Method, device and computer readable storage medium for determining wind power of wind power station

Abstract

This specification provides a method, device and computer-readable storage medium for determining the wind power of a wind power station. Based on this method, before specific implementation, a first prediction model based on RNN can be obtained and trained based on alternative features; then, according to a preset processing rule, the shap value of the alternative feature is determined using the first prediction model; according to the shap value, the target feature is screened out from the alternative features; then, based on the target feature training, a second prediction model based on TCN is obtained to support the realization of medium- and short-term predictions of wind power at multiple moments in a continuous time period. During specific implementation, the data value of the target feature of the target wind power station at the current target moment can be obtained first; then, the wind power of the target wind power station at the current target moment is determined by processing the data value of the target feature using the second prediction model. Thereby, the wind power of the target wind power station at multiple moments in the current continuous time period can be determined more efficiently and accurately.

Description

Technical Field

[0001] The present invention relates to the field of new energy technology, and in particular to a method, device and computer-readable storage medium for determining wind power of a wind power station. Background Art

[0002] With the advancement and development of new energy technologies, wind power generation has begun to attract more and more attention and attention.

[0003] However, due to the instability and intermittency of wind energy itself, the wind power output of wind power stations fluctuates relatively greatly, and it is often difficult to accurately and real-time predict the wind power of wind power stations in advance, which in turn affects the subsequent wind power station grid distribution, power station energy storage, market adjustments, and related decision-making.

[0004] To address the above problems, no effective solution has been proposed yet. Summary of the invention

[0005] The present specification provides a method, device and computer-readable storage medium for determining the wind power of a wind power station. By utilizing a first prediction model based on RNN, a second prediction model based on TCN that supports medium- and short-term predictions of wind power at multiple moments in a continuous time period with better results can be constructed and trained. The second prediction model can then be used to efficiently and accurately determine the wind power of the target wind power station at multiple moments in the current continuous time period.

[0006] This specification provides a method for determining wind power of a wind power station, including:

[0007] Obtain data values ​​of target characteristics of the target wind power station at the current target time;

[0008] Processing the data value of the target feature using the second prediction model to obtain a corresponding target prediction result; wherein the second prediction model is a TCN model trained using the second type of sample data about the target feature and the first prediction model according to a preset training rule, and the first prediction model is an RNN model trained using the first type of sample data about the candidate feature;

[0009] According to the target prediction results, the wind power of the target wind power station at the current target moment is determined.

[0010] In one embodiment, the method further comprises:

[0011] According to the first screening rule, matching features are screened out from the feature set about the wind power station as candidate features;

[0012] According to the preset training rules, the candidate features are used to construct corresponding reconstruction features; and an initial first prediction model based on RNN is constructed; wherein the initial first prediction model at least includes an initial first input layer, an initial RNN hidden layer, and an initial first output layer;

[0013] According to the reconstructed features and the sample data set, a first type of sample data about the candidate features that matches is constructed;

[0014] Using the first type of sample data to train an initial first prediction model to obtain a first prediction model that meets the requirements;

[0015] According to the preset processing rules, the shap value of the candidate feature is determined by using the first prediction model;

[0016] According to the second screening rule and the shap values ​​of the candidate features, a target feature meeting the requirements is screened out from the candidate features;

[0017] According to the preset training rules, using the target features and the sample data set, a second type of sample data matching the target features is constructed; and an initial second prediction model based on TCN is constructed; wherein the initial second prediction model at least includes an initial second input layer, an initial TCN layer, and an initial second output layer;

[0018] The second type of sample data is used to train the initial second prediction model to obtain a second prediction model that meets the requirements.

[0019] In one embodiment, according to the first screening rule, matching features are screened out from the feature set related to the wind power station as candidate features, including:

[0020] Obtain environmental parameters of the target wind power station in the current time period;

[0021] According to the first screening rule and the environmental parameters of the target wind power station in the current time period, matching features are screened out from the feature set related to the wind power station as candidate features.

[0022] In one embodiment, the alternative features include: 1-period lag data of wind power, 2-period lag data of wind power, 3-period lag data of wind power, 96-period lag data of wind power, 97-period lag data of wind power, 98-period lag data of wind power, 35040-period lag data of wind power, 35041-period lag data of wind power, 35042-period lag data of wind power, current moment data of wind speed, 1-period lag data of wind speed, 2-period lag data of wind speed, 3-period lag data of wind speed, 96-period lag data of wind speed, 97-period lag data of wind speed, 98-period lag data of wind speed, 35040-period lag data of wind speed, 35041-period lag data of wind speed, 35042-period lag data of wind speed, current moment data of temperature, and current moment data of humidity.

[0023] In one embodiment, according to a preset processing rule, the shap value of the candidate feature is determined using the first prediction model, including:

[0024] According to the preset processing rules, the sample data set is used to construct test data about the candidate features;

[0025] Modifying at least one candidate feature in the test data to obtain modified data; and combining the test data and the modified data to establish a test data set;

[0026] Processing the test data set using the first prediction model to obtain corresponding prediction results for multiple types of situations;

[0027] According to the prediction results of various types of situations, the shap values ​​of the candidate features are determined.

[0028] In one embodiment, the shap value of the candidate feature is determined according to the following formula:

[0029]

[0030] in, is the shap value of the candidate feature numbered i, N represents the set of candidate features in the test data set, the dimension of N is M, S represents a subset of the candidate feature numbered i extracted based on N, and the dimension of S is |S|, f x (S) represents the average value of the prediction results of the corresponding situation type obtained by processing the test data about S in the test data set using the first prediction model, It represents the average value of the prediction results of the corresponding situation type obtained by processing the test data about S and the union of the candidate features numbered i in the test data set using the first prediction model.

[0031] In one embodiment, after the initial second prediction model is trained using the second type of sample data to obtain a second prediction model that meets the requirements, the method further includes:

[0032] Obtain data values ​​of target characteristics of the target wind power station at multiple moments in the previous target time period; query the power generation record of the target wind power station to obtain measured values ​​of wind power of the target wind power station at multiple moments in the previous target time period;

[0033] Determine the predicted values ​​of wind power of the target wind power station at multiple moments in the previous target time period by processing the data values ​​of the target characteristics of the target wind power station at multiple moments in the previous target time period using the second prediction model;

[0034] Calculating the model accuracy of the second prediction model according to the measured values ​​of the wind power of the target wind power station at multiple moments in the previous target time period and the predicted values ​​of the wind power of the target wind power station at multiple moments in the previous target time period;

[0035] Check whether the model accuracy is greater than or equal to the preset accuracy threshold;

[0036] When it is determined that the model accuracy is less than a preset accuracy threshold, the second prediction model is adjusted and updated according to a preset adjustment rule.

[0037] This specification also provides a device for determining wind power of a wind power station, comprising:

[0038] An acquisition module, used to acquire data values ​​of target characteristics of a target wind power station at a current target moment;

[0039] A processing module, used to process the data value of the target feature using the second prediction model to obtain a corresponding target prediction result; wherein the second prediction model is a TCN model trained using the second type of sample data about the target feature and the first prediction model according to a preset training rule, and the first prediction model is an RNN model trained using the first type of sample data about the candidate feature;

[0040] The determination module is used to determine the wind power of the target wind power station at the current target moment according to the target prediction result.

[0041] This specification also provides a data processing method, including:

[0042] According to the first screening rule, matching features are screened out from the feature set about the wind power station as candidate features;

[0043] According to the preset training rules, the candidate features are used to construct corresponding reconstruction features; and an initial first prediction model based on RNN is constructed; wherein the initial first prediction model at least includes an initial first input layer, an initial RNN hidden layer, and an initial first output layer;

[0044] According to the reconstructed features and the sample data set, a first type of sample data about the candidate features that matches is constructed;

[0045] Using the first type of sample data to train an initial first prediction model to obtain a first prediction model that meets the requirements;

[0046] According to the preset processing rules, the shap value of the candidate feature is determined by using the first prediction model;

[0047] According to the second screening rule and the shap values ​​of the candidate features, a target feature meeting the requirements is screened out from the candidate features;

[0048] According to the preset training rules, using the target features and the sample data set, a second type of sample data matching the target features is constructed; and an initial second prediction model based on TCN is constructed; wherein the initial second prediction model at least includes an initial second input layer, an initial TCN layer, and an initial second output layer;

[0049] The second type of sample data is used to train an initial second prediction model to obtain a second prediction model that meets the requirements; wherein the second prediction model is used to predict the wind power of the wind power station.

[0050] The present specification also provides a computer-readable storage medium on which computer instructions are stored. When the instructions are executed by a processor, the relevant steps of the method for determining the wind power of a wind power station are implemented.

[0051] Based on the method, device and computer-readable storage medium for determining wind power in a wind power station provided in this specification, before specific implementation, matching alternative features can be screened out first, and corresponding reconstruction features can be constructed; then, the reconstruction features and the sample data set can be used to construct a first type of sample data about the alternative features; the first type of sample data can be used to train an initial first prediction model based on RNN, and a first prediction model that meets the requirements can be quickly trained; then, according to a preset processing rule, the first prediction model is used to determine the shap value of the alternative feature; and according to a second screening rule and the shap value of the alternative feature, standard features with better effects are screened out from the alternative features; then, according to a preset training rule, the target feature and the sample data set are used to construct a matching second type of sample data about the target feature; and an initial second prediction model based on TCN is constructed; the second type of sample data is used to train the initial second prediction model to obtain a second prediction model that supports the realization of medium- and short-term predictions of wind power at multiple moments in a continuous time period. In specific implementation, the data value of the target characteristic of the target wind power station at the current target moment can be obtained first; then the wind power of the target wind power station at the current target moment can be determined by processing the data value of the target characteristic using the second prediction model. This can better adapt to the wind power station supervision scenario, and can more efficiently and accurately determine the specific wind power of the target wind power station at multiple moments in a continuous time period, so as to provide valuable reference for relevant decisions of the target wind power station, and thus effectively reduce decision-making errors, ensure the safety of the power distribution network of the target wind power station, and also help to improve the market benefits of the target wind power station. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of this specification, the drawings required for use in the embodiments will be briefly introduced below. The drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0053] Figure 1 It is a flowchart of a method for determining wind power of a wind power station provided by an embodiment of this specification;

[0054] Figure 2 It is a schematic diagram of an embodiment of a method for determining wind power of a wind power station provided by an embodiment of this specification, in a scenario example;

[0055] Figure 3 It is a schematic diagram of an embodiment of a method for determining wind power of a wind power station provided by an embodiment of this specification, in a scenario example;

[0056] Figure 4It is a schematic diagram of an embodiment of a method for determining wind power of a wind power station provided by an embodiment of this specification, in a scenario example;

[0057] Figure 5 It is a schematic diagram of an embodiment of a method for determining wind power of a wind power station provided by an embodiment of this specification, in a scenario example;

[0058] Figure 6 It is a schematic diagram of the structure of a server provided by an embodiment of this specification;

[0059] Figure 7 It is a schematic diagram of the structure of a device for determining wind power of a wind power station provided by an embodiment of this specification;

[0060] Figure 8 It is a schematic diagram of an embodiment of a method for determining wind power of a wind power station provided by an embodiment of this specification, in a scenario example;

[0061] Fig. 9 It is a schematic diagram of an embodiment of a method for determining wind power of a wind power station provided by an embodiment of this specification, in a scenario example. DETAILED DESCRIPTION

[0062] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0063] It should be noted that the user-related information and data involved in the embodiments of this specification are all information and data authorized by the user or fully authorized by relevant parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data comply with relevant laws, regulations and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entrances for users or relevant parties to choose to authorize or refuse.

[0064] It should also be noted that in the embodiments of this specification, certain software, components, models and other existing solutions in the industry may be mentioned. They should be regarded as exemplary. Their purpose is only to illustrate the feasibility of implementing the technical solution of this application, but it does not mean that the applicant has or will necessarily use the solution.

[0065] See also Figure 1As shown, the embodiment of this specification provides a method for determining the wind power of a wind power station. The method may include the following contents when it is specifically implemented:

[0066] S101: Obtaining data values ​​of target characteristics of a target wind power station at a current target moment;

[0067] S102: Processing the data value of the target feature using the second prediction model to obtain a corresponding target prediction result; wherein the second prediction model is a TCN model trained using the second type of sample data about the target feature and the first prediction model according to a preset training rule, and the first prediction model is an RNN model trained using the first type of sample data about the candidate feature;

[0068] S103: Determine the wind power at multiple moments in the current target moment of the target wind power station according to the target prediction result.

[0069] Among them, the above-mentioned second prediction model can be specifically understood as a neural network model that is pre-trained according to preset training rules and comprehensively uses the RNN model, TCN model, and shap value technology. It is suitable for wind power station supervision scenarios and can predict the wind power at the corresponding moment based on the input data values ​​of multiple features.

[0070] The above-mentioned RNN model (Recurrent Neural Network) may specifically refer to a recursive neural network that takes sequence data as input, performs recursion in the direction of sequence evolution, and all nodes (recurrent units) are connected in a chain.

[0071] The above-mentioned TCN model (Temporal Convolutional Networks) may specifically refer to an algorithm model used to solve time series prediction.

[0072] The above-mentioned shap value technology (Shapley Additive exPlanations) can specifically refer to a method of explaining individual predictions, which can be used as a model explanation.

[0073] In this embodiment, first of all, considering that the wind power at a certain moment is often not only associated with the relevant features at that moment, but also associated with the relevant features at other moments before that moment, that is, the characteristic factors that affect the actual wind power are time-dependent. Based on the above considerations, an RNN model with a relatively fast model training speed can be introduced and used to learn and capture the time dependency in the input data based on the sequence form of the candidate features, and then through the state update and transmission mechanism inside the RNN, save and accumulate historical information, and then combine the above information for abstraction and analysis to achieve the prediction of wind power at a certain moment, and obtain a first prediction model that can preliminarily predict wind power based on the temporal nature of the input features.

[0074] Next, considering that the number of alternative features that need to be obtained and used when predicting wind power based on the above-mentioned first prediction model is large, on the one hand, the processing burden of the model is increased, affecting the overall operating efficiency of the model; on the other hand, the alternative features used will also include features that have no practical effect on the predicted wind power, and may even bring about error effects and other adverse effects, affecting the accuracy of the final output prediction results. Based on the above considerations, the shap value technology can be introduced, and at the same time, combined with the first prediction model based on RNN obtained by training based on alternative features, the contribution of each alternative feature used in the first prediction model in predicting wind power based on the time dependence of the feature can be quantitatively and objectively evaluated, so as to accurately select the alternative features that really have a good positive effect on wind power prediction from a large number of alternative features with a wide variety of types, and at the same time eliminate the alternative features that will bring negative effects, and determine the target features with better prediction effect and smaller error.

[0075] Finally, in actual wind power plant management scenarios, it is often required to accurately predict wind power at multiple moments in a continuous time period (for example, wind power every 15 minutes every day) in order to timely adjust the power distribution network and related market prices, which places high demands on the model's medium- and short-term prediction effect. Therefore, a TCN model is also introduced to replace the previously used RNN model, and the model is trained accordingly using data based on target features, so that the model can take advantage of the performance advantages of the TCN model and effectively learn to capture patterns and long-term dependencies at different scales in the sequence, so as to better model relatively longer-term time correlations, and train to obtain a second prediction model that can more accurately predict wind power at each moment in a medium- to short-term continuous time period, so as to meet the scenario requirements of wind power plant supervision scenarios.

[0076] The above-mentioned current target time can be specifically understood as a certain time of the wind power to be predicted, for example, it can be 10:00 of the current day, or it can be 9:00 of the previous day.

[0077] In specific implementation, the wind speed, temperature, and humidity of the target wind power station at the current target time can be collected; at the same time, the historical monitoring records of the target wind power station are queried, and with the current target time as the reference point, the 1st period lag data of wind power, the 2nd period lag data of wind power, the 3rd period lag data of wind power, the 96th period lag data of wind power, the 97th period lag data of wind power, the 98th period lag data of wind power, the 35040th period lag data of wind power, and the 97th period lag data of wind power are queried. The data of the last 35041 periods, the data of wind power lagged 35042 periods, the data of wind speed lagged 1 period, the data of wind speed lagged 2 periods, the data of wind speed lagged 3 periods, the data of wind speed lagged 96 periods, the data of wind speed lagged 97 periods, the data of wind speed lagged 98 periods, the data of wind speed lagged 35040 periods, the data of wind speed lagged 35041 periods, and the data of wind speed lagged 35042 periods; and then combine the above data to obtain the data value of the target characteristic of the target wind power station at the current target moment.

[0078] Among them, lag 1 can be understood as the moment 15 minutes before the current target moment, lag 2 can be understood as the moment 30 minutes before the current target moment, and so on. Lag 96 specifically refers to the same moment of the previous day with respect to the current target moment. Lag 3540 specifically refers to the same moment of the same day of the previous year with respect to the current target moment. This is because through the collation and analysis of historical data and model testing and screening, it is found that the wind power at a certain moment is often closely related to the moment and the relevant factors of the above historical moments about the moment. Therefore, by obtaining and using the data value of the target feature of the above current target moment, the wind power of the target wind power station at the current target moment can be predicted more accurately.

[0079] Among them, the wind speed, temperature, and humidity at the current target moment can be simply recorded as basic features, and the lagged data such as the three-period lagged data of wind power can be simply recorded as lagged features about the basic features.

[0080] In specific implementation, the data value of the target feature at the current target moment can be converted into a corresponding sequence form; then the data value of the target feature at the current target moment in sequence form is used as a model input, input into the second prediction model, and the model is run to obtain the corresponding model output as the target prediction result; finally, the wind power of the target wind power station at the current target moment can be determined based on the target prediction result.

[0081] After obtaining the data value of the target feature of the target wind power station at the current target moment, the integrity of the data value of the target feature can also be detected to determine whether there is a missing value of the target feature. When it is detected that the data value of the target feature is missing, the target feature can be marked as a missing feature; then the feature data value at the same time of the previous target time period corresponding to the missing feature is queried and obtained as the data value of the missing feature to fill and repair the data value of the target feature, so as to obtain a relatively complete and effective data value of the target feature at the current target moment.

[0082] Based on the above embodiments, it can be better adapted to the wind power station supervision scenario, and a second prediction model can be obtained by training multiple technologies such as the joint RNN model, TCN model, shap value technology, etc., to accurately predict the wind power of the target wind power station at the current target moment, effectively reducing the prediction error.

[0083] In some embodiments, different natural factors may significantly change the wind power of a wind power station in different seasons. For example, in spring and summer, there is more rain and the environmental humidity changes greatly, and the effect of humidity on the change of wind power will be relatively more significant. At this time, when predicting the wind power at the current target moment, in addition to collecting and using the current target moment data of humidity, you can further collect and use the lag 1 data of humidity, the lag 2 data of humidity, and the lag 35040 data of humidity, so as to more fully and comprehensively consider the impact of humidity factors on the wind power at the current target moment, and more accurately predict the wind power at the current target moment.

[0084] For example, in autumn and winter, there is less rain, the air is relatively dry for a long time, the environmental humidity is relatively stable, and the impact of humidity on wind power changes is relatively small. At this time, when predicting the wind power at the current target time, it is only necessary to collect and use the current target time data of humidity, and there is no need to collect and use other historical lagged data of humidity. This can effectively reduce the overall data processing and more efficiently and accurately predict the wind power at the current target time.

[0085] For example, in autumn and winter, since wind speed and wind direction change frequently and have poor stability, the lagged data on wind speed and wind power can be further added to more fully and comprehensively consider the specific changes in wind speed and wind direction, and more accurately predict the wind power at the current target moment.

[0086] Based on the above embodiment, the above method of obtaining the data value of the target characteristic of the target wind power station at the current target moment may include: obtaining the current season information; querying the preset characteristic template set according to the current season information, and determining the preset characteristic template matching the current season as the target characteristic template; determining the target characteristic at the current target moment according to the target characteristic template, and obtaining the data value of the target characteristic of the target wind power station at the current target moment.

[0087] The preset feature template set includes a plurality of preset feature templates, each preset feature template corresponds to at least one season, and each preset feature template includes at least target features that match the corresponding season.

[0088] The above-mentioned preset feature template set can be specifically constructed in the following manner: in different seasons, adjust the features and use the second prediction model to conduct experimental tests, and collect corresponding experimental test data; then perform clustering processing on the experimental test data based on seasonal information to determine the common features that have better effects for each season and combine them to obtain preset feature templates for each season, thereby constructing a preset feature template set.

[0089] In some embodiments, the wind power of the target wind power station at multiple moments in the current target time period can be predicted in the above manner, for example, the wind power at each moment every 15 minutes on the day; and then, the target wind power station can be matched with the power grid distribution operation according to the wind power of the target wind power station at multiple moments in the current target time period, so that the wind power generated by the target wind power station can be safely and accurately fed into the power grid. In addition, the total electric energy generated by the target wind power station in the current target time period can be estimated according to the wind power of the target wind power station at multiple moments in the current target time period; and based on the preset protocol rules, the reward data of the electric energy generated in the current time period can be adjusted according to the total electric energy generated in the current target time period; and, in combination with the reward data of the electric energy generated in the current time period, the revenue data of the target wind power station can be estimated; and decisions on the expansion and operation of the target wind farm can be made according to the revenue data of the target wind power station.

[0090] In some embodiments, see Figure 2 As shown, when the method is implemented specifically, it may also include the following contents:

[0091] S1: According to the first screening rule, select matching features from the feature set about the wind power station as candidate features;

[0092] S2: According to the preset training rules, using the candidate features, constructing corresponding reconstruction features; and constructing an initial first prediction model based on RNN; wherein the initial first prediction model at least includes an initial first input layer, an initial RNN hidden layer, and an initial first output layer;

[0093] S3: constructing the first type of sample data matching the candidate features according to the reconstructed features and the sample data set;

[0094] S4: using the first type of sample data to train an initial first prediction model to obtain a first prediction model that meets the requirements;

[0095] S5: Determine the shap value of the candidate feature using the first prediction model according to the preset processing rule;

[0096] S6: According to the second screening rule and the shap values ​​of the candidate features, a target feature meeting the requirements is screened out from the candidate features;

[0097] S7: According to the preset training rules, using the target features and the sample data set, construct a second type of sample data matching the target features; and construct an initial second prediction model based on TCN; wherein the initial second prediction model at least includes an initial second input layer, an initial TCN layer, and an initial second output layer;

[0098] S8: Use the second type of sample data to train the initial second prediction model to obtain a second prediction model that meets the requirements.

[0099] The shap value can be specifically understood as an indicator parameter based on the shap value technology, which quantitatively evaluates the specific contribution of the feature to the model prediction result by explaining the prediction result.

[0100] Based on the above embodiments, according to the preset training rules, multiple technologies such as RNN model, TCN model, shap value technology, etc. can be used in combination to efficiently construct and train a second prediction model that is suitable for wind power station supervision scenarios and can perform continuous medium- and short-term wind power predictions more accurately.

[0101] In some embodiments, the above-mentioned filtering out matching features from the feature set about the wind power station as candidate features according to the first filtering rule may include the following contents during specific implementation:

[0102] S1: Obtain the environmental parameters of the target wind power station in the current time period;

[0103] S2: According to the first screening rule and the environmental parameters of the target wind power station in the current time period, matching features are screened out from the feature set related to the wind power station as candidate features.

[0104] The above environmental parameters may specifically include one or more of the following: terrain parameters, longitude and latitude parameters, height information of surrounding objects, and the like.

[0105] In specific implementation, seasonal information of the current time period can also be obtained, and then matching features can be screened out as candidate features from the feature set of the wind power station based on the first screening rule, the environmental parameters of the target wind power station in the current time period, and the seasonal information of the current time period.

[0106] Based on the above embodiments, matching candidate features with better effects can be determined more accurately and efficiently.

[0107] In some embodiments, the alternative features may specifically include: 1-period lag data of wind power, 2-period lag data of wind power, 3-period lag data of wind power, 96-period lag data of wind power, 97-period lag data of wind power, 98-period lag data of wind power, 35040-period lag data of wind power, 35041-period lag data of wind power, 35042-period lag data of wind power, current moment data of wind speed, 1-period lag data of wind speed, 2-period lag data of wind speed, 3-period lag data of wind speed, 96-period lag data of wind speed, 97-period lag data of wind speed, 98-period lag data of wind speed, 35040-period lag data of wind speed, 35041-period lag data of wind speed, 35042-period lag data of wind speed, current moment data of temperature, current moment data of humidity, etc.

[0108] Of course, it should be noted that the above-listed candidate features are only for illustrative purposes. In specific implementation, other feature combinations may be used as candidate features according to specific circumstances and processing requirements.

[0109] In some embodiments, during specific implementation, it is also possible to first select matching basic features from a feature set about a wind power station according to environmental parameters and / or seasonal information, as well as more important and commonly used lag features about the basic features as initial features; then perform correlation analysis based on the initial features, and determine alternative features based on the results of the correlation analysis.

[0110] In specific implementation, multiple lag feature sequences about multiple basic features and corresponding wind power sequences can be constructed based on the initial features by querying historical monitoring records or sample data sets; then, corresponding correlation analysis can be performed based on the multiple lag features and the corresponding wind power sequences, and multiple correlation analysis results corresponding to each lag feature can be calculated; wherein the correlation analysis results at least include: Spearman rank correlation coefficient.

[0111] Then, based on the correlation analysis results, the correlation degree between the basic characteristics associated with different lag characteristics and wind power can be distinguished; and based on the correlation degree, different basic characteristics and lag characteristics related to the basic characteristics are treated differently to determine the alternative characteristics.

[0112] The Spearman's rank correlation coefficient can be a non-parametric correlation measure that can be used to evaluate the degree of dependence between two variables without requiring the data to follow a normal distribution. Specifically, the data can be calculated based on the ranks (i.e., the positions after sorting) of the two variables.

[0113] Specifically, according to the results of the correlation analysis, when it is determined that the Spearman rank correlation coefficient is greater than or equal to 0.6, it can be judged that the corresponding basic feature and the lag feature about the basic feature are strongly correlated with the wind power. At this time, more lag features can be further added on the basis of the current lag feature. When it is determined that the Spearman rank correlation coefficient is greater than or equal to 0.4 and less than 0.6, it can be judged that the corresponding basic feature and the lag feature about the basic feature are moderately correlated with the wind power. At this time, the basic feature can be retained and the current lag feature is not changed. When it is determined that the Spearman rank correlation coefficient is greater than or equal to 0.2 and less than 0.4, it can be judged that the corresponding basic feature and the lag feature about the basic feature are weakly correlated with the wind power. At this time, the basic feature can be retained and some lag features can be deleted. When it is determined that the Spearman rank correlation coefficient is less than 0.2, it can be judged that the corresponding basic feature and the lag feature about the basic feature are not correlated with the wind power. At this time, the basic feature and the corresponding lag feature can be deleted.

[0114] In some embodiments, when constructing the reconstruction feature, considering that the input data of the RNN model is usually in three-dimensional form, the candidate features can be converted into corresponding three-dimensional sequences according to the preset training rules to obtain the corresponding reconstruction features. Furthermore, according to the reconstruction features, the sample data set can be queried to obtain the data values ​​of the corresponding features for filling, and the first type of sample data about the candidate features that matches can be obtained.

[0115] The above three-dimensional form can be specifically expressed as: [seq_len, batch_size, input_dim].

[0116] Specifically, seq_len represents the sequence length, that is, the number of time steps; batch_size represents the amount of data input to the RNN model at one time; input_dim represents the input feature dimension of each time step.

[0117] In specific implementation, for wind power station supervision scenarios, combined with corresponding supervision requirements, the above time step can be set to 15 minutes, that is, the corresponding time resolution is 15 minutes.

[0118] Accordingly, first, seq_len can be set to 96 according to the time resolution and regulatory requirements (for example, if you want to predict the wind power for 24 hours a day).

[0119] Next, considering that a larger batch_size can speed up model training, it will also consume relatively more computing resources. In order to balance model training efficiency and server operation stability, batch_size can be set to 50.

[0120] Finally, we can set input_dim to 21 based on the number of features we have determined as candidate features.

[0121] When constructing the initial first prediction model, the constructed initial first prediction model at least includes an initial first input layer, an initial RNN hidden layer, and an initial first output layer.

[0122] The input of the initial RNN hidden layer includes the state of the output at the previous moment and the data input at the current moment.

[0123] In specific implementation, based on the above model structure, the data input into the first prediction model is processed by the first input layer and then input into the RNN hidden layer; in the RNN hidden layer, the corresponding state data can be obtained by processing according to the following formula:

[0124]

[0125] Among them, h t Represents the state data of the RNN hidden layer output at the current moment, h t-1 represents the state data of the RNN hidden layer output at the previous moment, f represents the activation function of the RNN hidden layer, U represents the weight matrix between the state output at the previous moment to the hidden layer, W represents the weight matrix between the first input layer and the RNN hidden layer, b represents the initial bias, and x t The input data at the current moment that is processed by the first input layer and input to the RNN hidden layer.

[0126] In specific implementation, based on the above model structure, the state data processed and output by the RNN hidden layer will be input into the second output layer. In the first output layer, the corresponding predicted value can be obtained and output according to the following formula:

[0127] y t =g(V•h t )

[0128] Among them, y t is the predicted value, g represents the activation function of the first output layer, and V represents the weight matrix between the RNN hidden layer and the first output layer.

[0129] Specifically, the above W also includes the weight matrix from the hidden layer to the hidden layer itself (also called loop weight or state weight). In this way, the hidden layer state of the previous time step can be transferred to the current time step based on W, thereby realizing the circulation and transfer of information. In turn, the RNN hidden layer can capture the time dependency in the sequence data, and then save and accumulate historical information through the state update and transfer mechanism, and abstractly process and analyze this information to generate a prediction for the current moment.

[0130] When training the first prediction model, the first type of sample data can be continuously input into the initial first prediction model for learning, so as to continuously stop and update the relevant model parameters (for example, parameters such as U, W, and V) in the model until the model converges and a first prediction model that meets the requirements is obtained.

[0131] For further information, see Figure 3 As shown, the above RNN hidden layer may also include multiple connected substructures, wherein each substructure may share parameters such as U, W, and V, thereby reducing the parameters that need to be learned in the network and improving the model training efficiency of the first prediction model.

[0132] In some embodiments, see Figure 4 As shown, the above-mentioned method of determining the shap value of the candidate feature by using the first prediction model according to the preset processing rule may include the following contents during specific implementation:

[0133] S1: According to the preset processing rules, use the sample data set to construct test data about the candidate features;

[0134] S2: obtaining modified data by modifying at least one candidate feature in the test data; and combining the test data and the modified data to establish a test data set;

[0135] S3: Processing the test data set using the first prediction model to obtain corresponding prediction results of multiple types of situations;

[0136] S4: Determine the shap value of the candidate feature according to the prediction results of various types of situations.

[0137] Based on the above embodiments, SHAP technology can be effectively utilized to quantitatively and accurately evaluate the contribution of each candidate feature to model prediction by calculating the SHAP value of the candidate feature. Then, the SHAP value can be used as a screening basis to screen and determine the target features with relatively better effects and relatively higher accuracy.

[0138] In specific implementation, for each test data, one candidate feature in the test data can be deleted to obtain corresponding multiple modified data; then the test data and the modified test data set can be combined. The above test data set contains both test data with complete candidate features and test data missing one candidate feature.

[0139] In some embodiments, during specific implementation, the shap value of the candidate feature may be determined according to the following formula:

[0140]

[0141] in, is the shap value of the candidate feature numbered i, N represents the set of candidate features in the test data set, the dimension of N is M, S represents a subset of the candidate feature numbered i extracted based on N, and the dimension of S is |s|, f x (S) represents the average value of the prediction results of the corresponding situation type obtained by processing the test data about S in the test data set using the first prediction model, It represents the average value of the prediction results of the corresponding situation type obtained by processing the test data about S and the union of the candidate features numbered i in the test data set using the first prediction model.

[0142] In some embodiments, when specifically screening facial features, a preset shap value reference value for basic features and a preset shap value reference value for lagging features can be determined according to the second screening rule; then it can be determined whether the type of the current candidate feature is a basic feature or a lagging feature; then a preset shap value reference corresponding to the current candidate feature is determined, and it is detected whether the shap value of the current candidate feature is greater than the preset shap value reference value. When the shap value of the candidate feature is greater than the preset shap value reference value, the candidate feature can be determined as a target feature.

[0143] Among them, the preset shap value reference value can be determined in the following manner. The shap value of each candidate feature is calculated using the sample data set and the first prediction model. According to the type of the candidate feature, the shap value of the candidate feature is split into two data groups, which are respectively recorded as the first data group corresponding to the basic feature and the second data group corresponding to the lagging feature; then the shap values ​​in the first data group and the second data group that are less than the preset tolerance threshold (i.e., the contribution is obviously too low) and the shap values ​​with negative values ​​(i.e., have a negative effect on the model prediction) are filtered to obtain the processed first data group and the processed second data group; then the data distribution statistics are performed on the processed first data group and the data statistics are performed on the processed second data group respectively, and the preset shap value reference value corresponding to the basic feature and the preset shap value reference value corresponding to the lagging feature are determined according to the data distribution statistics results.

[0144] In some embodiments, during specific implementation, the sample data set can be queried according to the target feature, and the data value of the corresponding feature can be obtained for filling, so as to obtain the second type of sample data matching the target feature.

[0145] When constructing the initial second prediction model, the initial TCN layer at least includes an initial causal convolution layer and an initial dilated convolution layer. In addition, the initial TCN layer also includes a residual module connected by a residual. The initial second output layer can be a linear layer with a dimension of 64.

[0146] Specifically, the causal convolution layer is used to extract and process key features based on the causal relationship by performing corresponding feature operations on the feature sequence output after processing by the second input layer, obtain the corresponding first intermediate feature and input it to the dilated convolution layer. The dilated convolution layer will further associate and process the features by performing corresponding feature operations on the first intermediate feature to obtain the second intermediate feature. Finally, the TCN layer can perform corresponding operations based on the intermediate features to determine the final prediction result; and output the model through the second output layer.

[0147] The above-mentioned residual module constrains the causal convolution layer and the dilated convolution layer, so as to continuously and specifically modify the network parameters of the causal convolution layer and the dilated convolution layer during the model training process to obtain a second prediction model that meets the requirements.

[0148] Specifically, the feature sequence output by the second input layer can be expressed as: X = (x1, x2, ... x t The convolution kernel (or filter) used in the TCN layer can be in the following form: F = (f1, f2, ... f k ).

[0149] In the causal convolutional layer, feature operations can be performed according to the following formula to calculate the value of x t Causal convolution at: .

[0150] In the dilated convolutional layer, feature operations can be performed according to the following formula to calculate the value of x t Dilated convolution at: , where d is the expansion rate.

[0151] During specific training, the residual module can be operated according to the following formula: H(x)=F(x)+x, the model's input x is weighted and fused into the model's output F(x) to calculate the corresponding residual, and the network parameters are continuously adjusted and updated according to the residual until the model converges.

[0152] Finally, the above TCN layer can calculate the final prediction result according to the following formula: y=Activation(F(x)+x). Among them, y represents the prediction result, X represents the input data of the TCN layer, and Activation represents the activation function of the TCN layer.

[0153] When training the second prediction model, the optimizer can be Adam, the learning rate can be set to 0.001, the loss function can be MAE, and the preset number of training rounds can be set to 500. During the training process, the loss function value can be calculated based on the prediction results; then the model parameters are optimized based on the loss function to finally obtain a second prediction model that meets the requirements.

[0154] In some embodiments, after the initial second prediction model is trained using the second type of sample data to obtain a second prediction model that meets the requirements, refer to Figure 5 As shown, when the method is implemented specifically, it may also include the following contents:

[0155] S1: Obtain data values ​​of target characteristics of the target wind power station at multiple moments in the previous target time period; query the power generation record of the target wind power station to obtain the measured values ​​of wind power of the target wind power station at multiple moments in the previous target time period;

[0156] S2: Determine the predicted values ​​of wind power of the target wind power station at multiple moments in the previous target time period by processing the data values ​​of the target characteristics of the target wind power station at multiple moments in the previous target time period using the second prediction model;

[0157] S3: calculating the model accuracy of the second prediction model according to the measured values ​​of the wind power of the target wind power station at multiple moments in the previous target time period and the predicted values ​​of the wind power of the target wind power station at multiple moments in the previous target time period;

[0158] S4: Check whether the model accuracy is greater than or equal to a preset accuracy threshold;

[0159] S5: When it is determined that the model accuracy is less than a preset accuracy threshold, adjust and update the second prediction model according to a preset adjustment rule.

[0160] Specifically, in order to adapt to the wind power station supervision scenario, the multiple moments in the target time period may be 96 moments at intervals of 15 minutes in a day. The preset accuracy threshold may be 85%.

[0161] In specific implementation, combined with the specific wind power station supervision scenario, the model accuracy of the second prediction model can be calculated according to the following formula:

[0162]

[0163] Among them, Acc day-ahead represents the model accuracy, Cap is the maximum startup capacity of the target wind power station in the target time period, and P i is the measured value of wind power at the target wind power station at time i, P i ' is the predicted value of wind power of the target wind power station at time i, and n is the total number of time.

[0164] Based on the above embodiments, it is possible to accurately detect and determine whether the second prediction model currently trained meets the scenario requirements of the wind power station supervision scenario; and when it is determined that the second prediction model does not meet the scenario requirements, the second prediction model can be adjusted and updated in a targeted manner to obtain a second prediction model that meets the scenario requirements.

[0165] In some embodiments, during specific implementation, when it is determined that the model accuracy is greater than or equal to a preset accuracy threshold, it can be determined that the second prediction model meets the scenario requirements of the wind power station supervision scenario for the target wind power station, and then the second prediction model can be used to predict the wind power of the target wind power station in real time, and based on the predicted value of the wind power of the target wind power station, guide relevant decisions such as power grid distribution and reward data formulation for the target wind power station.

[0166] When it is determined that the model accuracy is less than the preset accuracy threshold, the shap value of the target feature can be queried and obtained; and the target feature that has the greatest impact on the model prediction and belongs to the first level range is screened out according to the shap value as the first priority target feature; the remaining target features except the first priority target feature are determined as the second priority target feature. Then, according to the preset adjustment rule, the model parameters corresponding to the first priority target feature in the second prediction model are adjusted to obtain the first adjusted model; then the model accuracy is re-determined based on the first adjusted model, and the model accuracy is detected whether it is less than the preset accuracy threshold. When it is determined that it is less than the preset accuracy threshold, it is further detected whether it is greater than or equal to the preset first tolerance threshold. If it is greater than or equal to the preset accuracy tolerance threshold, the model parameters corresponding to the second priority target feature in the second prediction model can be adjusted until the model accuracy is greater than or equal to the preset accuracy threshold, so as to obtain a second prediction model that meets the requirements. On the contrary, if it is less than the preset accuracy tolerance threshold, the model parameters corresponding to the first priority target feature in the second prediction model are continuously adjusted until the model accuracy is greater than or equal to the preset accuracy tolerance threshold. Therefore, by making targeted adjustments to the second prediction model, a second prediction model that is targeted at the target wind power station and meets specific wind power station supervision requirements can be quickly updated.

[0167] In some embodiments, when it is determined that the model accuracy is less than a preset accuracy threshold, the loss data of the wind power reward data of the target power station can be calculated based on the model accuracy; and then, based on the loss data of the wind power reward data, it is determined whether the second prediction model needs to be adjusted and updated.

[0168] In specific implementation, the target power station's assessment electricity can be calculated as the loss data of wind power reward data according to the following formula: Assessment electricity = (85% - Acc day-ahead )*PN*0.4 (hours).

[0169] In specific implementation, it is possible to detect whether the data value of the loss data is greater than a preset loss risk value, and when it is determined to be greater than the preset loss risk value, trigger the adjustment and update of the second prediction model. The preset loss risk value can be flexibly set according to the scenario requirements of the target wind power station.

[0170] In some embodiments, after determining the wind power of the target wind power station at the current target moment, the electric energy storage plan of the target wind power station can be finely adjusted in a targeted manner according to the wind power of the target wind power station at the current target moment. Specifically, when it is determined that the wind power of the target wind power station at the current target moment is greater than or equal to the first power reference value, the electric energy generated by the wind power station can be divided into the first electric energy and the second electric energy according to the first distribution ratio; wherein the first electric energy is fed into the power grid, and the second electric energy is converted into the potential energy of the compressed gas through the compressed air energy storage system and stored in the electric energy storage system. When it is determined that the wind power of the target wind power station at the current target moment is less than the first power reference value and greater than or equal to the second power reference value, the electric energy generated by the wind power station can be fed into the power grid. When it is determined that the wind power of the target wind power station at the current target moment is less than the second power reference value, all the electric energy generated by the wind power station is fed into the power grid, and at the same time, the compressed gas stored in the electric energy storage system is converted into electric energy through the compressed air energy storage system and fed into the power grid for supplementation. Thereby, the power grid distribution network can be made relatively more stable and reliable.

[0171] As can be seen from the above, the method for determining the wind power of a wind power station provided in the embodiment of this specification can screen out matching alternative features and construct corresponding reconstruction features before specific implementation; then use the reconstruction features and sample data sets to construct the first type of sample data about the alternative features; use the first type of sample data to train the initial first prediction model based on RNN to obtain a first prediction model that meets the requirements; according to the preset processing rules, use the first prediction model to determine the shap value of the alternative features; according to the second screening rules and the shap value of the alternative features, select the standard features with better effects from the alternative features; then, according to the preset training rules, use the target features and sample data sets to construct matching second type of sample data about the target features; and construct the initial second prediction model based on TCN; use the second type of sample data to train the initial second prediction model to obtain a second prediction model that supports the realization of short-term prediction of wind power at multiple consecutive moments in a time period. In specific implementation, first obtain the data value of the target feature of the target wind power station at the current target moment; then determine the wind power of the target wind power station at the current target moment by processing the data value of the target feature using the second prediction model. In this way, the wind power of the target wind power station at multiple consecutive moments in the current time period can be determined more efficiently and accurately to guide decisions related to the target wind power station and effectively reduce decision errors.

[0172] The embodiment of this specification also provides a data processing method. When the method is implemented specifically, it may include the following contents:

[0173] S1: According to the first screening rule, matching features are screened out from the feature set about the wind power station as candidate features;

[0174] S2: According to the preset training rules, using the candidate features, constructing corresponding reconstruction features; and constructing an initial first prediction model based on RNN; wherein the initial first prediction model at least includes an initial first input layer, an initial RNN hidden layer, and an initial first output layer;

[0175] S3: constructing the first type of sample data matching the candidate features according to the reconstructed features and the sample data set;

[0176] S4: using the first type of sample data to train an initial first prediction model to obtain a first prediction model that meets the requirements;

[0177] S5: Determine the shap value of the candidate feature using the first prediction model according to the preset processing rule;

[0178] S6: According to the second screening rule and the shap values ​​of the candidate features, a target feature meeting the requirements is screened out from the candidate features;

[0179] S7: According to the preset training rules, using the target features and the sample data set, construct a second type of sample data matching the target features; and construct an initial second prediction model based on TCN; wherein the initial second prediction model at least includes an initial second input layer, an initial TCN layer, and an initial second output layer;

[0180] S8: Using the second type of sample data to train an initial second prediction model to obtain a second prediction model that meets the requirements; wherein the second prediction model is used to predict the wind power of the wind power station.

[0181] Based on the above embodiment, based on the above embodiment, according to the preset training rules, the RNN model, TCN model, shap value technology and other technologies can be used together to efficiently construct and train a second prediction model that is suitable for the wind power station supervision scenario and can accurately perform continuous prediction of medium- and short-term wind power. Then, when the wind power station is subsequently supervised, the second prediction model can be used to efficiently and accurately determine the wind power of the wind power station, providing a valuable reference for the wind power station's power grid distribution, market decision-making, etc.

[0182] This specification embodiment provides a server, referring to Figure 6 The server includes a network communication port 601, a processor 602 and a memory 603, and the above structures are connected through internal cables so that each structure can perform specific data interaction.

[0183] The network communication port 601 may be used to obtain data values ​​of target characteristics of the target wind power station at the current target moment;

[0184] The processor 602 can be specifically used to use the second prediction model to process the data value of the target feature to obtain a corresponding target prediction result; wherein the second prediction model is a TCN model trained according to a preset training rule using the second type of sample data about the target feature and the first prediction model, and the first prediction model is an RNN model trained using the first type of sample data about the alternative feature; based on the target prediction result, the wind power of the target wind power station at the current target moment is determined.

[0185] The memory 603 may be specifically used to store corresponding instruction programs, as well as data values ​​of target characteristics of the target wind power station at the current target moment, target prediction results and other related data.

[0186] Based on the above method, the relevant structural performance of the server can be effectively utilized, the data processing speed of the electronic equipment can be improved, and the data processing related to the determination of the wind power of the wind power station can be efficiently realized.

[0187] In this embodiment, the network communication port 601 can be a virtual port that is bound to different communication protocols so that different data can be sent or received. For example, the network communication port can be a port responsible for web data communication, a port responsible for FTP data communication, or a port responsible for email data communication. In addition, the network communication port can also be a physical communication interface or communication chip. For example, it can be a wireless mobile network communication chip, such as GSM, CDMA, etc.; it can also be a Wifi chip; it can also be a Bluetooth chip.

[0188] In this embodiment, the processor 602 may be implemented in any appropriate manner. For example, the processor may take the form of a microprocessor or a processor and a computer-readable medium storing a computer-readable program code (such as software or firmware) executable by the (micro)processor, a logic gate, a switch, an application specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller, etc. This specification does not limit this.

[0189] In this embodiment, the memory 603 may include multiple levels. In a digital system, anything that can store binary data can be a memory; in an integrated circuit, a circuit with a storage function but no physical form is also called a memory, such as RAM, FIFO, etc.; in a system, a storage device with a physical form is also called a memory, such as a memory stick, TF card, etc.

[0190] The embodiment of this specification also provides a computer-readable storage medium based on the above-mentioned method for determining the wind power of the wind power station, wherein the computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed, the following steps are implemented: obtaining the data value of the target feature of the target wind power station at the current target moment; using a second prediction model to process the data value of the target feature to obtain a corresponding target prediction result; wherein the second prediction model is a TCN model trained according to a preset training rule using the second type of sample data about the target feature and the first prediction model, and the first prediction model is an RNN model trained using the first type of sample data about the alternative feature; according to the target prediction result, determining the wind power of the target wind power station at the current target moment.

[0191] In this embodiment, the storage medium includes but is not limited to a random access memory (RAM), a read-only memory (ROM), a cache, a hard disk (HDD) or a memory card. The memory may be used to store computer program instructions. The network communication unit may be an interface for network connection communication set in accordance with the standard specified by the communication protocol.

[0192] In this embodiment, the functions and effects specifically implemented by the program instructions stored in the computer-readable storage medium can be explained in comparison with other implementations and will not be described in detail here.

[0193] The embodiments of the present specification also provide a computer program product, which at least includes a computer program, and when the computer program is executed by a processor, it implements the following method steps: obtaining the data value of the target feature of the target wind power station at the current target moment; using a second prediction model to process the data value of the target feature to obtain a corresponding target prediction result; wherein the second prediction model is a TCN model trained according to a preset training rule using the second type of sample data about the target feature and the first prediction model, and the first prediction model is an RNN model trained using the first type of sample data about the alternative feature; based on the target prediction result, determining the wind power of the target wind power station at the current target moment.

[0194] See also Figure 7 As shown, the embodiment of this specification also provides a device for determining wind power of a wind power station, which may specifically include the following structural modules:

[0195] The acquisition module 701 may be specifically used to acquire the data value of the target characteristic of the target wind power station at the current target moment;

[0196] The processing module 702 may be specifically used to process the data value of the target feature using the second prediction model to obtain a corresponding target prediction result; wherein the second prediction model is a TCN model trained using the second type of sample data about the target feature and the first prediction model according to a preset training rule, and the first prediction model is an RNN model trained using the first type of sample data about the candidate feature;

[0197] The determination module 703 may be specifically used to determine the wind power of the target wind power station at the current target moment according to the target prediction result.

[0198] In some embodiments, the device may further include a model training module, which may be used to: select matching features from a feature set about a wind power station as candidate features according to a first screening rule; construct corresponding reconstruction features using the candidate features according to a preset training rule; and construct an initial first prediction model based on RNN; wherein the initial first prediction model includes at least an initial first input layer, an initial RNN hidden layer, and an initial first output layer; construct a matching first class of sample data about the candidate features according to the reconstruction features and the sample data set; train the initial first prediction model using the first class of sample data to obtain a matching first prediction model. A first prediction model that meets the requirements; according to a preset processing rule, the first prediction model is used to determine the shap value of the alternative feature; according to a second screening rule and the shap value of the alternative feature, a target feature that meets the requirements is screened out from the alternative features; according to a preset training rule, the target feature and the sample data set are used to construct a matching second type of sample data about the target feature; and an initial second prediction model based on TCN is constructed; wherein the initial second prediction model at least includes an initial second input layer, an initial TCN layer, and an initial second output layer; the initial second prediction model is trained using the second type of sample data to obtain a second prediction model that meets the requirements.

[0199] In some embodiments, when the above-mentioned model training module is implemented, matching features can be selected as candidate features from a feature set related to the wind power station according to the first screening rule in the following manner: obtaining environmental parameters of the target wind power station in the current time period; and selecting matching features as candidate features from a feature set related to the wind power station according to the first screening rule and the environmental parameters of the target wind power station in the current time period.

[0200] In some embodiments, the alternative features may specifically include: 1-period lag data of wind power, 2-period lag data of wind power, 3-period lag data of wind power, 96-period lag data of wind power, 97-period lag data of wind power, 98-period lag data of wind power, 35040-period lag data of wind power, 35041-period lag data of wind power, 35042-period lag data of wind power, current moment data of wind speed, 1-period lag data of wind speed, 2-period lag data of wind speed, 3-period lag data of wind speed, 96-period lag data of wind speed, 97-period lag data of wind speed, 98-period lag data of wind speed, 35040-period lag data of wind speed, 35041-period lag data of wind speed, 35042-period lag data of wind speed, current moment data of temperature, current moment data of humidity, etc.

[0201] In some embodiments, when the above-mentioned model training module is implemented, the shap value of the alternative feature can be determined by using the first prediction model according to the preset processing rules in the following manner: according to the preset processing rules, test data about the alternative feature is constructed using a sample data set; modified data is obtained by modifying at least one alternative feature in the test data; and the test data and the modified data are combined to establish a test data set; the test data set is processed using the first prediction model to obtain corresponding prediction results for multiple situation types; and the shap value of the alternative feature is determined based on the prediction results for multiple situation types.

[0202] In some embodiments, when the above model training module is implemented, the shap value of the candidate feature can be determined according to the following formula:

[0203]

[0204] in, is the shap value of the candidate feature numbered i, N represents the set of candidate features in the test data set, the dimension of N is M, S represents a subset of the candidate feature numbered i extracted based on N, and the dimension of S is |S|, f x (S) represents the average value of the prediction results of the corresponding situation type obtained by processing the test data about S in the test data set using the first prediction model, It represents the average value of the prediction results of the corresponding situation type obtained by processing the test data about S and the union of the candidate features numbered i in the test data set using the first prediction model.

[0205] In some embodiments. After the initial second prediction model is trained with the second type of sample data to obtain a second prediction model that meets the requirements, when the model training module is specifically implemented, it can also be used to: obtain the data values ​​of the target characteristics of the target wind power station at multiple moments of the previous target time period; query the power generation record of the target wind power station to obtain the measured values ​​of the wind power of the target wind power station at multiple moments of the previous target time period; determine the predicted values ​​of the wind power of the target wind power station at multiple moments of the previous target time period by processing the data values ​​of the target characteristics of the target wind power station at multiple moments of the previous target time period with the second prediction model; calculate the model accuracy of the second prediction model based on the measured values ​​of the wind power of the target wind power station at multiple moments of the previous target time period and the predicted values ​​of the wind power of the target wind power station at multiple moments of the previous target time period; detect whether the model accuracy is greater than or equal to a preset accuracy threshold; when it is determined that the model accuracy is less than the preset accuracy threshold, adjust and update the second prediction model according to the preset adjustment rules.

[0206] It should be noted that the units, devices or modules described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. For the convenience of description, the above devices are described separately by functions divided into various modules. Of course, when implementing this specification, the functions of each module can be implemented in the same or more software and / or hardware, or the modules that implement the same function can be implemented by a combination of multiple sub-modules or sub-units. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0207] It can be seen from the above that the device for determining the wind power of a wind power station provided based on the embodiment of this specification can be well adapted to the wind power station supervision scenario, and can more efficiently and accurately determine the wind power of the target wind power station at multiple consecutive moments in the current time period, so as to guide decisions related to the target wind power station and effectively reduce decision errors.

[0208] In a specific scenario example, the method for determining wind power of a wind power station provided in this specification can be applied to implement wind power prediction based on TCN and SHAP values. For the specific implementation process, please refer to the following content.

[0209] In this scenario example, considering that the temporal convolutional network (TCN) is a deep learning model that supports processing time series data, it can effectively capture long-term dependencies and patterns in sequence data, and can show strong advantages in the field of wind power forecasting. However, the simple TCN model is still insufficient in explaining the model prediction results and key features, and then consider combining other technologies to enhance the interpretability of the model.

[0210] Considering that SHAP value (Shapley Additive exPlanations) is a model interpretation technology that can quantitatively evaluate the contribution of each feature to the model prediction results, it provides a powerful tool for the interpretability of the model. We then consider combining SHAP value with TCN algorithm, which can not only improve the accuracy of short-term wind power forecast, but also analyze and explain the key features of the model through SHAP value, providing a scientific basis for feature screening and model optimization of wind power forecast.

[0211] Based on the above considerations, the applicant proposed a wind power prediction method and system based on TCN and SHAP value. Figure 8 As shown. Among them, the system can realize the automation of the whole process from data preprocessing to model prediction by supporting the docking of big data platform, integrating modules such as data cleaning, feature screening and prediction model building. At the same time, by combining TCN algorithm and SHAP value analysis technology, it not only improves the accuracy of short-term wind power prediction, but also enhances the interpretability and adaptability of the model, providing strong support for the safe, stable and economical operation of wind power system.

[0212] For details, see Figure 8 As shown, the wind power prediction system based on TCN and SHAP value has a data cleaning module, a feature screening module, a prediction model building module, and a prediction effect evaluation module.

[0213] The operational steps for predicting wind power based on the above system may specifically include: selecting historical power data of the site to be predicted and historical data such as wind speed and temperature at the corresponding time from the big data platform, and performing data cleaning and preprocessing to ensure data quality and consistency; constructing a basic model of the TCN algorithm (for example, an initial second prediction model) based on the data items determined in step S1, and training and tuning the model in combination with cross-validation and extreme value optimization methods for short-term prediction of wind power.

[0214] Specifically, when training the basic model of the TCN algorithm, the short-term prediction of wind power can be achieved by combining the TCN algorithm with SHAP value analysis. At the same time, the SHAP value is used to process the RNN algorithm model (for example, the first prediction model) to discover and explain the key features of the model, and provide guidance for the screening of short-term prediction features of wind power (for example, screening and determining target features from alternative features).

[0215] In addition, this system has high scalability and adaptability, and can be optimized and adjusted according to the type of site, bringing better prospects for use. Specifically, the model architecture of the system has good flexibility, such as being able to integrate other models well: TCN can be integrated with other machine learning or deep learning models (such as LSTM, GRU, etc.) to form a hybrid model to further improve prediction accuracy; the system also has good system scalability, and the system adopts a modular design, so that each module is relatively independent, which is easy to maintain and upgrade. In this way, when it is necessary to introduce a new prediction algorithm or data source (to refine the actual business situation scenario), it is only necessary to modify or expand the corresponding module.

[0216] Specifically, it is also possible to perform anomaly detection on wind power data, identify and locate abnormal values. These abnormal values ​​may be caused by equipment failure, adverse weather conditions or data collection errors, and appropriate measures need to be taken to correct the detected abnormal values ​​or supplement missing values.

[0217] In this scenario example, first, select the candidate feature set D for short-term wind power prediction and consider P T-1 , P T-2 , P T-3 , P T-96 , P T-97 , P T-98 , P T-35040 , P T-35041 , P T-35042 , W T , W T-1 , W T-2 , W T-3 , W T-96 , W T-97 , W T-98 , W T-35040 , W T-35041 , W T-35042, Temp, and Hum, namely, lag 1 period data of wind power, lag 2 period data of wind power, lag 3 period data of wind power, lag 96 period data of wind power, lag 97 period data of wind power, lag 98 period data of wind power, lag 35040 period data of wind power, lag 35041 period data of wind power, and lag 35042 period data of wind power; current moment data of wind speed, lag 1 period data, lag 2 period data of wind speed, lag 3 period data of wind speed, lag 96 period data of wind speed, lag 97 period data of wind speed, lag 98 period data of wind speed, lag 35040 period data of wind speed, lag 35041 period data of wind speed, and lag 35042 period data of wind speed; current moment data of temperature and current moment data of humidity. That is, for the short-term prediction of wind power, it is believed that the wind power at this moment is most correlated with the wind power at the previous moment, the previous two moments, the previous three moments, the same moment of the previous day, the moment before the same moment of the previous day, the two moments before the same moment of the previous day, the same moment of the previous year, the moment before the same moment of the previous year, and the two moments before the same moment of the previous year, as well as the wind speed at this moment, the previous moment, the previous two moments, the previous three moments, the same moment of the previous day, the moment before the same moment of the previous day, the two moments before the same moment of the previous day, the same moment of the previous year, the moment before the same moment of the previous year, and the two moments before the same moment of the previous year, and the temperature and humidity at this moment (that is, according to the first screening rule, matching features are screened out from the feature set about wind power stations as alternative features).

[0218] Next, reconstruct the input features of the model: reconstruct the 21 features in the above-mentioned candidate feature set into three-dimensional features that can be recognized by RNN (that is, according to the preset training rules, use the candidate features to construct the corresponding reconstructed features). Fig. 9 As shown in the figure, the input data of RNN is usually a three-dimensional tensor with a shape of [seq_len, batch_size, input_dim]. seq_len represents the sequence length, that is, the number of time steps; batch_size represents the number of samples input to RNN at a time; input_dim represents the input feature dimension of each time step.

[0219] In the specific implementation, determine seq_len: This depends on the prediction requirements and the time resolution of the data. The resolution of the wind power prediction data is 15 minutes, and it is hoped to predict the wind power in the next 24 hours. Take seq_len as 96. Determine batch_size: It depends on the computing resources and model training requirements. A larger batch_size can speed up training, but it may also require more memory. Set it to 50. Determine input_dim: That is, the number of features in the candidate feature set is 21.

[0220] Then, based on the above reconstructed features, an RNN prediction model can be constructed (i.e., an initial first prediction model based on RNN is constructed). In specific implementation, the output function of the RNN hidden layer includes the state at the previous moment and the input state at this moment, so the output of the hidden layer is expressed in the following form:

[0221]

[0222] Among them, f is the activation function at the hidden layer. Function U is the weight matrix between the state value at the previous moment to the hidden layer, W is the weight matrix from the input layer to the hidden layer, and b is the initial bias value. t is the input, h t is the output.

[0223] The above matrix V is the function from the hidden layer to the output layer, g is the activation function of the output layer, y t is the predicted value of the output.

[0224] The predicted value of RNN output can be expressed as: t =g(V•h t ).

[0225] Among them, x is the input layer, y is the output layer, h is the hidden layer, and t refers to the number of calculations; V, W, U are weights.

[0226] Specifically, U: weight matrix from input to hidden layer. Converts input data into a form that can be processed by hidden layer. V: weight matrix from hidden layer to output. Converts the state of hidden layer into output data. W: weight matrix from hidden layer to hidden layer itself (also called loop weight or state weight). Transfers the hidden layer state of the previous time step to the current time step, thus realizing the circulation and transmission of information.

[0227] The role of the above RNN hidden layer is to capture the time dependency in the sequence data, save and accumulate historical information through the state update and transmission mechanism, and abstractly process this information to generate a prediction for the current time step.

[0228] See also Figure 3 As shown in the figure, the circular arrows represent the self-connections of the hidden layer. In RNN, each layer shares the parameters U, V, and W, which reduces the parameters that need to be learned in the network and improves learning efficiency.

[0229] Furthermore, the shap values ​​of the 18 features in the candidate feature set can be calculated based on the RNN model trained in the previous step. Based on the shap value technology, the following steps are included:

[0230] S1: Define the benchmark value. In order to calculate the Shap value, you need to define a benchmark value. The benchmark value is a reference point, which represents the prediction result without any feature influence. The average prediction value of the entire training set is selected as the benchmark value.

[0231] S2: Calculate the Shap value for each sample and each feature. The Shap value measures the contribution of the feature to the model prediction result. It iteratively adds or removes features, observes the changes in the prediction results, and calculates the contribution of different feature combinations.

[0232] S3: Explain the prediction results. The Shap value explains the impact of each feature on the prediction results, as well as the contribution of different feature combinations to the prediction results.

[0233] Based on the above ideas, consider assuming that the i-th sample is x i , the jth feature of the i-th sample is x ij , the model's predicted value for the sample is y i , the baseline of the entire model, that is, the mean of all sample target variables is y base , then the SHAP value obeys the following equation:

[0234] y i =y base +f(x i1 )+ f(x i2 )+……+ f(x ij )

[0235] Among them, f(x ij ) is x ij SHAP value of f(x i1 ) is the final predicted value y of the first feature in the i-th sample i The SHAP value for each feature represents the change in the expected model prediction when conditioning on that feature. For each feature, the SHAP value illustrates the contribution to illustrate the difference between the average model prediction and the actual prediction for an instance. When f(x i1 )>0, it means that the feature improves the prediction value, otherwise, it means that the feature reduces the contribution.

[0236] Consider again that the Shap value can be interpreted as an additive feature attribution method, interpreting the model's predicted value as a linear function of a binary variable:

[0237]

[0238] Among them, f(x) represents the predicted value, , indicates how many of the D candidate features are included in the prediction model. For a certain sample, if feature k is not in the prediction model, the SHAP value of the corresponding feature is 0, that is, , indicating that the feature k will not be attributed to the sample and has no contribution to the final prediction value.

[0239] For a specific sample, the SHAP value of feature i The calculation includes the following:

[0240]

[0241] Among them, N is the set of all features in the training set, its dimension is M, S is a subset extracted from N, its dimension is |S|, f x (S) means that only the feature set S is used to calculate the average value of the sample; It expresses that based on the feature set S, feature i is added to calculate the average value of the sample; It is the weight of the difference between the sample values ​​when feature i is included and when feature i is not included in the corresponding feature subset S.

[0242] In addition, since multiple feature combinations can be extracted from all features N to form a subset S, the SHAP value of feature i is a comprehensive score that enumerates all possible feature subsets S, taking into account the influence of other features on feature i in addition to itself.

[0243] The shap value of each feature in the candidate feature set is compared with the set threshold, and the features above the threshold are placed in the feature subset.

[0244] Among them, the above threshold is determined in the following way: First, calculate the SHAP value of all features and sort the features according to the size of the SHAP value. Then, observe the sorted feature list and analyze which features have significantly higher SHAP values ​​than other features. These features usually have a greater impact on the model prediction results. Starting from a model containing all features, gradually exclude features with lower SHAP values, retrain the model after each exclusion and evaluate its performance. Observe the trend of model performance as the number of features decreases, and stop excluding features when the model performance begins to decline significantly. At this point, a suitable threshold can be determined based on the SHAP values ​​of the remaining features. The threshold can be stored and can be called for feature screening later.

[0245] Then, a short-term prediction model (eg, a second prediction model) may be constructed using features in the feature subset (eg, target features).

[0246] In specific implementation, the above feature subset of the selected station can be input into TCN. For the TCN layer, for the filter F=(f1,f2,……f k ), feature sequence X=(x1,x2,……x t ), then in x t The calculation formula of the causal convolution layer is as follows: .

[0247] When the dilation rate is set to d, t The dilated convolution at is shown below: .

[0248] The residual module designs the network as H(x)=F(x)+x, and weightedly fuses the model's input x into the model's output F(x). The final output formula of TCN is as follows: y=Activation(F(x)+x).

[0249] Activation is the activation function. The above TCN output is connected to the Linear layer, and the output dimension of the Linear layer is set to 64.

[0250] During specific training, the model is trained according to the selection of the features and the historical sequence of wind power. During the training process, the optimizer selects Adam, a learning rate of 0.001, a loss function MAE, and 500 training rounds. The trained model is parameter optimized according to the training results to obtain a short-term prediction model for wind power of the target station. The wind power prediction model obtained through training can be used to predict future wind power, and the predicted value of future wind power of the target station can be obtained.

[0251] Finally, the effect of the short-term wind power prediction model can also be evaluated (equivalent to detecting whether the second prediction model meets the scenario requirements of the wind power station supervision scenario).

[0252] Specifically, taking Region A as an example, according to the assessment requirements of the “two detailed rules” in Region A, the accuracy of the wind farm’s next-day power forecast should be greater than or equal to 85%. The accuracy calculation formula is as follows:

[0253] .

[0254] If the accuracy is lower than 85% (for example, the preset accuracy threshold), the next day's forecast accuracy assessment power consumption is calculated according to the following formula: Assessment Power Consumption = (85% - Acc day-ahead )*PN*0.4 (hours).

[0255] Among them, Acc day-aheadis the prediction accuracy for the next day; Cap is the maximum operating capacity of the new energy power station on the assessment day (unit: MW); P i is the actual power of the new energy power station at time i (unit: MW), and the power restriction period P i is the available power of the new energy power station at time i (unit: MW); P i ' is the predicted power of the new energy power station at time i (unit: MW); n is the number of predicted points for the day's assessment (96 points); PN is the installed capacity of the new energy power station (unit: MW).

[0256] Furthermore, based on the prediction results of the above model, after obtaining the actual wind power results the next day, the prediction model effect evaluation module can be input to obtain the prediction accuracy and the assessed power volume, and the model can be optimized and adjusted in time to improve the accuracy, which can reduce the power volume assessed by the "two rules".

[0257] In addition, according to the spot trading situation of the selected power plant, the specific transaction income of the relevant market can be calculated according to the following settlement formula = day-ahead fee + deviation fee = Q day-ahead × P day-ahead + (Q real-time - Q day-ahead) × P real-time, where Q day-ahead is the day-ahead electricity volume settled based on the day-ahead power forecast report, P day-ahead is the day-ahead electricity price, and Q real-time is the real-time settlement electricity volume obtained from the real-time power curve of the wind farm. If the accuracy of the day-ahead power forecast curve is improved, the income from spot trading will be improved.

[0258] Furthermore, in this scenario example, the differences in wind power of wind power plants in different seasons are also taken into account. Historical data of the area where the wind farm is located can be called up, including wind power, wind speed, temperature, humidity, etc. These data should cover multiple seasons in order to analyze seasonal characteristics.

[0259] After collecting relevant historical data, you can preprocess the historical data first, including: 1. Data cleaning: remove outliers, missing values ​​and other data that do not meet the requirements to ensure the integrity and accuracy of the data. 2. Data normalization: convert data of different dimensions into the same dimension for subsequent data analysis and model building. 3. Seasonal division: divide the data set into different subsets according to seasonal characteristics (such as spring, summer, autumn, winter) for seasonal feature screening and model building.

[0260] In addition, when performing feature screening, you can also 1. Introduce correlation analysis: Use statistical methods to analyze the correlation between meteorological data and power generation data to determine the key factors affecting wind power. 2. Introduce seasonal feature screening: For different seasons, screen out the characteristic variables that have the greatest impact on wind power. For example, in winter, you may pay more attention to the stability of wind speed and wind direction, while in summer, you may pay more attention to the effects of temperature and humidity. Specifically, the sequence and lagged sequence of the feature library (wind speed, temperature, humidity) can be correlated with the wind power sequence, and the lagged autocorrelation analysis of the wind power sequence can be performed, using the Spearman rank correlation coefficient.

[0261] Among them, Spearman's rank correlation coefficient is a non-parametric correlation measure used to assess the degree of dependence between two variables without requiring the data to follow a normal distribution. It is calculated based on the ranks (i.e., the positions after sorting) of the two variables.

[0262] Through the above scenario examples, the method for determining the wind power of a wind power station provided in this manual is verified. 1) An innovative combination of multiple technologies is achieved: the temporal convolutional network (TCN) is combined with the SHAP value analysis technology and applied to the short-term prediction of wind power. This innovative combination of methods improves the prediction accuracy and model interpretability. 2) A modular prediction system is constructed, which includes a data cleaning module, a feature screening module, and a prediction model building module. The modules are made to work together to achieve full process automation from data processing to model prediction. 3) Data preprocessing and cleaning can be achieved: strict cleaning and preprocessing of historical data are emphasized to ensure data quality and consistency, providing a reliable data foundation for model training. 4) SHAP value application is introduced: SHAP value analysis technology is used to reveal and explain the key features of the model, provide a scientific basis for feature screening of wind power prediction, and enhance the interpretability and reliability of the model. 5) Introduced a prediction effect evaluation module: Added a prediction effect evaluation module to the wind power prediction system, focusing on the accuracy of the two detailed assessments, the accuracy of the electricity calculation, and the improvement of spot trading revenue, which can significantly improve the operational efficiency and economic benefits of the wind farm. 6) High scalability and adaptability: The system has high scalability and flexibility, and can be optimized and adjusted according to the specific conditions of different wind farms, which improves the actual application value of the system.

[0263] Although the present specification provides method operation steps as described in the embodiments or flow charts, more or fewer operation steps may be included based on conventional or non-creative means. The order of steps listed in the embodiments is only one way of executing the order of many steps and does not represent the only order of execution. When the device or client product in practice is executed, it can be executed in sequence or in parallel according to the method shown in the embodiments or the drawings (for example, a parallel processor or a multi-threaded processing environment, or even a distributed data processing environment). The terms "include", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, product or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed, or also includes elements inherent to such a process, method, product or device. In the absence of more restrictions, it is not excluded that there are other identical or equivalent elements in the process, method, product or device including the elements. The words first, second, etc. are used to represent the name, and do not represent any particular order.

[0264] Those skilled in the art also know that, in addition to implementing the controller in a purely computer-readable program code, the controller can be made to implement the same function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, such a controller can be considered as a hardware component, and the devices for implementing various functions included therein can also be considered as structures within the hardware component. Or even, the devices for implementing various functions can be considered as both software modules for implementing the method and structures within the hardware component.

[0265] This specification may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc. that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer-readable storage media including storage devices.

[0266] Through the description of the above embodiments, it can be known that those skilled in the art can clearly understand that the present specification can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present specification can essentially be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a mobile terminal, a server, or a network device, etc.) to execute the methods described in each embodiment of the present specification or some parts of the embodiments.

[0267] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. This specification can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc.

[0268] Although the present specification is described through embodiments, those skilled in the art will appreciate that there are many modifications and changes to the present specification without departing from the spirit of the present specification, and it is intended that the appended claims include these modifications and changes without departing from the spirit of the present specification.

Claims

1. A method for determining wind power of a wind power station, characterized in that: include: Obtain data values ​​of target characteristics of the target wind power station at the current target time; Processing the data value of the target feature using the second prediction model to obtain a corresponding target prediction result; wherein the second prediction model is a TCN model trained using the second type of sample data about the target feature and the first prediction model according to a preset training rule, and the first prediction model is an RNN model trained using the first type of sample data about the candidate feature; According to the target prediction result, the wind power of the target wind power station at the current target time is determined; The method also includes: according to the first screening rule, selecting matching features from the feature set about the wind power station as candidate features; according to the preset training rule, using the candidate features to construct corresponding reconstruction features; and constructing an initial first prediction model based on RNN; wherein the initial first prediction model at least includes an initial first input layer, an initial RNN hidden layer, and an initial first output layer; according to the reconstruction features and the sample data set, constructing a matching first category of sample data about the candidate features; using the first category of sample data to train the initial first prediction model to obtain a first prediction model that meets the requirements; according to the preset processing rule, using the first prediction model to determine the shap value of the candidate features; according to the second screening rule and the shap value of the candidate features, selecting the target features that meet the requirements from the candidate features; according to the preset training rule, using the target features and the sample data set, constructing Create a matching second type of sample data about the target feature; and construct an initial second prediction model based on TCN; wherein the initial second prediction model at least includes an initial second input layer, an initial TCN layer, and an initial second output layer; use the second type of sample data to train the initial second prediction model to obtain a second prediction model that meets the requirements; wherein, according to a preset processing rule, use the first prediction model to determine the shap value of the alternative feature, including: according to the preset processing rule, use the sample data set to construct test data about the alternative feature; obtain the modified data by modifying at least one alternative feature in the test data; and combine the test data and the modified data to establish a test data set; use the first prediction model to process the test data set to obtain corresponding prediction results for multiple types of situations; determine the shap value of the alternative feature based on the prediction results for multiple types of situations.

2. The method according to claim 1, characterized in that According to the first screening rule, matching features are screened out from the feature set about the wind power station as candidate features, including: Obtain environmental parameters of the target wind power station in the current time period; According to the first screening rule and the environmental parameters of the target wind power station in the current time period, matching features are screened out from the feature set related to the wind power station as candidate features.

3. The method according to claim 1, characterized in that The alternative features include: 1-period lagged data of wind power, 2-period lagged data of wind power, 3-period lagged data of wind power, 96-period lagged data of wind power, 97-period lagged data of wind power, 98-period lagged data of wind power, 35040-period lagged data of wind power, 35041-period lagged data of wind power, 35042-period lagged data of wind power, current moment data of wind speed, 1-period lagged data of wind speed, 2-period lagged data of wind speed, 3-period lagged data of wind speed, 96-period lagged data of wind speed, 97-period lagged data of wind speed, 98-period lagged data of wind speed, 35040-period lagged data of wind speed, 35041-period lagged data of wind speed, 35042-period lagged data of wind speed, current moment data of temperature, and current moment data of humidity.

4. The method according to claim 1, characterized in that According to the following formula, determine the shap value of the candidate feature: in, is the shap value of the candidate feature numbered i, N represents the set of candidate features in the test data set, the dimension of N is M, S represents a subset of the candidate feature numbered i extracted based on N, and the dimension of S is , represents the average value of the prediction results of the corresponding situation type obtained by processing the test data about S in the test data set using the first prediction model, It represents the average value of the prediction results of the corresponding situation type obtained by processing the test data about S and the union of the candidate features numbered i in the test data set using the first prediction model.

5. The method according to claim 1, characterized in that After the initial second prediction model is trained using the second type of sample data to obtain a second prediction model that meets the requirements, the method further includes: Obtain data values ​​of target characteristics of the target wind power station at multiple moments in the previous target time period; query the power generation record of the target wind power station to obtain measured values ​​of wind power of the target wind power station at multiple moments in the previous target time period; Determine the predicted values ​​of wind power of the target wind power station at multiple moments in the previous target time period by processing the data values ​​of the target characteristics of the target wind power station at multiple moments in the previous target time period using the second prediction model; Calculating the model accuracy of the second prediction model according to the measured values ​​of the wind power of the target wind power station at multiple moments in the previous target time period and the predicted values ​​of the wind power of the target wind power station at multiple moments in the previous target time period; Check whether the model accuracy is greater than or equal to the preset accuracy threshold; When it is determined that the model accuracy is less than a preset accuracy threshold, the second prediction model is adjusted and updated according to a preset adjustment rule.

6. A device for determining wind power of a wind power station, characterized in that: include: An acquisition module, used to acquire data values ​​of target characteristics of a target wind power station at a current target moment; A processing module, used to process the data value of the target feature using the second prediction model to obtain a corresponding target prediction result; wherein the second prediction model is a TCN model trained using the second type of sample data about the target feature and the first prediction model according to a preset training rule, and the first prediction model is an RNN model trained using the first type of sample data about the candidate feature; A determination module is used to determine the wind power of the target wind power station at the current target moment according to the target prediction result; The device is also used to select matching features from a feature set about a wind power station as candidate features according to a first screening rule; construct corresponding reconstruction features using the candidate features according to a preset training rule; and construct an initial first prediction model based on RNN; wherein the initial first prediction model at least includes an initial first input layer, an initial RNN hidden layer, and an initial first output layer; construct a matching first category of sample data about the candidate features according to the reconstruction features and the sample data set; train the initial first prediction model using the first category of sample data to obtain a first prediction model that meets the requirements; determine the shap value of the candidate features using the first prediction model according to the preset processing rule; select the target features that meet the requirements from the candidate features according to the second screening rule and the shap value of the candidate features; construct a first prediction model using the target features and the sample data set according to the preset training rule; Create a matching second type of sample data about the target feature; and construct an initial second prediction model based on TCN; wherein the initial second prediction model at least includes an initial second input layer, an initial TCN layer, and an initial second output layer; use the second type of sample data to train the initial second prediction model to obtain a second prediction model that meets the requirements; wherein, according to a preset processing rule, use the first prediction model to determine the shap value of the alternative feature, including: according to the preset processing rule, use the sample data set to construct test data about the alternative feature; obtain the modified data by modifying at least one alternative feature in the test data; and combine the test data and the modified data to establish a test data set; use the first prediction model to process the test data set to obtain corresponding prediction results for multiple types of situations; determine the shap value of the alternative feature based on the prediction results for multiple types of situations.

7. A data processing method, characterized in that: include: According to the first screening rule, matching features are screened out from the feature set about the wind power station as candidate features; According to the preset training rules, the candidate features are used to construct the corresponding reconstruction features; And construct an initial first prediction model based on RNN; wherein the initial first prediction model at least includes an initial first input layer, an initial RNN hidden layer, and an initial first output layer; According to the reconstructed features and the sample data set, a first type of sample data about the candidate features that matches is constructed; Using the first type of sample data to train an initial first prediction model to obtain a first prediction model that meets the requirements; According to the preset processing rules, the shap value of the candidate feature is determined by using the first prediction model; According to the second screening rule and the shap values ​​of the candidate features, a target feature meeting the requirements is screened out from the candidate features; According to the preset training rules, using the target features and the sample data set, a second type of sample data matching the target features is constructed; and an initial second prediction model based on TCN is constructed; wherein the initial second prediction model at least includes an initial second input layer, an initial TCN layer, and an initial second output layer; Using the second type of sample data to train an initial second prediction model to obtain a second prediction model that meets the requirements; wherein the second prediction model is used to predict the wind power of the wind power station; Among them, according to the preset processing rules, the shap value of the alternative feature is determined using the first prediction model, including: according to the preset processing rules, using the sample data set to construct test data about the alternative feature; by modifying at least one alternative feature in the test data, the modified data is obtained; and the test data and the modified data are combined to establish a test data set; using the first prediction model to process the test data set to obtain corresponding prediction results of multiple situation types; based on the prediction results of multiple situation types, the shap value of the alternative feature is determined.

8. A computer-readable storage medium, characterized in that: Computer instructions are stored thereon, and when the instructions are executed by a processor, the steps of the method described in any one of claims 1 to 5 or 7 are implemented.

Images