Renewable energy power multi-step prediction method based on Transform-SWBiGRU

Abstract

The invention discloses a renewable energy power multi-step prediction method based on Transform-SWBiGRU, and belongs to the technical field of new energy power prediction. The method comprises the steps of obtaining a historical generated power sequence and a historical meteorological feature sequence of a photovoltaic power station to be subjected to power prediction, and performing preprocessing; performing nonlinear analysis on multi-dimensional meteorological data in the meteorological characteristics by combining a Chatterj correlation coefficient and a feature screening algorithm of a sliding time window strategy, and extracting key meteorological data; then, a hybrid prediction model fusing Transform and SWBiGRU is constructed, then key meteorological data and historical photovoltaic power serve as input, and the model is trained and optimized; and finally, predicting the photovoltaic power of a plurality of time steps in the future by using the trained model. According to the invention, the scheduling reliability of photovoltaic power generation can be obviously improved.

Description

Technical Field

[0001] This invention belongs to the field of new energy power prediction technology, specifically relating to a multi-step prediction method for renewable energy power based on Transformer-SWBiGRU. Background Technology

[0002] As global climate issues become increasingly severe, the energy structure is accelerating its transformation towards cleaner and lower-carbon energy, driving the rapid growth of photovoltaic (PV) installed capacity. However, PV power output is affected by multiple factors such as irradiance, temperature, and humidity, exhibiting intermittency and fluctuations, posing a significant challenge to the stable operation of the power grid. Therefore, high-precision PV power forecasting has become crucial for improving the grid's absorption capacity.

[0003] Currently, the mainstream methods for multi-step photovoltaic power prediction include recursive multi-step methods, direct multi-step methods, and multi-output prediction methods. Recursive multi-step methods (such as models based on RNNs or GRUs) can model time series through recursive structures, but errors accumulate significantly in long-term predictions. Direct multi-step methods (such as multimodal methods combining satellite imagery and historical data) can improve robustness, but require separate modeling for each prediction step, resulting in parameter redundancy and neglecting temporal correlations. Multi-output prediction methods (such as multiquantile LSTMs or MIMO neural networks) can simultaneously output predictions for multiple time points, enhancing consistency, but the models are complex and require high data and computing power. Therefore, model structure design has a decisive impact on prediction performance.

[0004] Of course, feature extraction is also a crucial step in improving prediction accuracy. Traditional linear correlation analysis methods (such as Pearson and Spearman coefficients) can only identify global monotonic relationships and are difficult to handle nonlinear dependencies. Mutual information (MI) has some ability to capture nonlinearities, but its robustness is poor. Chatterjee correlation coefficient (CCC) can assess nonlinear relationships between variables, but it is not sensitive enough to local dynamic changes.

[0005] In summary, existing methods still have limitations in terms of long-term prediction consistency, nonlinear feature extraction, and dynamic response capabilities. There is an urgent need to construct a hybrid prediction model that can integrate global modeling and local feature perception, and possesses high robustness and adaptability. Summary of the Invention

[0006] To address the problems in the prior art, this invention provides a multi-step prediction method for renewable energy power based on Transformer-SWBiGRU.

[0007] The technical solution of the present invention is as follows:

[0008] This invention discloses a multi-step renewable energy power prediction method based on Transformer-SWBiGRU for predicting the power generation of photovoltaic power plants; the method includes:

[0009] 1) Obtain the historical power generation sequence and historical meteorological characteristic sequence of the photovoltaic power station to be predicted, and perform preprocessing to obtain the historical power generation-meteorological characteristic sequence; among which, the meteorological characteristics include multidimensional meteorological data;

[0010] 2) Based on the sliding time window strategy, multiple window sequences are dynamically extracted from the historical power generation-meteorological feature sequence. The Chatterjee correlation coefficient is used to calculate the nonlinear dependence strength between each meteorological data in each window sequence and photovoltaic power. Several dimensions of meteorological data with nonlinear dependence strength greater than the preset correlation threshold are selected as relevant meteorological features.

[0011] 3) Construct a hybrid prediction model that integrates Transformer and SWBiGRU, update the historical power generation-meteorological feature sequence based on the relevant meteorological features, and train the hybrid prediction model using the updated historical power generation-meteorological feature sequence; wherein, the hybrid prediction model includes an encoding stage and a decoding stage, the encoding stage is used to capture the global dependency relationship between meteorological data and power generation as well as the temporal features of power generation, and the decoding stage adopts a parallel multi-step decoding mechanism to generate power generation prediction values ​​for multiple future time points;

[0012] 4) Obtain the relevant meteorological characteristics at the location of the photovoltaic power station and input them into the trained hybrid prediction model to obtain the predicted power generation of the photovoltaic power station at multiple future time points.

[0013] Further, step 2) includes:

[0014] 21) Set the minimum length, maximum length, and sliding step size of the sliding time window, and set the correlation threshold;

[0015] 22) Initialize the optimal CCC value between each meteorological data and power generation to 0;

[0016] 23) Starting from the time series starting point of the historical power generation-meteorological characteristic sequence, slide the time window according to the set step size. For the position of the sliding time window, use the Chatterjee correlation coefficient to calculate the nonlinear dependence strength between each meteorological data and power generation, and update the current optimal CCC value corresponding to each meteorological data.

[0017] 24) Perform a sliding traversal of the entire time series of historical power generation-meteorological characteristic sequence, and repeat step 23) to obtain the optimal CCC value corresponding to each meteorological data.

[0018] 25) Compare the optimal CCC value corresponding to each meteorological data with the correlation threshold, retain the meteorological data corresponding to the multiple optimal CCC values ​​that are greater than the correlation threshold, and use them as relevant meteorological features.

[0019] Further, in step 3), when training the hybrid prediction model using the updated historical power generation-meteorological feature sequence, the encoding stage of the hybrid prediction model includes:

[0020] First, the updated historical power generation-meteorological feature sequence is divided into multiple first sequence segments. Then, each first sequence segment is transformed by a learnable linear projection to obtain the corresponding projected feature vector. Next, the ReLU activation function is used to process each projected feature vector to obtain the first feature vector corresponding to each first sequence segment. The first feature vectors corresponding to each first sequence segment are input into the encoding layer of the Transformer to obtain the feature matrix corresponding to each first sequence segment. Then, multiple parallel BiGRU units recursively iterate the feature matrix corresponding to each first sequence segment according to the time order of each first sequence segment to obtain a hidden feature, which is then output to the decoding stage of the hybrid prediction model.

[0021] Further, in step 3), when training the hybrid prediction model using the updated historical power generation-meteorological feature sequence, the decoding stage of the hybrid prediction model includes:

[0022] First, the updated historical power generation-meteorological feature sequence is divided into multiple second sequence segments according to the preset future photovoltaic power prediction time series interval. Then, a corresponding relative position code and channel position code are generated for each second sequence segment. Based on the relative position code and channel position code of the second sequence segment, a position embedding of the second sequence segment is generated. The position embedding of each second sequence segment is fused with the hidden features output from the encoding stage to obtain feature pairs of each second sequence segment. Next, multiple BiGRU units process the feature pairs of each second sequence segment in parallel to obtain the second feature vector of each second sequence segment. Then, a Dropout layer is used to regularize all the second feature vectors to obtain multiple regularized second feature vectors. Then, a learnable linear projection is used to transform each regularized second feature vector to obtain multiple feature matrices. All feature matrices are concatenated in time sequence to obtain the predicted power generation values ​​at multiple future time points.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] 1) The feature selection algorithm that combines Chatterjee correlation coefficient (CCC) and sliding time window (TWSE) overcomes the technical problems of traditional linear correlation analysis (such as Pearson and Spearman coefficients) in failing to capture nonlinear dependencies, poor robustness of mutual information (MI), and difficulty of single CCC in perceiving time-series dynamic changes. It can dynamically mine meteorological data that are strongly correlated with photovoltaic power, effectively remove redundant information, reduce feature selection differences, and improve model training efficiency and input data quality.

[0025] 2) Construct a hybrid prediction model that integrates Transformer and SWBiGRU. In the encoding stage, Transformer captures the global dependency relationship between meteorological data and power generation, and BiGRU extracts time series features through segmented recursive iteration. This overcomes the limitation of a single model in taking into account both global correlation and local time series patterns, and achieves efficient modeling of complex time series data.

[0026] 3) This invention can effectively screen out the most representative meteorological factors and reduce the differences in the feature selection process; combined with the Transformer-SWBiGRU combined prediction model, it can effectively extract power generation characteristics and efficiently model complex time series data. Attached Figure Description

[0027] Figure 1 This is a flowchart of the multi-step renewable energy power prediction method based on Transformer-SWBiGRU of the present invention;

[0028] Figure 2 This is a framework diagram of the hybrid prediction model that integrates Transformer and SWBiGRU of the present invention;

[0029] Figure 3 This is a comparison chart of the actual and predicted power generation values ​​of test set 1 in one embodiment of the present invention;

[0030] Figure 4 This is a comparison chart of the actual and predicted power generation values ​​of test set 2 in one embodiment of the present invention;

[0031] Figure 5 This is a comparison chart of the actual and predicted power generation values ​​of test set 3 in one embodiment of the present invention;

[0032] Figure 6 This is a comparison chart of the actual and predicted power generation values ​​of test set 4 in one embodiment of the present invention. Detailed Implementation

[0033] The present invention will be further described and illustrated below with reference to specific embodiments. The embodiments described are merely examples of the content of this disclosure and do not limit the scope of the invention. The technical features of each embodiment in the present invention can be combined accordingly, provided that there is no mutual conflict.

[0034] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-step prediction method for renewable energy power based on Transformer-SWBiGRU, so as to solve the problems of input feature redundancy, frequent iteration and error accumulation in the existing multi-step prediction methods, and improve prediction accuracy and reliability.

[0035] To address the problem of redundant meteorological features, this invention proposes a feature selection algorithm that combines the CCC (Chatterjee Correlation Coefficient) and TWSE (Time Window Shrinking and Expanding) strategies. This algorithm improves model training efficiency by performing nonlinear analysis on the relationship between multivariate meteorological characteristic sequences and photovoltaic power output to uncover strongly correlated features.

[0036] To address the issues of large cumulative error and frequent iterations in recursive prediction in multi-step prediction tasks, this invention proposes a multi-step prediction model based on Transformer-Stepwise Iteration Bidirectional Recurrent Neural Network (SWBiGRU) and combines it with a parallel prediction strategy to significantly improve prediction accuracy.

[0037] The present invention comprises a multi-step prediction method for renewable energy power based on Transformer-SWBiGRU, which combines feature selection algorithms of CCC and TWSE strategies with a multi-step prediction model based on Transformer-segmented iterative bidirectional gated cyclic units.

[0038] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0039] like Figure 1 As shown, the multi-step renewable energy power prediction method based on Transformer-SWBiGRU of the present invention includes the following steps:

[0040] Step 1: Obtain the historical power generation sequence and historical meteorological feature sequence of the photovoltaic power station to be predicted, and perform preprocessing to obtain the historical power generation-meteorological feature sequence; among which, the meteorological features include multi-dimensional meteorological data.

[0041] Among them, the historical power generation sequence includes the power generation of the photovoltaic power station as a function of time during the historical period, and the historical meteorological characteristic sequence includes the total radiation, direct radiation, diffuse radiation, ambient temperature, air pressure and relative humidity of the location of the photovoltaic power station during the historical period. The meteorological data are total radiation, direct radiation, diffuse radiation, ambient temperature, air pressure or relative humidity.

[0042] Preprocessing historical power generation sequences involves filling in missing values ​​and handling outliers. Preprocessing historical meteorological feature sequences involves filling in missing values ​​and handling outliers for total radiation, direct radiation, diffuse radiation, ambient temperature, air pressure, and relative humidity.

[0043] The historical power generation-meteorological characteristic sequence includes multiple historical power generation-meteorological characteristic data pairs arranged in chronological order. The historical power generation-meteorological characteristic data pairs are the power generation, total radiation, direct radiation, diffuse radiation, ambient temperature, air pressure, and relative humidity at the same time.

[0044] Step 2: Based on a sliding time window strategy, multiple historical power generation-meteorological feature window sequences are dynamically selected from the historical power generation-meteorological feature sequences. The Chatterjee correlation coefficient is used to calculate the nonlinear dependence strength between each meteorological data point and photovoltaic power in each historical power generation-meteorological feature window sequence. Meteorological data corresponding to multiple nonlinear dependence strengths greater than a preset correlation threshold are selected and used as relevant meteorological features. The goal of this step is to identify relevant time windows within the time series intervals. The CCC (Chatterjee correlation coefficient) index is calculated within the identified window. While CCC has significant advantages in measuring nonlinear dependencies between variables, its static nature makes it ineffective in capturing temporal correlations in dynamically evolving systems. Meanwhile, the Time Window Shrinking and Expanding (TWSE) strategy, based on a sliding window, effectively locates key intervals in time-series data, but has limitations in quantitatively analyzing dependency strength within the window. To combine the advantages of both, this invention embeds CCC into the TWSE framework: the photovoltaic power-meteorological characteristic sequence is segmented using a sliding window mechanism, and the CCC value is dynamically calculated within each window to capture the evolution of dependencies over time. This fusion strategy not only enhances the expressive power of time-series dependency modeling but also significantly improves the accuracy and adaptability of the prediction model.

[0045] Step two includes the following steps:

[0046] First, set the minimum length of the sliding time window ( ), maximum length ( The system uses a sliding step size and a correlation threshold to determine whether a feature is "strongly correlated".

[0047] Then, the nonlinear dependence strength between meteorological data and photovoltaic power in a single historical power generation-meteorological feature window sequence is calculated using the Chatterjee correlation coefficient. Specifically: 1) Initialization: Initialize the optimal CCC value between each meteorological data point and power generation to 0, and initialize the window information corresponding to each optimal CCC value to be empty. 2) Window sliding: Starting from the beginning of the historical power generation-meteorological feature sequence, slide the time window according to the set step size. For each window position: extract the historical power generation-meteorological feature window sequence (i.e., subsequence), and then calculate the CCC: use the Chatterjee correlation coefficient formula to calculate the nonlinear dependence strength between each meteorological data point and power generation. Update the optimal CCC value: compare the nonlinear dependence strength between each meteorological data point and power generation with the current optimal CCC value corresponding to the meteorological data. If the nonlinear dependence strength between each meteorological data point and power generation is greater than the current optimal CCC value corresponding to the meteorological data, then the current nonlinear dependence strength between the meteorological data point and power generation is taken as the updated optimal CCC value corresponding to that meteorological data, and the time range of the current sliding time window corresponding to the meteorological data is taken as the updated window information corresponding to that meteorological data. 3) Perform a sliding traversal of the entire historical power generation-meteorological characteristic sequence over time, and repeat step 2) to obtain the optimal CCC value and window information corresponding to each meteorological data. 4) Compare the optimal CCC value corresponding to each meteorological data with a correlation threshold, and retain the meteorological data corresponding to multiple optimal CCC values ​​that are greater than the correlation threshold. The window information is the photovoltaic power-meteorological data sequence.

[0048] Furthermore, to further reduce computational complexity, this invention sets a larger window range. Candidate sets in time series intervals The process involved screening and excluding those that did not meet the requirements.

[0049] Step 3: Construct a hybrid prediction model integrating Transformer and SWBiGRU. Update the historical power generation-meteorological feature sequence based on the relevant meteorological features obtained in Step 2), and train the hybrid prediction model using the updated historical power generation-meteorological feature sequence to obtain the trained hybrid prediction model. The hybrid prediction model includes an encoding stage and a decoding stage. The encoding stage is used to capture the global dependency between meteorological data and power generation, as well as the temporal characteristics of power generation. The decoding stage employs a parallel multi-step decoding mechanism to generate power generation predictions for multiple future time points.

[0050] The architecture of a hybrid prediction model integrating Transformer and SWBiGRU is as follows: Figure 2 As shown in the diagram, in the hybrid prediction model, the Transformer is responsible for extracting the global dependencies between segments, while BiGRU focuses on modeling temporal recursion patterns and uses a sliding window mechanism to process features in segments, thereby improving training efficiency. Finally, a parallel multi-step decoding mechanism is designed to accelerate inference speed while ensuring prediction stability. The two are connected sequentially to construct a deep learning framework that combines global perception and local dynamic response capabilities. During the modeling process, the input sequence is first divided into multiple segments, and the Transformer layer captures the long-distance dependencies between different segments to extract global feature information. Subsequently, the BiGRU layer further performs temporal modeling on these features to deeply capture the dynamic change patterns of the time series.

[0051] Finally, the output layer employs a Dropout regularization strategy to mitigate the risk of overfitting, while a linear mapping layer with learnable parameters transforms high-dimensional features into the final prediction result, enabling multi-step prediction of photovoltaic power (i.e., power generation).

[0052] During the encoding phase, the hybrid prediction model replaces the original point-to-point iterative approach with a segmented processing approach, effectively reducing the number of iterations from L to L / w, as detailed below.

[0053] a) Sequence segmentation and projection

[0054] For the updated historical power generation-meteorological characteristic sequence It can be divided into multiple first sequence segments. ,in, represents the window length of the updated historical power generation-meteorological characteristic sequence, and w represents the window length of each first sequence segment. This indicates the number of the first sequence segments. Then, these first sequence segments... It will be achieved through a learnable linear projection. Convert to Then, the ReLU activation function is applied, where d represents the hidden state dimension of the BiGRU unit. Next, the transformed vector... The input is fed into the Transformer architecture. This architecture allows the model to focus on the relationships between different sequence segments in the input sequence, constructing a global representation by establishing the dependency between each query and all keys, effectively capturing long-distance dependencies between input features; essentially, the representation of each feature is influenced by all other features in the sequence.

[0055] b) Recursive encoding

[0056] Next, BiGRU is used for recursive iteration to learn and extract. The hidden features obtained in the final step are used to effectively capture the temporal relationships in the sequence data, thereby enriching the model's understanding of time series changes. After n recursive iterations, the hidden features are... The original sequence is already included. All temporal features are hidden. This hidden feature will be passed to the decoding stage for subsequent inference and prediction steps.

[0057] During the decoding phase, the hybrid prediction model employs a parallel multi-step prediction strategy, further reducing the number of iterations from H / w to 1. The specific process is as follows:

[0058] A) Location embedding

[0059] During the decoding phase, the updated historical power generation-meteorological feature sequence needs to be divided into multiple second sequence segments according to the preset future photovoltaic power prediction time series interval. Due to recursion interruption, the order relationship between the second sequence segments is lost. Furthermore, the Transformer itself cannot capture the order of the input second sequence segments. This invention utilizes position embedding to maintain the local dependencies in the sequence. Specifically, this invention generates a corresponding position embedding for each second sequence segment, represented as follows: This is to identify the location of each second sequence segment. Here, This represents the number of windows obtained by segmenting a pre-defined time series interval of future photovoltaic power to be predicted. Each location is embedded... It is achieved by encoding relative positions. and channel position encoding The sequence is composed of multiple segments. The location code represents the predicted location of each second sequence segment within the updated historical power generation-meteorological characteristic sequence. Simultaneously, the channel location code represents the channel's position among all channels within the current second sequence segment.

[0060] B) Parallel Decoding

[0061] During the decoding phase, the same BiGRU unit used in the encoding phase is shared. Specifically, the final state obtained during the encoding phase... It will be copied m times and embedded in m positions. These pairs are then combined. These pairs are subsequently processed in parallel by BiGRU units. This parallel processing generates m output vectors, each of length d, representing... It is worth noting that this method differs from traditional recursive processing because the computation of each vector is independent of the results of previous time steps. Therefore, parallel computation within the sample is achieved, thereby improving inference speed. Furthermore, prediction error does not accumulate with increasing iterations, thus improving prediction accuracy.

[0062] C) Prediction and Sequence Recovery

[0063] A Dropout layer is used to randomly discard a portion of values ​​for regularization. The regularized result is then transformed through a learnable linear projection layer. Finally, all of them Reshaped into , This represents the final forecast result, i.e., the predicted power generation at multiple points in the future.

[0064] Step 4: Obtain the relevant meteorological characteristics of the location of the photovoltaic power station to be predicted, and input them into the trained hybrid prediction model to obtain the predicted power generation value of the photovoltaic power station at multiple future time points.

[0065] In a specific embodiment of the present invention, photovoltaic datasets Test1, Test2, Test3, and Test4 are constructed using data from the last week of the 2nd, 5th, 8th, and 11th months of 2017 from actual operational measurement data from a certain location. These datasets are then used as test datasets to verify the effectiveness of the present invention.

[0066] The specific data structure of the photovoltaic dataset includes total radiation (horizontal plane), direct radiation (horizontal plane), diffuse radiation (horizontal plane), module temperature, ambient temperature, air pressure, relative humidity, and power generation. This invention uses the photovoltaic power in the test dataset as an input variable to predict the photovoltaic power for the next week.

[0067] To rigorously evaluate the predictive performance of this invention, root mean square error was used. Mean absolute error Accuracy and pass rate As an evaluation indicator.

[0068] like Figure 3 , Figure 4 , Figure 5 , Figure 6 As shown in Table 1, they illustrate the prediction results of the present invention. Figures 3-6In the diagram, blue information represents predicted values, and red information represents actual values. Overall, this invention demonstrates good generalization ability and stability in highly challenging multi-step prediction scenarios. Specifically, in the photovoltaic dataset Test1 (i.e., Figure 3 In the test set 1), the present invention performs particularly well, with a root mean square error of [missing value]. The mean absolute error is 0.0924. The accuracy is 0.0471. For 90.7635% and The pass rate was 95.6909%. This was achieved in the photovoltaic dataset Test2 (i.e....). Figure 4 In test set 2), although the performance is slightly lower than that of the photovoltaic dataset Test1, it still maintains a high level, with a four-mean-square error. The mean absolute error is 0.1167. The accuracy rate is 0.0637. For 88.3323% and The pass rate was 93.6107%. This was achieved in the photovoltaic dataset Test3 (i.e....). Figure 5 Test set 3) and photovoltaic dataset Test4 (i.e. Figure 6 In the test set 4), the prediction performance of the present invention decreased slightly, but the overall performance remained within an acceptable range. Analysis shows that the present invention has high accuracy in multi-step prediction tasks and strong robustness in different scenarios. Furthermore, the proposed model overcomes the limitations of traditional BiGRU models, which are prone to getting stuck in deep iterations and producing random predictions.

[0069] Table 1 Prediction Indicators for Different Photovoltaic Data Sets

[0070]

[0071] To address the problems of redundant input features, frequent iterations, and error accumulation in existing multi-step prediction methods, this invention proposes a multi-step renewable energy power prediction method based on Transformer-SWBiGRU. Specifically, the proposed CCC-TWSE algorithm, compared to traditional Pearson and Spearman correlation analysis methods, further filters environmental temperature features, possesses stronger nonlinear structure awareness capabilities, and effectively improves prediction accuracy. Furthermore, the hybrid prediction model integrating Transformer and SWBiGRU in this invention comprehensively outperforms mainstream methods such as Transformer, FEDformer, Informer, and BiGRU in modeling long-term dependencies, improving prediction accuracy, and controlling error propagation, effectively addressing the problems of large accumulated errors and frequent iterations inherent in recursive prediction.

[0072] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. Those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A multi-step renewable energy power prediction method based on Transformer-SWBiGRU, used to predict the power generation of photovoltaic power plants; characterized in that, The method includes: 1) Obtain the historical power generation sequence and historical meteorological characteristic sequence of the photovoltaic power station to be predicted, and perform preprocessing to obtain the historical power generation-meteorological characteristic sequence; among which, the meteorological characteristics include multidimensional meteorological data; 2) Based on the sliding time window strategy, multiple window sequences are dynamically extracted from the historical power generation-meteorological feature sequence. The Chatterjee correlation coefficient is used to calculate the nonlinear dependence strength between each meteorological data in each window sequence and photovoltaic power. Several dimensions of meteorological data with nonlinear dependence strength greater than the preset correlation threshold are selected as relevant meteorological features. 3) Construct a hybrid prediction model that integrates Transformer and SWBiGRU, update the historical power generation-meteorological feature sequence based on the relevant meteorological features, and train the hybrid prediction model using the updated historical power generation-meteorological feature sequence; wherein, the hybrid prediction model includes an encoding stage and a decoding stage, the encoding stage is used to capture the global dependency relationship between meteorological data and power generation as well as the temporal features of power generation, and the decoding stage adopts a parallel multi-step decoding mechanism to generate power generation prediction values ​​for multiple future time points; 4) Obtain the relevant meteorological characteristics at the location of the photovoltaic power station and input them into the trained hybrid prediction model to obtain the predicted power generation of the photovoltaic power station at multiple future time points.

2. The multi-step renewable energy power prediction method based on Transformer-SWBiGRU according to claim 1, characterized in that, In step 1), the historical power generation sequence includes the power generation of the photovoltaic power station arranged by time within a historical period; the historical meteorological characteristic sequence includes the total radiation, direct radiation, diffuse radiation, ambient temperature, air pressure and relative humidity of the location of the photovoltaic power station arranged by time within a historical period. The meteorological data are total radiation, direct radiation, diffuse radiation, ambient temperature, air pressure or relative humidity.

3. The multi-step renewable energy power prediction method based on Transformer-SWBiGRU according to claim 1, characterized in that, In step 1), the historical power generation-meteorological characteristic sequence includes multiple historical power generation-meteorological characteristic data pairs arranged in chronological order. The historical power generation-meteorological characteristic data pairs are the power generation, total radiation, direct radiation, diffuse radiation, ambient temperature, air pressure and relative humidity at the same time.

4. The multi-step renewable energy power prediction method based on Transformer-SWBiGRU according to claim 1, characterized in that, Step 2) includes: 21) Set the minimum length, maximum length, and sliding step size of the sliding time window, and set the correlation threshold; 22) Initialize the optimal CCC value between each meteorological data and power generation to 0; 23) Starting from the time series starting point of the historical power generation-meteorological characteristic sequence, slide the time window according to the set step size. For the position of the sliding time window, use the Chatterjee correlation coefficient to calculate the nonlinear dependence strength between each meteorological data and power generation, and update the current optimal CCC value corresponding to each meteorological data. 24) Perform a sliding traversal of the entire time series of historical power generation-meteorological characteristic sequence, and repeat step 23) to obtain the optimal CCC value corresponding to each meteorological data. 25) Compare the optimal CCC value corresponding to each meteorological data with the correlation threshold, retain the meteorological data corresponding to the multiple optimal CCC values ​​that are greater than the correlation threshold, and use them as relevant meteorological features.

5. The multi-step renewable energy power prediction method based on Transformer-SWBiGRU according to claim 4, characterized in that, In step 23), when the calculated nonlinear dependence strength between each meteorological data and the power generation is greater than the current optimal CCC value corresponding to the meteorological data, the optimal CCC value corresponding to each meteorological data is updated.

6. The multi-step renewable energy power prediction method based on Transformer-SWBiGRU according to claim 1, characterized in that, In step 3), when training the hybrid prediction model using the updated historical power generation-meteorological feature sequence, the encoding stage of the hybrid prediction model includes: First, the updated historical power generation-meteorological feature sequence is divided into multiple first sequence segments. Then, each first sequence segment is transformed by a learnable linear projection to obtain the corresponding projected feature vector. Next, the ReLU activation function is used to process each projected feature vector to obtain the first feature vector corresponding to each first sequence segment. The first feature vectors corresponding to each first sequence segment are input into the encoding layer of the Transformer to obtain the feature matrix corresponding to each first sequence segment. Then, multiple parallel BiGRU units recursively iterate the feature matrix corresponding to each first sequence segment according to the time order of each first sequence segment to obtain a hidden feature, which is then output to the decoding stage of the hybrid prediction model.

7. The multi-step renewable energy power prediction method based on Transformer-SWBiGRU according to claim 6, characterized in that, In step 3), when training the hybrid prediction model using the updated historical power generation-meteorological feature sequence, the decoding stage of the hybrid prediction model includes: First, the updated historical power generation-meteorological feature sequence is divided into multiple second sequence segments according to the preset future photovoltaic power prediction time series interval. Then, a corresponding relative position code and channel position code are generated for each second sequence segment. Based on the relative position code and channel position code of the second sequence segment, a position embedding of the second sequence segment is generated. The position embedding of each second sequence segment is fused with the hidden features output from the encoding stage to obtain feature pairs of each second sequence segment. Next, multiple BiGRU units process the feature pairs of each second sequence segment in parallel to obtain the second feature vector of each second sequence segment. Then, a Dropout layer is used to regularize all the second feature vectors to obtain multiple regularized second feature vectors. Then, a learnable linear projection is used to transform each regularized second feature vector to obtain multiple feature matrices. All feature matrices are concatenated in time sequence to obtain the predicted power generation values ​​at multiple future time points.

8. The multi-step renewable energy power prediction method based on Transformer-SWBiGRU according to claim 7, characterized in that, The BiGRU units in the encoding and decoding stages of the hybrid prediction model share parameters.

Images