention.
• 5 shows an application example of a system according to the invention with an optimized model selection and a state machine with monitoring unit.
• 6 shows a diagram of the time course of real measurement data and model measurement data during training of a data model.

The drawings are merely schematic and not to scale. Identical or similar parts are generally designated by the same reference numerals.

1 shows a power transformer with various physical sensors 11. The sensors 11 measure different parameters and are mounted at different areas of the power transformer 10. Temperature sensors 12 measure different temperatures at various locations on the power transformer 10. Examples include the inlet and outlet temperatures of the radiator and heat exchanger, the temperature in the OLTC motor drive, the temperature in the OLTC oil tank, the temperature in the transformer tank, the contact temperature of the electrical connections, and the temperature of the winding and core. Pressure sensors 14 measure pressure conditions, such as the pressure in the transformer tank or the pressure in the OLTC tank.

Flow sensors 16 measure, for example, the flow rate in the heat exchanger or the pump. A vibration sensor 20 is designed to measure vibrations in the OLTC, for example, during switching operations. A level sensor 22 is designed to measure the oil level in the expansion tank 32. Humidity sensors 24 measure, for example, moisture in the boiler oil or the oil moisture content in the OLTC tank. Gas sensors 26 can provide measurement data regarding gas compositions such as DGA (Dissolved Gas Analysis) or measure collected gas quantities.

In 2 The basic principle of the substitute calculation of measurement data 34, 40 is shown in simplified form in a virtual sensor 38. A virtual sensor 38 is intended to mean a mathematical model that substitutely maps the functionalities of a physical sensor 11. Specifically, for example, the behavior of output variables as a function of input variables 36 is to be mapped. The input variables 36 can be physical measured variables such as temperature, pressure, electrical variables, and others. In the case of a sensor, the output variables are the associated real measurement data 34. An input variable 36, for example a specific temperature, is fed to the physical sensor 11.

The physical sensor 11 generates real measurement data 34, which are made available to a higher-level system for further processing. In the event of failure or malfunction of the physical sensor 11, a digital twin or virtual sensor 38, i.e. a digital or mathematical representation of the behavior of the physical sensor 11, is intended to provide measurement data. The virtual sensor 38 learns from the behavior of the various input variables 36 and the associated real measurement data 34 and generates a mathematical representation of the functional or statistical relationships. As a result, the virtual sensor 38 calculates model measurement values 40 as a substitute. It is fundamentally possible for the virtual sensor 38 and the physical sensor 11 to work in parallel and simultaneously provide real measurement data 34 and model measurement data 40. With the help of the state machine (monitoring unit 44) described below, the present invention further solves the problem of deciding in which situation real measurement data 34 or model measurement data 40 are passed on to a higher-level control unit, for example a monitoring system for power transformers.

In 3 A method 100 according to the invention for continuously providing measurement data 34, 40 of a measured variable of a power transformer 10 is presented. In a step 110, a control unit 42 requests measurement data 34, 40 of a measured variable from a sensor 11, 38. In step 120, a plurality of data models 50 are provided in a model database 46. These provided data models 50 are suitable for calculating model measurement data 40 of a measured variable as a function of reference measurement data 48 of reference measured variables. Reference measured variables can be physical, mathematical, or other measured variables that are related in any mathematical, statistical, functional, or otherwise representable way. The purpose of these reference measured variables and data models 50 is to calculate model measurement data 40 from third variables without requiring a physical sensor 11 of this associated measured variable.

In step 130, the data models 50 are pre-trained to enable a more qualified selection of a most suitable data model 50. In other words, the data models 50 are already more accurate due to initial pre-training and thus allow a more precise assessment of the best suitability for a virtual sensor 38. The pre-training 130 is performed on a plurality of data models 50 based on training data 52, wherein the training data 52 comprises reference measurement data 48 of reference measured variables and/or real measurement data 34 of the measured variable. The pre-training 130 of the data models 50 is performed in such a way that parameters of the data models 50 are adjusted so that a deviation of the model measurement data 40 from the real measurement data 34 of the training data 52 is reduced.

In step 140, a preference model 54 is selected by a simulation unit 56. The preference model 54 is understood to be a most suitable data model 50 in which a model error is the smallest. According to one example, the selection 140 is performed by sequentially testing the individual models 50 by calculating a corresponding model error by the simulation unit 56. According to one example, the selection criterion for a preference model 54 is processing speed, energy efficiency, storage space requirements, and/or resource requirements instead of the model error.

Step 150 relates to training the preference model 54 with training data 52 by the simulation unit 56, wherein the training data 54 comprises reference measurement data 48 of reference measurement variables and/or real measurement data 34 of the measurement variable. The training 150 of the preference model 54 is carried out in such a way that parameters of the preference model 54 are adjusted such that a deviation of the model measurement data 40 calculated by the preference model 54 from the real measurement data 34 of the training data 52 is minimized. In a subsequent step 160, a monitoring unit calculates 160 a deviation between the model measurement data 40 of the measurement variable calculated by the preference model 54 and the real measurement data 34 of the measurement variable.

This deviation is intended to serve as a measure of the extent to which real measurement data 34 generated by a physical sensor 11 differs from the corresponding model measurement data 40 of a data model 50 trained for this sensor, which serves as a benchmark. The deviation can relate to various criteria, such as amplitude, signal shape, noise content, superimposed signals, error rates, data patterns, and the like. This determined deviation is used in step 170 as the basis for comparing the deviation with a threshold value 58. This threshold value is to be considered the input variable for this step 170. If this threshold value 58 is undershot, the monitoring unit 44 transmits real measurement data 34 to a higher-level control unit 42 in step 180.

In other words, this is the case where the physical sensor 11 functions correctly and delivers plausible measurement data 34. If the result of the comparison in step 170 is that the calculated deviation reaches or exceeds the threshold value 58, the monitoring The control unit 44 transmits model measurement data 40 of the measured variable to the control unit 42. This applies to the case where the physical sensor 11 has either failed or is malfunctioning, and in this case, model measurement data 40 is transmitted as a substitute. Consequently, regardless of whether the physical sensor is functioning correctly or incorrectly, a higher-level control unit 42 receives the necessary measurement data.

In one example, the calculation 160 of the deviation is performed continuously in the sense of automated monitoring and follows the principle of a state machine. In other words, the monitoring unit continuously calculates and repeats the deviations between model measurement data 40 and real measurement data 34. This is indicated by the dashed line in 3 which, after transmitting 180 the real measurement data 34 or model measurement data 40, recalculates 160 the deviation. Additionally, step 200 describes the case where the monitoring unit 44 detects a model error. This means that the cause lies, among other things, in the preference model 54 itself. Therefore, in step 200, the preference model 54 is retrained by the simulation unit 56, with the retraining 200 being performed in such a way that a deviation between the model measurement data 40 of the measured variable and the real measurement data of the measured variable is reduced.

4 1 schematically shows a system 41 for providing measurement data 34, 40 with a simulation unit 56, a monitoring unit 44, and a control unit 42 according to the invention. This simplified representation is intended to explain the essential technical components and their interaction. A control unit 42 can be understood as a higher-level instance, for example, for monitoring and controlling a power transformer 10. In order to perform this control and monitoring function, the control unit 42 requires sensor data from the relevant equipment 10 in order to derive decisions based thereon. Thus, the control unit 42 requests 110 measurement data 36, 40 from the monitoring unit 44. A physical sensor 11 supplies real measurement data 34 to the monitoring unit 44 and to a simulation unit 56.

The simulation unit 56 manages and trains the data models 50, which are stored in a model database 46. To train the data models 50, the simulation unit 56 is provided with real measurement data 34 from a physical sensor 11, historical real measurement data 49, and reference measurement data 48 of reference measurement variables. The simulation unit 56 is designed to adapt the parameters of the data models 50 through training such that a model error is reduced during training. The simulation unit 56 is configured to select a preference model 54 from the totality of the optionally pre-trained data models 50. At the same time, the simulation unit 56 is configured to calculate model measurement values 40 based on a selected data model 50, for example, a preference model 54.

These model measured values 40 are made available to the monitoring unit 44, which then calculates 160 the deviation between the real measured data 34 and the model measured data 40. If necessary, the monitoring unit 44 can send a request for retraining to the simulation unit 56, analogous to step 200 of the data model 50. Depending on the degree of deviation with respect to a threshold value 58, which can be understood as an input variable, the monitoring unit 44 returns real measured data 34 or model measured data 40 to the control unit 42.

In 5 A further example of a system 41 for providing measurement data 34, 40 is shown. The blocks shown are to be understood as exemplary process steps of a method according to the invention. In step A, a first initial pre-training 130 of a plurality of data models 50 takes place. In step B, the best data model 50 (preference model 54) is selected 140 together with a further training step 150, which results in a reduced model error a). In step C, condition monitoring or anomaly detection takes place such that, for example, a deviation occurs between real measurement data 34 and model measurement data 40. This condition monitoring in step C can mean, for example, that real measurement data 34 of a physical sensor 11 has a high degree of noise c).

In step D, the monitoring unit 44 ensures that, as an alternative, corresponding model measurement values 40 with low signal-physical cleanness and low noise b) are generated. In a further scenario, a sensor error d) is detected by the condition monitoring, i.e., the monitoring unit 44. Here, too, corresponding model measurement data 40 are calculated as an alternative using the associated preference model 54, and the sensor 38 is recognized by the higher-level system as repaired or functional e). In a further case, the anomaly detection or condition monitoring detects in step C that the preference model 54 has an increasingly large model error i). Here, in step F, retraining 200 of the preference model 56 is initiated, resulting in a reduced model error h).

A combination of model error and noisy signal is also shown in step E, where, on the one hand, substitute model measurement data 40 is generated via the preference model 54, but also a retraining 200 of the preference model 54 takes place. Generally speaking, an abstraction layer is provided around each physical sensor group. This abstraction layer switches between the various states of a sensor 11 in defined but flexible sequences. This abstraction layer can basically be used for any power transformer 10. The advantage of this system is its independence from specific sensors 11, since these can be mapped using models. At the same time, physical sensors 11 can be replaced by data models 50, which can reduce costs and complexity. At the same time, data that is normally only available offline can be included in the modeling and provide additional measurement data 40.

In 6 As an example, a temporal progression of real measurement data 34 and model measurement data 40 during training, i.e., pre-training 130, training 150, or post-training 200, of a data model 50 is shown. In the specific example, the hotspot temperature and the oil temperature are shown in the upper part, each with the real measurement data 34 and the associated model measurement data 40. The electrical load 60 and the ambient temperature 62 are plotted in the lower part as input variables for the model 50. It can be seen that a model error decreases over time and the real measurement data 34 and the model measurement data 40 become more similar in their values. The data model 50 therefore learns from the real measurement data 34 over time by adjusting the parameters of the data model 50 until the deviations are acceptably small.

In one example (not shown), the preference model 54 is an IEEE Std C57-91-2011 Annex G model. The reference measured variables are the hotspot temperature, top oil temperature, and ambient temperature 62, as well as the electrical load 60 of the power transformer 10. The state machine or monitoring unit 44, in conjunction with the simulation unit 56, uses the calculated deviation between real measurement data 34 and model measurement data 40 to determine whether a cooling device of a power transformer 11 is switched on or off, or has failed. In one example, the deviation 160 is calculated by the monitoring unit 44 using an artificial neural network. This has the advantage that nonlinear decision thresholds and threshold values 58 can be implemented.

Additionally, it should be noted that "comprising" does not exclude other elements or steps, and "one" or "an" does not exclude a plurality. Furthermore, it should be noted that features or steps described with reference to one of the above embodiments may also be used in combination with other features or steps of other embodiments described above. Reference signs in the claims are not to be considered as limitations.

List of reference symbols

100

Procedure for the continuous provision of measurement data

110

Requesting measurement data

120

Providing a variety of data models

130

Pre-training the data models

140

Selecting a preference model

150

Training the preference model

160

Calculating a deviation

170

Comparing the deviation with a threshold

180

Transmitting the model measurement data to the control unit

190

Transmitting real measurement data to the control unit

200

Retraining the preference model

10

Power transformers Power transformer

11

physical sensor

12

Temperature sensor

14

pressure sensor

16

flow sensor

18

Torque sensor

20

Vibration sensor

22

Level sensor

24

Humidity sensor

26

Gas sensor

28

Capacitance measuring sensor

30

Sensor for electrical quantities

32

Expansion tank

34

Real measurement data

36

Input variables

38

virtual sensor (model)

40

Model measurement data

41

System for providing measurement data

42

Control unit

44

monitoring unit

46

Model database

48

Reference measurement data

49

Real measurement data, historical

50

Data model

52

Training data

54

Preference model

56

Simulation unit

58

Threshold

60

electrical load

62

Ambient temperature

QUOTES CONTAINED IN THE DESCRIPTION

This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Cited patent literature

• US 2017011612A1 [0003]

Claims (16)

A method (100) for continuously providing measurement data (40, 34) relating to a measured variable of a power transformer (10), comprising the steps of: - requesting (110) measurement data (36, 40) of a measured variable by a control unit (42) of the power transformer (10); - providing (120) a plurality of data models (50) in a model database (46), wherein the data models (50) are suitable for calculating model measurement data (40) of a measured variable as a function of reference measurement data (48) of reference measured variables; - selecting (140) a preference model (54) from the plurality of provided models (50) by a simulation unit (56), wherein the selection (140) is based on the smallest possible deviation of the calculated model measurement data (40) of the preference model (54) from the associated real measurement data (34) of this measured variable; - Training (150) the preference model (54) with training data (52) by the simulation unit (56), wherein the training data (52) comprises reference measurement data (48) of reference measurement variables and/or real measurement data (34) of the measurement variable; wherein the training (150) of the preference model (54) is carried out in such a way that parameters of the preference model (54) are adjusted such that a deviation of the model measurement data (40) calculated by the preference model (54) from the real measurement data (34, 49) of the training data is minimized; - Calculating (160), with a monitoring unit (44), a deviation between the model measurement data (40) of the measurement variable calculated by the preference model (54) and real measurement data (34) of the measurement variable; - transmitting (180) the model measurement data (40) of the measured variable by the monitoring unit (44) to the control unit (42) of the transformer (10) when the deviation reaches or exceeds a defined threshold value (58), or transmitting (190) the real measurement data (34) to the control unit (42) of the power transformer (10) when the calculated deviation falls below the threshold value (58). Procedure (100) according to Claim 1 , further comprising the step: - retraining (200) the preference model (54) by the simulation unit (56) when the deviation between real measurement data (34) of the measured variable and model measurement data (40) reaches or exceeds a defined threshold value (58); wherein the retraining (200) is carried out in such a way that a deviation between model measurement data (40) of the measured variable and real measurement data (34) of the measured variable is reduced. Method (100) according to one of the preceding claims, wherein the training (150, 200) of the preference model (54) and the adaptation of the parameters of the preference model (54) are repeated after defined time intervals. Method (100) according to one of the preceding claims, wherein the providing (120) of the data models (50) comprises an IEC 60076-7 model, ETF model (Estimate Transfer Function), a model based on linear regression and/or an IEEE Std C57-91-2011 Annex G model. The method (100) according to one of the preceding claims, wherein the preference model (54) is an IEC-60076-7 model, and the simulation unit (56) calculates the parameters of the preference model (54) by solving the differential equation of the preference model (54), wherein the parameters are determined while minimizing model error, penalty term (short-term change), and penalty term (initial value change); wherein the determination of the parameters is based on historical or current measurement data (34, 49) of ambient temperature (62), maximum winding temperature, top oil temperature, and electrical load (60). Method (100) according to one of the preceding claims, wherein the measured variable is a hotspot temperature of a power transformer (10) and the reference measured variable is an average oil temperature, a top oil temperature, an electrical load of the power transformer and/or an ambient temperature (62). Method (100) according to one of the preceding claims, wherein the measured variable is a degree of polymerization of the paper insulation of a power transformer (10). Method (100) according to one of the preceding claims, wherein the measured variable relates to a gas analysis (DGA) of a power transformer (10). Method (100) according to one of the preceding claims, wherein, prior to the selection of the preference model (54), a - pre-training (130) of a plurality of data models (50) is carried out based on training data (52), wherein the training data (52) comprises reference measurement data (48) of reference measurement variables and/or real measurement data (34) of the measurement variable; wherein the pre-training (130) of the data models is carried out such that parameters of the data models (50) are adapted such that a deviation of the model measurement data (40) from the real measurement data (34) of the training data (52) is reduced; Method (100) according to one of the preceding claims, wherein the deviation of the model measurement data (40) of the preference model (54) from the associated real measurement data (34) of this measurement variable describes a noise component in the real measurement data (34). Method (100) according to one of the preceding claims, wherein the deviation of the model measurement data (40) of the preference model (54) from the associated real measurement data (34) of this measurement variable describes an amplitude, measurement data pattern and/or signal shape of the real measurement data (34). Method (100) according to one of the preceding claims, wherein the real measurement data (34) are sensor-detected measurement data. Method (100) according to one of the preceding claims, wherein the real measurement data (34) are measurement data determined offline in a laboratory. A computer program comprising instructions which, when executed by a processor, cause the processor to carry out the computer-implemented method (100) according to one of the Claims 1 until 13 to execute. A computer-readable medium comprising instructions which, when executed by a processor, cause the processor to perform the computer-implemented method (100) according to one of the Claims 1 until 13 to execute. A system (41) for providing measurement data (34, 40), comprising a monitoring unit (44) configured to calculate a deviation between the model measurement data (40) of the measured variable calculated by the preference model (54) and real measurement data (34) of the measured variable and to transmit model measurement data (40) and/or real measurement data (34) to a control unit of the power transformer (10); a control unit (42) for monitoring a power transformer (10), configured to request measurement data (34, 40) of a measured variable of the power transformer (10) from a monitoring unit (44) and to receive real measurement data (34) and/or model measurement data (40) of the measured variable from the monitoring unit; a simulation unit (56) configured to select and train a preference model (54) from the plurality of provided data models (50) and to calculate corresponding model measurement data (40) on the basis of data models (50); a model database (46) configured to store data models (50) of a measured variable; wherein the system (41) is designed to carry out the method (100) according to one of the Claims 1 until 13 to execute.