Wind power cluster power day-ahead prediction method and device based on causal relationship analysis

Abstract

The invention discloses a wind power cluster power day-ahead prediction method and device based on causal relationship analysis, and relates to the field of wind power prediction.The method comprises the steps that NWP wind speed features of all stations in a wind power cluster are screened by setting a causal relationship algorithm, and an available feature set is obtained; constructing a weight distribution mode according to the installed capacity information and the meteorological information; performing weight distribution on the available feature set according to the weight distribution mode to obtain an effective feature set; constructing a causal sensitive scene and a causal loss function based on the causal relationship; nesting a causal loss function to a set position of the prediction model, and constructing an improved prediction model; performing mixed training on the improved prediction model through a causal sensitive scene and a set scene; and inputting the effective feature set into the trained improved prediction model for prediction processing to obtain a wind power cluster power prediction result. The method can improve the prediction accuracy, can reflect the dynamic characteristics of the system, and tracks the future power trend.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and particularly to a day-ahead prediction method and device for wind power cluster power based on causal relationship analysis. Background Art

[0002] In recent years, driven by policy support and a significant reduction in the costs of solar photovoltaic power generation and wind power generation, the power generation capacity has increased rapidly. The application of wind energy in the international energy landscape shows a large-scale growth trend, and it remains the greatest potential for increasing the global renewable energy installed capacity. The International Energy Agency (IEA) reported that the wind power generation increased by a record 26.5 billion kWh (a 14% increase) in 2022, reaching over 210 billion kWh. China continues to lead in the new wind power generation capacity, adding 37 GW in 2022, of which 7,000 MW was added to offshore wind farms.

[0003] Currently, due to the intermittency, volatility, and randomness of wind energy, the large-scale grid connection of wind power poses a great safety challenge to the power system. Therefore, as a necessary means to ensure the safe and stable operation of the power system, wind power prediction needs to further improve the prediction accuracy. In wind power cluster power prediction, the high-dimensional information input composed of multi-site characteristics will provide redundant information to the prediction model. Some studies achieve the effect of feature processing through some correlation-based feature processing methods. However, with the increase in system complexity, correlation analysis is difficult to meet the modeling requirements. There are not only direct interactions between multiple variables, but also indirect interactions bridged by intermediate variables, and the influence relationship usually has asymmetry. Traditional correlation analysis methods are difficult to handle indirect relationships.

[0004] Therefore, how to invent a wind power cluster power prediction method that takes into account causal relationship analysis, improves the prediction accuracy, reflects the dynamic characteristics of the system, and tracks the future power trend has become an urgent problem to be solved. Summary of the Invention

[0005] To this end, the present invention provides a day-ahead prediction method and device for wind power cluster power based on causal relationship analysis, which embeds causal relationships in all stages of wind power prediction, can more reasonably select features and divide weights, more carefully and dynamically divide different weather scenarios, and improve the prediction accuracy. At the same time, it can also reflect the dynamic characteristics of the system and track the future power trend.

[0006] To achieve the above object, the present invention provides the following technical solutions: A day-ahead prediction method for wind power cluster power based on causal relationship analysis, including: Screening the NWP wind speed characteristics of each site in the wind power cluster through a set causal relationship algorithm to obtain an available feature set; Construct a weight allocation method based on the installed capacity information and meteorological information; allocate weights to the available feature set according to the weight allocation method to obtain an effective feature set; Construct a causality-sensitive scenario and a causality loss function based on the causal relationship; Nest the causality loss function into a set position of the prediction model to construct an improved prediction model; perform hybrid training on the improved prediction model through the causality-sensitive scenario and the set scenario to obtain a trained improved prediction model; Input the effective feature set into the trained improved prediction model for prediction processing to obtain a wind power cluster power prediction result.

[0007] As an optimal solution of the day-ahead prediction method for wind power cluster power considering causal relationship analysis, in the process of screening the NWP wind speed characteristics of each station in the wind power cluster through the set causal relationship algorithm, select Granger and CCM as the basic calculation methods; the screening steps of the available feature set are as follows: Obtain the P value and F-test value through Granger calculation, and judge the initial causal intensity threshold through the P value and the F-test value; According to the initial causal intensity threshold, combine the P value and the F-test value with the CCM value respectively to perform a preliminary screening on the NWP wind speed characteristics of each station in the wind power cluster to obtain a preliminary screening result; Determine the upper limit of the number of retained features according to the P value; Determine the final causal intensity threshold through the F-test value and the CCM value; obtain the available feature set according to the final causal intensity threshold.

[0008] As an optimal solution of the day-ahead prediction method for wind power cluster power considering causal relationship analysis, the expression for performing a preliminary screening on the NWP wind speed characteristics of each station in the wind power cluster by combining the P value and the F-test value with the CCM value respectively is: ; ; ; In the formula, is the feature selection result based on the Granger P value; is the feature selection result based on the Granger F-test value; is the feature selection result based on the CCM value; is the result of CCM; is a constant; is a constant; According to the final causal strength threshold, the expression for obtaining the available feature set is: ; In the formula, is the final selection result.

[0009] As an optimal solution of the day-ahead wind power cluster power prediction method based on causal relationship analysis, in the process of constructing the weight allocation method according to the installed capacity information and the meteorological information, the NWP wind speed feature of the wind power cluster in the available feature set is used as the initial feature group; the steps of constructing the weight allocation method are as follows: When considering weather factors, based on the NWP forecast wind speed of the wind farms corresponding to the initial feature group, the CCM causal score of each feature is used as the weight index; the values of the CCM causal score higher than the set threshold are extracted and normalized to the weather factor weight. When considering the capacity factor weight, based on the installed capacity of the wind farms corresponding to the initial feature group, the proportion of the installed capacity of each wind farm is used as the weight index; the proportion of the installed capacity is normalized to the capacity factor weight. The weather factor weight and the capacity factor weight are combined to construct the influence weight of each wind farm on the entire wind power cluster.

[0010] As an optimal solution of the day-ahead wind power cluster power prediction method based on causal relationship analysis, the expression of the weight allocation method is: ; ; ; In the formula, is the weight of the i th feature based on the installed capacity information; is the causal weight of the i th feature based on the weather information; is the final weight of the i th feature; is the installed capacity of the station corresponding to the i th feature; is the total installed capacity; is the i th feature selected by the feature selection method of the present invention; C is a constant threshold, set to 0.05; n is the feature cluster.

[0011] As an optimal solution to the method for the day-ahead prediction of the power of a wind power cluster considering causal relationship analysis, the causal loss function is the transfer entropy causal relationship loss function; the expression of the transfer entropy causal relationship loss function is: ; In the formula, is the transfer entropy from y to x; predicted output value; is the actual value; is the corresponding probability distribution.

[0012] As an optimal solution to the method for the day-ahead prediction of the power of a wind power cluster considering causal relationship analysis, error analysis is performed on the prediction results through the root mean square error, mean absolute error, and correlation coefficient; The expression of the root mean square error is: ; In the formula, RMSE is the root mean square error; is the actual value; is the predicted value; is the average value of the actual values in the corresponding time period; is the average value of the predicted values in the corresponding time period; is the installed capacity of the corresponding wind farm; n is the number of samples in the corresponding prediction time period; The expression of the mean absolute error is: ; In the formula, MAE is the mean absolute error; The expression of the correlation coefficient is: ; In the formula, R is the correlation coefficient; is the average value of the actual values in the corresponding time period; is the average value of the predicted values in the corresponding time period.

[0013] The present invention also provides a device for the day-ahead prediction of the power of a wind power cluster considering causal relationship analysis. Based on the above method for the day-ahead prediction of the power of a wind power cluster considering causal relationship analysis, it includes: A feature screening module, configured to screen the NWP wind speed features of each station in the wind power cluster through a set causal relationship algorithm to obtain an available feature set; A weight allocation method construction and processing module, configured to construct a weight allocation method according to the installed capacity information and meteorological information; and allocate weights to the available feature set according to the weight allocation method to obtain an effective feature set; Causal sensitivity scenario and causal loss function construction module, used to construct a causal sensitivity scenario and a causal loss function based on causal relationships; Improved prediction model construction and training module, used to nest the causal loss function into a set position of the prediction model to construct an improved prediction model; perform hybrid training on the improved prediction model through the causal sensitivity scenario and the set scenario to obtain a trained improved prediction model; Wind power cluster power prediction result acquisition module, used to input the effective feature set into the trained improved prediction model for prediction processing to obtain a wind power cluster power prediction result.

[0014] As an optimal solution for the wind power cluster power day-ahead prediction device based on causal relationship analysis, in the feature screening module, during the process of screening the NWP wind speed features of each power station in the wind power cluster through the set causal relationship algorithm, Granger and CCM are selected as the basic calculation methods; the screening sub-module of the available feature set includes: P-value and F-test value calculation sub-module, used to calculate the P-value and F-test value through Granger and judge the initial causal intensity threshold through the P-value and the F-test value; Feature preliminary screening sub-module, used to combine the P-value and the F-test value with the CCM value respectively according to the initial causal intensity threshold to preliminarily screen the NWP wind speed features of each power station in the wind power cluster to obtain a preliminary screening result; Retained feature quantity upper limit determination sub-module, used to determine the upper limit of the retained feature quantity according to the P-value; Available feature set acquisition sub-module, used to determine the final causal intensity threshold through the F-test value and the CCM value; obtain the available feature set according to the final causal intensity threshold.

[0015] As an optimal solution for the wind power cluster power day-ahead prediction device based on causal relationship analysis, in the feature preliminary screening sub-module of the feature screening module, the expression for preliminarily screening the NWP wind speed features of each power station in the wind power cluster by combining the P-value and the F-test value with the CCM value respectively is: ; ; ; In the formula, is the feature selection result based on the Granger P-value; is the feature selection result based on the Granger F-test value; is the feature selection result based on the CCM value; is the result of CCM; is a constant; is a constant; According to the final causal strength threshold, the expression for obtaining the available feature set is: ; In the formula, is the final selection result.

[0016] As an optimal solution of the wind power cluster power day-ahead prediction device based on causal relationship analysis, in the weight assignment method construction and processing module, in the process of constructing the weight assignment method according to the installed capacity information and the meteorological information, the NWP wind speed feature of the wind power cluster in the available feature set is used as the initial feature group; the sub-module for constructing the weight assignment method includes: The weather factor weight acquisition sub-module is used to, when considering weather factors, use the NWP forecast wind speed of the wind farm corresponding to the initial feature group as the benchmark, and use the CCM causal score of each feature as the weight index; extract the values of the CCM causal score higher than the set threshold and normalize them into weather factor weights; The capacity factor weight acquisition sub-module is used to, when considering the capacity factor weight, use the installed capacity of the wind farm corresponding to the initial feature group as the benchmark, and use the proportion of the installed capacity of each wind farm as the weight index; normalize the proportion of the installed capacity into capacity factor weights; The wind farm influence weight construction sub-module is used to combine the weather factor weight and the capacity factor weight to construct the influence weight of each wind farm on the entire wind power cluster.

[0017] As an optimal solution of the wind power cluster power day-ahead prediction device based on causal relationship analysis, in the weight assignment method construction and processing module, the expression of the weight assignment method is: ; ; ; In the formula, is the weight of the i th feature based on the installed capacity information; is the causal weight of the i th feature based on the weather information; is the final weight of the i th feature; is the installed capacity of the station corresponding to the i th feature; is the total installed capacity; is the ia feature; C is a constant threshold, set to 0.05; n is a feature cluster.

[0018] As an optimal solution of the device for predicting the daily power of a wind power cluster based on causal relationship analysis, in the causal sensitive scenario and causal loss function construction module, the causal loss function is the transfer entropy causal relationship loss function; the expression of the transfer entropy causal relationship loss function is: ; In the formula, is the transfer entropy from y to x; predicted output value; is the actual value; is the corresponding probability distribution.

[0019] As an optimal solution of the device for predicting the daily power of a wind power cluster based on causal relationship analysis, in the module for obtaining the prediction result of the wind power cluster power, error analysis of the prediction result is performed through the root mean square error, mean absolute error, and correlation coefficient; The expression of the root mean square error is: ; In the formula, RMSE is the root mean square error; is the actual value; is the predicted value; is the average value of the actual values in the corresponding time period; is the average value of the predicted values in the corresponding time period; is the installed capacity of the corresponding wind farm; n is the number of samples in the corresponding prediction time period; The expression of the mean absolute error is: ; In the formula, MAE is the mean absolute error; The expression of the correlation coefficient is: ; In the formula, R is the correlation coefficient; is the average value of the actual values in the corresponding time period; is the average value of the predicted values in the corresponding time period.

[0020] The present invention has the following advantages: The present invention screens the NWP wind speed characteristics of each power station in the wind power cluster by setting a causal relationship algorithm to obtain an available feature set; constructs a weight allocation method according to the installed capacity information and meteorological information; allocates weights to the available feature set according to the weight allocation method to obtain an effective feature set; constructs a causal sensitive scenario and a causal loss function based on the causal relationship; nests the causal loss function into a set position of a prediction model to construct an improved prediction model; performs hybrid training on the improved prediction model through the causal sensitive scenario and a set scenario to obtain a trained improved prediction model; inputs the effective feature set into the trained improved prediction model for prediction processing to obtain a wind power cluster power prediction result. The present invention proposes a solution starting from the causal perspective for the problem of improving the accuracy of the current day-ahead time-scale wind power cluster power prediction, and embeds the causal relationship in each stage of the wind power prediction. In the feature engineering stage, the present invention uses a feature selection method of compound causal relationship (CompoundCausal Feature Selection, CCFS); in terms of the weights of features, through CI and a combined application method, and adding AM, the fitting ability of the model to features is improved; in the prediction stage, a causal relationship-based loss function TELoss is designed, and a causal sensitive scenario is constructed. Thereby, the accuracy of the wind power prediction is improved, and the effectiveness and universality of the present invention are also verified. Description of the Drawings

[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only exemplary, and for those of ordinary skill in the art, without creative efforts, other implementation drawings can also be obtained according to the provided drawings.

[0022] The structures, ratios, sizes, etc. illustrated in this specification are only used to cooperate with the content disclosed in the specification for those familiar with this technology to understand and read, and are not used to limit the limited conditions under which the present invention can be implemented. Therefore, they do not have a substantial technical meaning. Any modification of the structure, change in the proportional relationship, or adjustment of the size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope covered by the technical content disclosed in the present invention.

[0023] Figure 1 It is a schematic flow chart of the day-ahead prediction method for wind power cluster power based on considering causal relationship analysis provided in Embodiment 1 of the present invention; Figure 2 It is a schematic specific implementation flow chart of the day-ahead prediction method for wind power cluster power based on considering causal relationship analysis provided in Embodiment 1 of the present invention; Figure 3 It is a schematic diagram of the CCFS process in the wind power cluster power day-ahead prediction method based on causal relationship analysis provided in Embodiment 1 of the present invention; Figure 4 It is a schematic diagram for predicting and comparing each feature selection method in a possible embodiment provided in Embodiment 1 of the present invention; Figure 5 It is a schematic diagram for predicting and comparing different weight construction methods in a possible embodiment provided in Embodiment 1 of the present invention; Figure 6 It is a schematic diagram for predicting and comparing based on AM in a possible embodiment provided in Embodiment 1 of the present invention; Figure 7 It is a schematic diagram for predicting and comparing the loss function in the causal sensitivity scenario in a possible embodiment provided in Embodiment 1 of the present invention; Figure 8 It is a schematic diagram of the architecture of the wind power cluster power day-ahead prediction device based on causal relationship analysis provided in Embodiment 2 of the present invention. Specific Embodiments

[0024] The following specific embodiments illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0025] Embodiment 1

[0026] Refer to Figure 1 and Figure 2 , Embodiment 1 of the present invention provides a wind power cluster power day-ahead prediction method based on causal relationship analysis, including the following steps: S1. Screen the NWP wind speed characteristics of each power station in the wind power cluster through a set causal relationship algorithm to obtain an available feature set; S2. Construct a weight allocation method according to the installed capacity information and meteorological information; allocate weights to the available feature set according to the weight allocation method to obtain an effective feature set; S3. Construct a causal sensitivity scenario and a causal loss function based on causal relationship; S4. Nest the causal loss function into a set position of the prediction model to construct an improved prediction model; perform hybrid training on the improved prediction model through the causal sensitivity scenario and the set scenario to obtain a trained improved prediction model; S5. Input the effective feature set into the trained improved prediction model for prediction processing to obtain the wind power cluster power prediction result.

[0027] In this embodiment, in step S1, the NWP wind speed features of each power station in the wind power cluster are screened by setting a causal relationship algorithm to obtain an available feature set. Specifically, by combining Granger and CCM causality tests, a CCFS feature selection method is constructed, and for the causal relationship between the NWP wind speed features of each power station in the cluster and the total power of the cluster, an available feature set is selected. Specifically, as Figure 3 shown, Granger and CCM are selected as the basic calculation methods; the screening steps of the available feature set are as follows: S11. Calculate the P value and F-test value through Granger, and judge the initial causal intensity threshold through the P value and the F-test value. Specifically, use Granger to calculate the P value and F-test value to judge the causal intensity threshold.

[0028] S12. According to the initial causal intensity threshold, combine the P value and the F-test value with the CCM value respectively to preliminarily screen the NWP wind speed features of each power station in the wind power cluster to obtain a preliminary screening result. Specifically, the expression for preliminary screening is: ; ; ; In the formula, is the feature selection result based on the Granger P value; is the feature selection result based on the Granger F-test value; is the feature selection result based on the CCM value; is the result of CCM; is a constant; is a constant; S13. Determine the upper limit of the number of features to be retained according to the P value. Specifically, based on the P value in the Granger causality test, determine the upper limit of the number of features to be retained. S14. Determine the final causal intensity threshold through the F-test value and the CCM value; obtain the available feature set according to the final causal intensity threshold.

[0029] Specifically, by further determining a threshold for the F-test value and the CCM causal strength, the optimal feature combination can be retained, and the available feature set can be obtained. The expression for obtaining the available feature set is: ; In the formula, is the final selection result.

[0030] In this embodiment, in step S2, according to the installed capacity information and the meteorological information, a weight allocation method is constructed; according to the weight allocation method, the available feature set is weighted to obtain an effective feature set; Specifically, by comprehensively considering the influence of weather factors and capacity factors on the output of the wind power cluster, an influence weight of each wind farm on the overall wind power cluster is constructed, and this weight is applied to feature input, feature construction, and model construction.

[0031] Specifically, in the process of constructing the weight allocation method according to the installed capacity information and the meteorological information, the NWP wind speed feature of the wind power cluster in the available feature set is used as the initial feature group; the steps for constructing the weight allocation method are as follows: S21. When considering weather factors, based on the NWP forecast wind speed of the wind farm corresponding to the initial feature group, and using the CCM causal score of each feature as the weight index; extract the values of the CCM causal score higher than the set threshold and normalize them into weather factor weights; Among them, in order to better reflect the difference in CI between NWP wind speed features, 0.5 is selected as the lower threshold.

[0032] Specifically, the expression for the weather factor weight is: ; In the formula, is the causal weight of the i th feature based on weather information; is the installed capacity of the i th feature corresponding station; is the total installed capacity; is the i th feature selected by the feature selection method of the present invention; C is a constant threshold, set to 0.05; n is the feature cluster.

[0033] S22. When considering the capacity factor weight, based on the installed capacity of the wind farm corresponding to the initial feature group, and using the proportion of the installed capacity of each wind farm as the weight index; normalize the proportion of the installed capacity into the capacity factor weight; Specifically, the expression for the capacity factor weight is: ; In the formula, is the weight of the i th feature based on the installed capacity information.

[0034] S23. Combine the weather factor weight and the capacity factor weight to construct the influence weight of each wind farm on the entire wind power cluster.

[0035] Specifically, the expression of the influence weight of the wind farm is: ; In the formula, is the final weight of the i th feature.

[0036] In this embodiment, the influence weight of each wind farm on the entire wind power cluster is applied to the wind speed characteristics of each wind farm, and then a new representative feature is constructed to represent the wind speed change characteristics of the entire wind farm cluster. Finally, the constructed new wind speed characteristics are applied to different scenarios to verify their effectiveness. Alternatively, each wind farm is predicted separately, and the prediction results are superimposed with the weights to obtain the total prediction result of the cluster.

[0037] In this embodiment, in step S3, a causal sensitive scenario and a causal loss function are constructed based on the causal relationship; wherein, the causal loss function is a transfer entropy (TE) causal relationship loss function; the expression of the transfer entropy causal relationship loss function is: ; In the formula, is the transfer entropy from y to x; predicted output value; is the actual value; is the corresponding probability distribution.

[0038] In this embodiment, in step S4, the causal loss function is nested at the set position of the prediction model to construct an improved prediction model; the improved prediction model is hybrid-trained through the causal sensitive scenario and the set scenario to obtain a trained improved prediction model; Specifically, the causal loss function is nested at the end of the prediction model as the loss function to control the fitting effect of the output result of the prediction model.

[0039] For the causal sensitive scenario, TELoss is adopted, and for other scenarios, MSELoss is adopted. The improved prediction model is hybrid-trained with the hybrid scenario to obtain a trained improved prediction model.

[0040] In this embodiment, in step S5, the effective feature set is input into the trained improved prediction model for prediction processing to obtain the wind power cluster power prediction result.

[0041] Specifically, the error analysis of the prediction result is carried out through the root mean square error, mean absolute error and correlation coefficient; The expression of the root mean square error is: ; In the formula, RMSE is the root mean square error; is the actual value; is the predicted value; is the average value of the actual values in the corresponding time period; is the average value of the predicted values in the corresponding time period; is the installed capacity of the corresponding wind farm; n is the number of samples in the corresponding prediction time period; The expression of the mean absolute error is: ; In the formula, MAE is the mean absolute error; The expression of the correlation coefficient is: ; In the formula, R is the correlation coefficient; is the average value of the actual values in the corresponding time period; is the average value of the predicted values in the corresponding time period.

[0042] In a possible embodiment, a photovoltaic power day-ahead prediction example is provided as follows: To improve the wind power prediction accuracy to the greatest extent, this embodiment selects multiple prediction models to perform day-ahead prediction on the photovoltaic power of the test set. The data set of a wind power cluster in Jilin Province, China is used. This data set contains the wind power measurement data and NWP data of a real wind power cluster with 50 wind farms, including meteorological factors (wind speed) (updated every 24 hours, 24-hour forecast in advance). The research objective of this embodiment is to predict the wind power for the next 1 day using wind power data and numerical weather prediction data. The total installed capacity is 5820 MW respectively. From 00:00 on January 1, 2021 to 23:45 on June 30, 2021, there are 180 days in total, and the time interval is 15 minutes. The first 5 months are used as the training set, and the 6th month is used as the test set.

[0043] T1. The composite causal relationship (CCFS) method is applied to select effective features. As Figure 3As shown in the figure, first, the Granger method is used to calculate the P-value and F-test value to determine the causal intensity threshold, which is then combined with the CCM calculation results respectively to preliminarily screen the NWP wind speed characteristics of each station in the cluster. Based on the P-value in the Granger causality test, the upper limit of the number of retained features is determined. Finally, by further determining the threshold for the F-test value and the CCM causal intensity, the optimal feature combination is retained.

[0044] To effectively verify the effectiveness of the feature selection of the present invention, the present invention uses GP, GF, CCM, and Pearson to calculate the CC or CI between the wind speeds of 50 stations and the total cluster power, and shows the prediction evaluation indicators of each calculation method after each feature selection method.

[0045] Table 1 Effectiveness of different feature selection methods

[0046] As shown in Table 1, the prediction evaluation indicators after each feature selection method are shown. Compared with the method of inputting the NWP wind speed as a whole and the Pearson calculation method, the feature selection methods based on causal relationship judgment can all obtain relatively low error indicators. After averaging the NWP of multiple stations, certain fluctuation information will be lost, which will lead to smoother prediction. Compared with inputting the NWP of all stations at the same time, there is a certain improvement in accuracy. The CCFS method proposed by the present invention has achieved the best prediction accuracy. Compared with other comparison methods, the RMSE has been reduced by 0.9 on average, which also shows the correctness of the composite causal feature selection idea of the present invention.

[0047] As Figure 4 shown, the prediction legends of each feature selection method are shown. The prediction curve of CCFS more effectively tracks the change trend of the measured power curve. Taking the high-output scenario as an example, the best prediction effect can also be obtained at the high-output level.

[0048] T2. Construct a weight allocation method through feature combination. Based on the features selected by the above CCFS, the present invention conducts experiments on weight construction to illustrate the effectiveness of the feature weights constructed in the present invention example. First, this embodiment defines three weight application methods, namely: Method A: After weighting according to the constructed weights, new features are constructed for input. Method B: After constructing features, they are input simultaneously with the original feature set. Method C: Weighting is carried out after prediction.

[0049] Table 2 Evaluation indicators of weight construction and application methods

[0050] As shown in Table 2, for the three weight construction methods, whether considering the causal relationship strength between the NWP wind speed of the station and the cluster power alone to construct the weight, or considering the installed capacity of the station alone to construct the weight, the prediction accuracy is relatively low. When comprehensively considering the causal relationship strength between the NWP wind speed of the station and the cluster power and the weight constructed by the installed capacity of the unit, the error evaluation index has certain improvement compared with the other two construction methods. The prediction legend is as Figure 5 shown. Compared with other methods, the proposed weight construction and application method enables the model to learn the input features more effectively and simultaneously tracks the change trend of the actual curve better.

[0051] To further enable the model to effectively learn features, the present invention uses an attention mechanism to further realize the distribution of feature weights in the prediction model. The present invention adds AM to the prediction model and conducts ablation experiments. Method D: Only perform prediction after using CCFS. Method E: Only use AM. Method F: Only use AM. Method G: Perform prediction using feature selection and AM. Method H: Perform prediction using CCFS, weight construction, and Attention mechanism fusion.

[0052] Table 3 Evaluation indicators for applying the attention mechanism

[0053] As shown in Table 3, directly applying the attention mechanism to the NWP wind speed input model of 50 stations, the obtained prediction accuracy is comparable to the prediction effect of the composite feature screening + weighted feature combination of the present invention. The prediction accuracy obtained by combining the attention mechanism with the screened features has a significant decline instead. However, when the composite feature screening + weighted feature combination method of the present invention is input into the LSTM benchmark model containing the attention mechanism, the prediction accuracy can still be improved to a certain extent, and the prediction accuracy at this time is the highest among all current prediction processes. This shows that the features constructed by the weight of the present invention bring more effective feature information, and the gain brought by this part of the additional information is more important than the information loss caused by feature elimination. As Figure 6 shown, the prediction based on AM effectively learns the weight construction method of the present invention. In the case where the above prediction methods are generally high, through the learning mechanism of AM, the model prediction curve can effectively track the change trend of the measured power while being closer to the measured power curve, effectively improving the prediction accuracy.

[0054] T3, TELoss, and Causal Sensitivity Scenarios. The present invention further improves the loss function of the existing causality-based regression method, constructs causal sensitivity scenarios for wind power prediction, with the expectation of further improving the prediction accuracy. It is compared with common regression loss functions (MSE, MAE), and the proposed causal sensitivity scenarios are further applied to improve the training effect to verify the effectiveness of the proposed loss function and causal sensitivity.

[0055] Table 4 Evaluation Metrics of Loss Functions and Scenarios

[0056] As shown in Table 4, by applying the causal sensitivity scenarios divided by the present invention to the prediction process with TE as the loss function, when dividing the causal sensitivity scenarios, the prediction accuracy is better than that without causal sensitivity scenarios. Among them, the prediction accuracy of the prediction model with causal effect sensitivity scenarios is significantly improved. It verifies that the method for dividing causal effect sensitivity scenarios proposed by the present invention can effectively improve the prediction accuracy when combined with the causal effect loss function.

[0057] As Figure 7 shown, the prediction legends obtained by using MSE, MAE, TELoss proposed by the present invention, and combining the loss functions under different scenarios. Figure 7 Clearly shows that the present invention can more effectively track the change trend of the actual power curve. In Figure 7 , the present invention also more detailedly shows the prediction effects of each loss function in the causal sensitivity scenarios. Compared with MSE and MAE, TELoss can more effectively make the prediction curve more consistent with the actual power curve in the causal sensitivity scenarios, and effectively track its change trend during the fluctuation period, better fitting the change trends of peaks and valleys. It also verifies that the combination of TELoss in the causal sensitivity scenarios proposed by the present invention and MSELoss in the conventional output scenarios can significantly improve the prediction effect.

[0058] T4. Verify the effectiveness through different prediction models. To further verify the universality of the proposed prediction method, this embodiment selects 6 commonly used prediction models in wind power prediction: basic model (LSTM), GRU, CNN, CNN-LSTM, BiLSTM, Transformer, CC, STDGCNN, and CCTransformer for prediction.

[0059] Table 5 Evaluation Metrics of Different Prediction Models

[0060] As shown in Table 5, they are the evaluation indicators of the prediction process of whether to adopt the proposed method under each prediction model. Compared with the prediction model without using the present invention, after considering the embedded causal relationship in the whole prediction process, the RMSE is reduced by 2.3% on average, and the MAE is reduced by 1.9% on average, verifying the effectiveness and universality of the present invention.

[0061] In summary, the present invention screens the NWP wind speed characteristics of each station in the wind power cluster by setting a causal relationship algorithm to obtain an available feature set; constructs a weight allocation method according to the installed capacity information and meteorological information; allocates weights to the available feature set according to the weight allocation method to obtain an effective feature set; constructs a causal sensitive scenario and a causal loss function based on the causal relationship; nests the causal loss function into the set position of the prediction model to construct an improved prediction model; performs hybrid training on the improved prediction model through the causal sensitive scenario and the set scenario to obtain a trained improved prediction model; inputs the effective feature set into the trained improved prediction model for prediction processing to obtain the wind power cluster power prediction result. Aiming at the problem of improving the accuracy of wind power cluster power prediction at the current day-ahead time scale, the present invention proposes a solution starting from the causal perspective and embeds causal relationships at each stage of wind power prediction. In the feature engineering stage, the present invention uses a feature selection method of compound causal relationship (Compound Causal Feature Selection, CCFS); in terms of the weights of features, through CI and combined application methods, and adding AM, the fitting ability of the model to features is improved; in the prediction stage, a causal relationship-based loss function TELoss is designed and a causal sensitive scenario is constructed. Thereby, the accuracy of wind power prediction is improved, and the effectiveness and universality of the present invention are also verified.

[0062] It should be noted that the method of the embodiments of the present disclosure can be executed by a single device, such as a computer or a server. The method of this embodiment can also be applied to a distributed scenario and completed by the cooperation of multiple devices. In such a distributed scenario, one of the multiple devices can only execute one or more steps of the method of the embodiments of the present disclosure, and these multiple devices will interact with each other to complete the described method.

[0063] It should be noted that some embodiments of the present disclosure have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be executed in a different order than in the above embodiments and still achieve the desired results. Additionally, the processes depicted in the figures do not necessarily require the specific order or sequential order shown to achieve the desired result. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0064] Embodiment 2

[0065] Refer to Figure 8 , Embodiment 2 of the present invention also provides a day-ahead wind power cluster power prediction device based on causal relationship analysis, including: Feature screening module 001, configured to screen the NWP wind speed features of each power station in the wind power cluster through a set causal relationship algorithm to obtain an available feature set; Weight assignment method construction and processing module 002, configured to construct a weight assignment method according to the installed capacity information and meteorological information; perform weight assignment on the available feature set according to the weight assignment method to obtain an effective feature set; Causal sensitive scenario and causal loss function construction module 003, configured to construct a causal sensitive scenario and a causal loss function based on causal relationship; Improved prediction model construction and training module 004, configured to nest the causal loss function into a set position of the prediction model to construct an improved prediction model; perform hybrid training on the improved prediction model through the causal sensitive scenario and the set scenario to obtain a trained improved prediction model; Wind power cluster power prediction result acquisition module 005, configured to input the effective feature set into the trained improved prediction model for prediction processing to obtain a wind power cluster power prediction result.

[0066] In this embodiment, in the feature screening module 001, in the process of screening the NWP wind speed features of each power station in the wind power cluster through the set causal relationship algorithm, Granger and CCM are selected as the basic calculation methods; the screening sub-module of the available feature set includes: P-value and F-test value calculation sub-module 011, configured to obtain a P-value and an F-test value through Granger calculation, and judge an initial causal intensity threshold through the P-value and the F-test value; Feature preliminary screening sub-module 012, configured to combine the P-value and the F-test value with the CCM value respectively according to the initial causal intensity threshold to perform preliminary screening on the NWP wind speed features of each power station in the wind power cluster to obtain a preliminary screening result; Retained feature quantity upper limit determination sub-module 013, configured to determine the upper limit of the retained feature quantity according to the P-value; Available feature set acquisition sub-module 014, configured to determine a final causal intensity threshold through the F-test value and the CCM value; obtain the available feature set according to the final causal intensity threshold.

[0067] In this embodiment, in the preliminary feature screening sub-module 012 of the feature screening module 001, the expression for preliminarily screening the NWP wind speed features of each power station in the wind power cluster by combining the P value and the F-test value with the CCM value respectively is as follows: ; ; ; In the formula, is the feature selection result based on the Granger P value; is the feature selection result based on the Granger F-test value; is the feature selection result based on the CCM value; is the result of CCM; is a constant; is a constant; According to the final causal strength threshold, the expression for obtaining the available feature set is: ; In the formula, is the final selection result.

[0068] In this embodiment, in the weight allocation method construction and processing module 002, in the process of constructing the weight allocation method according to the installed capacity information and the meteorological information, the NWP wind speed features of the wind power cluster in the available feature set are used as the initial feature group; the sub-modules for constructing the weight allocation method include: The weather factor weight acquisition sub-module 021 is used to, when considering weather factors, take the NWP forecast wind speed of the wind farm corresponding to the initial feature group as the benchmark, and use the CCM causal score of each feature as the weight index; extract the values of the CCM causal score higher than the set threshold and normalize them into weather factor weights; The capacity factor weight acquisition sub-module 022 is used to, when considering the capacity factor weight, take the installed capacity of the wind farm corresponding to the initial feature group as the benchmark, and use the proportion of the installed capacity of each wind farm as the weight index; normalize the proportion of the installed capacity into the capacity factor weight; The wind farm influence weight construction sub-module 023 is used to combine the weather factor weight and the capacity factor weight to construct the influence weight of each wind farm on the entire wind power cluster.

[0069] In this embodiment, in the weight allocation method construction and processing module 002, the expression of the weight allocation method is: ; ; ; In the formula, is the weight of the i th feature based on the installed capacity information; is the causal weight of the i th feature based on the weather information; is the final weight of the i th feature; is the installed capacity of the station corresponding to the i th feature; is the total installed capacity; is the i th feature selected by the feature selection method of the present invention; C is a constant threshold, set to 0.05; n is the feature cluster.

[0070] In this embodiment, in the causal sensitive scenario and causal loss function construction module 003, the causal loss function is the transfer entropy causal relationship loss function; the expression of the transfer entropy causal relationship loss function is: ; In the formula, is the transfer entropy from y to x; predicted output value; is the actual value; is the corresponding probability distribution.

[0071] In this embodiment, in the wind power cluster power prediction result acquisition module 005, the error analysis of the prediction result is performed by the root mean square error, mean absolute error and correlation coefficient; The expression of the root mean square error is: ; In the formula, RMSE is the root mean square error; is the actual value; is the predicted value; is the average value of the actual values in the corresponding time period; is the average value of the predicted values in the corresponding time period; is the installed capacity of the corresponding wind farm; n is the number of samples in the corresponding prediction time period; The expression of the mean absolute error is: ; In the formula, MAE is the mean absolute error; The expression of the correlation coefficient is: ; In the formula,R is the correlation coefficient; is the average value of the actual values in the corresponding time period; is the average value of the predicted values in the corresponding time period.

[0072] It should be noted that the information interaction, execution process, etc. among the above-mentioned system modules, due to being based on the same concept as the method embodiment in Embodiment 1 of the present application, have the same technical effects as the method embodiment of the present application. For specific content, reference can be made to the description in the method embodiment shown above in the present application, and details will not be elaborated here.

[0073] Embodiment 3 Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium, in which program codes for a day-ahead wind power cluster power prediction method based on causal relationship analysis are stored. The program codes include instructions for executing the day-ahead wind power cluster power prediction method based on causal relationship analysis in Embodiment 1 or any possible implementation manner thereof.

[0074] The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center integrating one or more available media. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium (for example, a solid state disk (SSD)).

[0075] Embodiment 4 Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor; The processor and the memory complete communication with each other through a bus; the memory stores program instructions executable by the processor, and the processor can execute the day-ahead wind power cluster power prediction method based on causal relationship analysis in Embodiment 1 or any possible implementation manner thereof by invoking the program instructions.

[0076] Specifically, the processor can be implemented by hardware or by software. When implemented by hardware, the processor can be a logic circuit, an integrated circuit, etc.; when implemented by software, the processor can be a general-purpose processor, which is implemented by reading software codes stored in the memory. The memory can be integrated in the processor or can exist independently outside the processor.

[0077] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present invention are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable systems. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from a website, computer, server, or data center to another website, computer, server, or data center by wire (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wirelessly (such as infrared, wireless, microwave, etc.).

[0078] Obviously, those skilled in the art should understand that the above-mentioned modules or steps of the present invention can be implemented by a general-purpose computing system. They can be concentrated on a single computing system or distributed on a network composed of multiple computing systems. Optionally, they can be implemented with program codes executable by the computing system. Thus, they can be stored in a storage system and executed by the computing system. And in some cases, the steps shown or described can be executed in a different order than here, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module to be implemented. In this way, the present invention is not limited to any specific combination of hardware and software.

[0079] Although the present invention has been described in detail above with general descriptions and specific embodiments, on the basis of the present invention, some modifications or improvements can be made, which are obvious to those skilled in the art. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of protection required by the present invention.

Claims

1. A method for predicting the daily power of a wind power cluster based on causal relationship analysis, characterized in that Including: Screen the NWP wind speed characteristics of each power station in the wind power cluster through setting a causal relationship algorithm to obtain an available feature set; Construct a weight allocation method according to the installed capacity information and meteorological information; allocate weights to the available feature set according to the weight allocation method to obtain an effective feature set; Construct a causal sensitive scenario and a causal loss function based on the causal relationship; Nest the causal loss function into the set position of the prediction model to construct an improved prediction model; perform hybrid training on the improved prediction model through the causal sensitive scenario and the set scenario to obtain a trained improved prediction model; Input the effective feature set into the trained improved prediction model for prediction processing to obtain the wind power cluster power prediction result.

2. The method for predicting the wind power cluster power on the day ahead based on considering causal relationship analysis according to claim 1, wherein In the process of screening the NWP wind speed characteristics of each power station in the wind power cluster through the set causal relationship algorithm, select Granger and CCM as the basic calculation methods; the screening steps of the available feature set are: Obtain the P value and F-test value through Granger calculation, and judge the initial causal intensity threshold through the P value and the F-test value; According to the initial causal intensity threshold, combine the P value and the F-test value with the CCM value respectively to preliminarily screen the NWP wind speed characteristics of each power station in the wind power cluster to obtain a preliminary screening result; Determine the upper limit of the number of retained features according to the P value; Determine the final causal intensity threshold through the F-test value and the CCM value; obtain the available feature set according to the final causal intensity threshold.

3. The method for predicting the wind power cluster power for the day-ahead based on causal relationship analysis according to claim 2, wherein The expression for preliminarily screening the NWP wind speed characteristics of each power station in the wind power cluster by combining the P value and the F-test value with the CCM value respectively is: ； ； ； In the formula, is the feature selection result based on the Granger P value; is the feature selection result based on the Granger F-test value; is the feature selection result based on the CCM value; is the result of CCM; is a constant; is a constant; The expression for obtaining the available feature set according to the final causal intensity threshold is: ； In the formula, is the final selection result.

4. The method for predicting the wind power cluster power on the day ahead based on causality analysis as claimed in claim 3, wherein In the process of constructing the weight allocation method according to the installed capacity information and the meteorological information, use the NWP wind speed characteristics of the wind power cluster in the available feature set as the initial feature group; the steps for constructing the weight allocation method are: When considering weather factors, use the NWP forecast wind speed of the wind farm corresponding to the initial feature group as the benchmark, and use the CCM causal score of each feature as the weight index; extract the values of the CCM causal score higher than the set threshold and normalize them as the weather factor weights; When considering the capacity factor weight, use the installed capacity of the wind farm corresponding to the initial feature group as the benchmark, and use the proportion of the installed capacity of each wind farm as the weight index; normalize the proportion of the installed capacity as the capacity factor weight; Combine the weather factor weight and the capacity factor weight to construct the influence weight of each wind farm on the entire wind power cluster.

5. The method for predicting the day-ahead power of a wind power cluster based on causal relationship analysis according to claim 4, wherein The expression of the weight allocation method is: ； ； ； Wherein, is the weight of the i th feature based on the installed capacity information; is the causal weight of the i th feature based on the weather information; is the final weight of the i th feature; is the installed capacity of the station corresponding to the i th feature; is the total installed capacity; is the i th feature selected by the feature selection method of the present invention; C is a constant threshold, set to 0.05; n is a feature cluster.

6. The method for predicting the day-ahead power of a wind power cluster based on causal relationship analysis according to claim 5, wherein The causal loss function is the transfer entropy causal relationship loss function; the expression of the transfer entropy causal relationship loss function is: ； In the formula, is the transfer entropy from y to x; is the predicted output value; is the actual value; is the corresponding probability distribution.

7. The method for predicting the wind power cluster power on the day ahead based on the analysis of causal relationship according to claim 6, wherein Perform error analysis on the prediction result through the root mean square error, mean absolute error and correlation coefficient; The expression of the root mean square error is: ； Wherein, RMSE is the root mean square error; is the actual value; is the predicted value; is the average value of the actual values in the corresponding time period; is the average value of the predicted values in the corresponding time period; is the installed capacity of the corresponding wind farm; n is the number of samples in the corresponding prediction time period; The expression of the mean absolute error is: ； Wherein, MAE is the mean absolute error; The expression of the correlation coefficient is: ； In the formula, R is the correlation coefficient; is the average value of the actual values in the corresponding time period; is the average value of the predicted values in the corresponding time period.

8. A wind power cluster power day-ahead prediction device based on causal relationship analysis, which adopts the wind power cluster power day-ahead prediction method based on causal relationship analysis according to any one of claims 1-7, characterized in that Including: A feature screening module, which is used to screen the NWP wind speed features of each power station in the wind power cluster through a set causal relationship algorithm to obtain an available feature set; A weight assignment method construction and processing module, which is used to construct a weight assignment method according to the installed capacity information and meteorological information; and assign weights to the available feature set according to the weight assignment method to obtain an effective feature set; A causal sensitive scenario and causal loss function construction module, which is used to construct a causal sensitive scenario and a causal loss function based on the causal relationship; An improved prediction model construction and training module, which is used to nest the causal loss function into the set position of the prediction model to construct an improved prediction model; and perform hybrid training on the improved prediction model through the causal sensitive scenario and the set scenario to obtain a trained improved prediction model; A wind power cluster power prediction result acquisition module, which is used to input the effective feature set into the trained improved prediction model for prediction processing to obtain the wind power cluster power prediction result.

9. The device for predicting the wind power cluster power on the day ahead based on the analysis of causality according to claim 8, wherein, In the feature screening module, during the process of screening the NWP wind speed features of each power station in the wind power cluster through the set causal relationship algorithm, Granger and CCM are selected as the basic calculation methods; the screening sub-module of the available feature set includes: A P-value and F-test value calculation sub-module, which is used to calculate the P-value and F-test value through Granger and judge the initial causal strength threshold through the P-value and the F-test value; A feature preliminary screening sub-module, which is used to combine the P-value and the F-test value with the CCM value respectively according to the initial causal strength threshold to preliminarily screen the NWP wind speed features of each power station in the wind power cluster to obtain a preliminary screening result; A retained feature quantity upper limit determination sub-module, which is used to determine the upper limit of the retained feature quantity according to the P-value; An available feature set acquisition sub-module, which is used to determine the final causal strength threshold through the F-test value and the CCM value; and obtain the available feature set according to the final causal strength threshold.

10. The device for predicting the day-ahead power of a wind power cluster based on causal relationship analysis according to claim 9, characterized in that, In the weight assignment method construction and processing module, during the process of constructing the weight assignment method according to the installed capacity information and the meteorological information, the NWP wind speed features of the wind power cluster in the available feature set are used as the initial feature group; The sub-module for constructing the weight assignment method includes: A weather factor weight acquisition sub-module, which is used to take the NWP forecast wind speed of the wind farm corresponding to the initial feature group as the benchmark and the CCM causal score of each feature as the weight index when considering weather factors; extract the values of the CCM causal score higher than the set threshold and normalize them into weather factor weights; A capacity factor weight acquisition sub-module, which is used to take the installed capacity of the wind farm corresponding to the initial feature group as the benchmark and the installed capacity ratio of each wind farm as the weight index when considering the capacity factor weight; normalize the installed capacity ratio into capacity factor weights; A wind farm influence weight construction sub-module, which is used to combine the weather factor weights and the capacity factor weights to construct the influence weight of each wind farm on the entire wind power cluster.

Images