New energy power generation system power prediction method and system combining CEEMDAN decomposition and Informer-BiGRU model

Abstract

The invention relates to a new energy power generation system power prediction method and system in combination with CEEMDAN decomposition and an Informer-BiGRU model. The prediction method comprises the following steps: firstly, collecting and preprocessing historical new energy power generation power data and related meteorological data, performing data cleaning and normalization processing, and analyzing feature correlation; then, the original power data is decomposed by using CEEMDAN, and sub-sequence data with different complexities are obtained; and then, constructing a combined prediction model combining Informer and BiGRU, and obtaining a prediction result of the subsequences through feature splicing and linear layer mapping. And finally, directly adding the prediction results of all the components to obtain a final prediction value. According to the method, the prediction precision can be effectively improved when complex new energy power generation power data is processed, and the method has relatively high robustness and generalization ability.

Description

Technical Field

[0001] The present invention relates to the field of power prediction of new energy power generation systems, and particularly to a power prediction method for new energy power generation systems combining CEEMDAN decomposition and Informer-BiGRU model. Background Art

[0002] In today's society, power prediction of new energy power generation systems plays a crucial role in aspects such as power dispatching, energy management, and grid stability. Accurate power prediction can not only optimize the allocation of power resources but also effectively reduce the phenomena of light curtailment and wind curtailment, improving the economic and environmental benefits of new energy power generation systems. With the continuous expansion of the scale of new energy power generation, the complexity of power prediction is also increasing. How to extract valuable information from multi-source data such as meteorological data, historical power data, and equipment operating status and conduct accurate prediction has become an urgent problem to be solved.

[0003] However, due to the characteristics of new energy power generation power data such as non-stationarity, non-linearity, complex long- and short-term dependence relationships, and significant multi-scale characteristics, traditional prediction methods (such as long short-term memory network LSTM, temporal convolutional network TCN, etc.) have certain limitations in dealing with complex new energy power generation power data. Traditional time series prediction methods (such as ARIMA, exponential smoothing method) usually assume that the data is stationary or weakly stationary, and it is difficult to effectively process the strongly non-stationary and multi-scale changing new energy power generation power data. These methods perform poorly in dealing with long-term dependence relationships and are difficult to capture complex time series characteristics. Classic RNN and LSTM structures can effectively model the long-term dependence relationships of time series data, but when facing ultra-long time series data, there are problems of gradient disappearance and high computational complexity. Although CNN can process time series data in parallel, it is still limited in capturing global dependencies. In recent years, the Transformer structure has performed well in time series modeling, but its original self-attention mechanism has a high computational complexity and is difficult to be directly applied to the prediction of new energy power generation power data.

[0004] Since new energy power generation power data often contains multi-scale fluctuation characteristics, directly inputting the original sequence into a deep learning model may lead to difficult learning for the model. Signal decomposition methods such as empirical mode decomposition (EMD), ensemble empirical mode decomposition (EEMD), and complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) can decompose complex time series into multiple intrinsic mode functions (IMFs), improving prediction stability and accuracy. Existing research shows that combining time series decomposition technology with deep learning models (such as LSTM, Transformer) can effectively improve prediction accuracy, but there is still a lack of a method for jointly modeling ultra-long time series and local short-term dependence characteristics. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to construct an efficient and accurate new energy power generation prediction method to address the limitations of traditional methods in dealing with complex new energy power generation data and provide more reliable prediction support for power dispatching and energy management.

[0006] To solve the above technical problems, the present invention proposes a new energy power generation system power prediction method combining CEEMDAN decomposition and Informer-BiGRU model, aiming to make full use of CEEMDAN decomposition to reduce the complexity of subsequences, the advantages of Informer in long-term dependence modeling, and the ability of BiGRU in short-term local feature modeling.

[0007] The new energy power generation system power prediction method proposed by the present invention includes the following steps: The present invention proposes a new energy power generation system power prediction method combining CEEMDAN decomposition and Informer-BiGRU model, including the following steps: Step S1: Collect and preprocess historical new energy power generation data and related meteorological data, perform data cleaning and normalization processing to obtain the original data set.

[0008] Step S2: Conduct a correlation analysis on the predicted power data and its characteristic data to screen out highly correlated characteristic data; it includes the following steps: S201: Calculate the Spearman correlation coefficient between new energy power generation and multiple influencing factors; S202: Visualize the correlation between variables through a correlation heat map, and screen out highly correlated characteristics as model inputs according to the Spearman correlation coefficient, while removing redundant characteristics.

[0009] Step S3: Perform CEEMDAN decomposition on the data set obtained in Step S1 to form subsequence data with multiple different frequency characteristics and time scale characteristics; specifically, it includes the following steps: (1) Initialization: Select appropriate white noise amplitude and number of iterations; (2) Add white noise: Add Gaussian white noise to the original power signal to construct multiple preprocessing sequences; (3) EMD decomposition: Perform EMD decomposition on each preprocessing sequence to obtain the first-order IMF component , and calculate its mean as the first-order IMF, and calculate the residual at the same time; (4) Iterative processing: Add the residual to new white noise, repeat the EMD decomposition to obtain the next-order IMF, and calculate the new residual; (5) Repeated iteration: Repeat the above process until the stopping condition is met, and finally obtain all IMF components; Add all IMF components and the residual to obtain the decomposition expression of the original power signal.

[0010] Step S4: Construct a combined prediction model integrating Informer and BiGRU, input the decomposed subsequence data and the filtered feature data into the combined prediction model, and obtain the prediction results of the subsequences through feature splicing and linear layer mapping; The specific steps include: (1) Combine the subsequences obtained by decomposition with other features screened by Spearman to form multiple new data sets and input them into the model; (2) Adopt the Informer structure for global time series modeling; (3) Adopt the BiGRU structure for short-time series dependence relationship modeling to obtain local power sequence features; (4) By fusing the long-term dependence features of Informer and the short-term dependence features of the output of BiGRU, splice and weight-combine the outputs of the two, and train different weight ratios to achieve the final prediction of the new energy power generation subsequence.

[0011] S5: Obtain the final predicted value by directly adding the prediction results of all components.

[0012] Further, for the method proposed by the present invention, the calculation formula of the Spearman correlation coefficient in step S2 is as follows: , where, represents the Spearman correlation coefficient between sequence S and sequence K, represents the rank difference of each pair in the two sequences, n represents the size of the data sample, The larger the value of, the stronger the correlation between sequence S and sequence K, and vice versa, the weaker the correlation.

[0013] Further, for the method proposed by the present invention, in step S3, perform CEEMDAN decomposition on the data set obtained in S1, and the specific calculation formula is as follows: By adding N Gaussian white noise sequences with a mean of 0 to the original power signal construct the sequence to be decomposed for a total of N experiments : , where, is the Gaussian white noise weight coefficient; is the white noise sequence added for the -th time, ; Pair Apply the EMD algorithm for decomposition to obtain the first mode component IMF and the first unique residual component : , , Add noise to the j stage residual component obtained after decomposition and continue to apply EMD for decomposition: , , Among them, j = 2, 3,..., N； Repeat the decomposition until the termination condition is met. The termination criterion is that the number of extreme points of the residual signal does not exceed 2 at most; Finally, the original power signal sequence is decomposed into N mode components and a residual term : .

[0014] Furthermore, in the method proposed by the present invention, step S4 constructs a combined prediction model that combines Informer and BiGRU. The formula of the Informer-BiGRU combined prediction model is as follows: (1) Adopt the Informer structure for global time series modeling: For the input sequence , first obtain the query matrix , key matrix and value matrix through a linear transformation: , , , Sparse self-attention output result and pass it to the self-attention distillation module: , , Among them, is the key vector dimension, A is an n×n matrix, that is, the attention score matrix; Compress , , matrix through a low-dimensional approximate representation: , , , Among them, , , is , , the compressed matrix; The core formula of generative decoding: , The encoder processes historical data to generate a fixed-dimensional representation, and the decoder directly generates all future prediction results. Among them, C is the context representation generated by the encoder, and Y is all the prediction results output by the decoder; (2) Use the BiGRU structure to model short-term temporal dependencies: , , , , ; Among them, is the activation value of the update gate, which determines the influence degree of the previous hidden state on the current moment ; is the weight matrix of the input data for the update gate; is the weight matrix of the previous hidden state for the update gate; is the bias term of the update gate; is the activation value of the reset gate; is the weight matrix of the input data for the reset gate; is the weight matrix of the previous hidden state for the reset gate; is the bias term of the reset gate; is the candidate hidden state at the current moment; is the weight matrix of the input data for the candidate hidden state; is the weight matrix of the previous hidden state for the candidate hidden state; is the bias term of the candidate hidden state; is the final state at the current moment; is the hidden state obtained by forward GRU calculation; is the hidden state obtained by backward GRU calculation; is the output of the bidirectional GRU at time t; (3) Feature splicing and weighted fusion are performed on the output matrices of Informer and BiGRU: , , The final power prediction result is formed through a fully connected layer: , Among them, , , are learning parameters and are obtained by self-learning through model training.

[0015] The training process of the Informer-BiGRU combined prediction model includes: (1) Use the PyTorch framework for construction and training; (2) Divide the training set, validation set, and test set in a ratio of 7:2:1 to enhance the generalization ability of the model; (3) Use the trial-and-error method to adjust the hyperparameters to improve the prediction accuracy of the model; (4) Use the mean absolute error MAE, root mean square error RMSE, and coefficient of determination R 2 Calculate the percentage of the power prediction error in the actual value to measure the relative error.

[0016] The calculation formula for the prediction model evaluation standard is as follows: , , , Among them, is the actual power test sample value; is the model power prediction result; is the average power; is the total number of power test samples.

[0017] In addition, the present invention also proposes an electronic system, including: at least one processor; and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method steps of the present invention.

[0018] By adopting the above method, the present invention has the following technical effects compared with the prior art: The present invention uses the CEEMDAN method to decompose new energy power generation data, decomposing complex time series signals into multiple intrinsic mode functions (IMFs) and a residual term to reduce the non-stationarity and complexity of the data. It uses Informer for long time series modeling, combined with a sparse self-attention mechanism, to effectively improve the computational efficiency of new energy power generation modeling. It uses BiGRU for short-term dependence modeling to enhance the model's ability to capture local time series features, thereby optimizing the short-term new energy power generation prediction effect. It fuses the prediction results of Informer and BiGRU, and uses a weighted fusion strategy to construct the final prediction result, achieving collaborative modeling of long-term and short-term dependence relationships. Finally, the predicted values of each IMF and the residual term are added together to obtain the final new energy power generation prediction result, improving the overall prediction accuracy and robustness. Description of the Drawings

[0019] To more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of specific applications or the prior art. Obviously, the drawings in the following description are only those of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0020] Figure 1 It is a flowchart of the present invention.

[0021] Figure 2 It is a curve graph of the photovoltaic power data of the present invention.

[0022] Figure 3 It is a heat map of the correlation between the power and each feature of the present invention.

[0023] Figure 4 It is a graph of each subsequence decomposed by CEEMDAN of the present invention.

[0024] Figure 5 It is a structural diagram of the Informer-BiGRU model of the present invention.

[0025] Figure 6 It is a schematic diagram of the sparse probability self-attention mechanism of Informer of the present invention.

[0026] Figure 7 It is a structural diagram of Informer of the present invention.

[0027] Figure 8 It is a structural diagram of BiGRU of the present invention.

[0028] Figure 9 It is a prediction result graph of Informer-BiGRU combined with CEEMDAN decomposition of the present invention. Detailed Embodiments

[0029] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0030] The present invention covers all alternatives, modifications, equivalent methods and solutions within its core idea and scope. In order to enable the public to have a more comprehensive understanding of the present invention, specific details are described in detail in the following preferred embodiments. However, for those skilled in the art, even without the description of these details, they can fully understand the content of the present invention. In addition, the drawings of the present invention are schematic diagrams and are not drawn exactly to scale. This is hereby stated.

[0031] As Figure 1 shown, the present invention proposes a power prediction method for a new energy power generation system combining CEEMDAN decomposition and Informer-BiGRU model, including the following steps: S1: Collect and preprocess data to form an original data set; S2: Conduct a correlation analysis on the predicted power data and its characteristic data, including irradiance, wind speed, wind direction, temperature, humidity, and pressure; S3: Perform CEEMDAN decomposition on the power data processed in S1 to form subsequence data with multiple different frequency characteristics and time-scale characteristics; S4: Based on the power sequence data processed in S2 and S3, input the decomposed subsequences and the filtered characteristic data into the CEEMDAN decomposition-based Informer-BiGRU combined prediction model; S5: Calculate the error between the power prediction value of the CEEMDAN decomposition-based Informer-BiGRU model and the actual power value, and use MSE, RMSE, and R 2 evaluation indicators for performance evaluation.

[0032] Furthermore, the preprocessing in S1 includes: S11: Data cleaning, using the median method for inspection and supplementation. Calculate the median of a section of the power sequence before the missing value and replace the missing value with it; S12: Normalize the data. Since the characteristic values (such as power and temperature) have different dimensions or magnitudes, linearly transform the data set for normalization processing, map the characteristic data to the interval [0,1], and the normalization processing calculation is: , where is the original data, is the normalized data, represents the minimum value of the original data, represents the maximum value of the original data; The inverse normalization processing calculation is: , wherein, is the normalized power prediction value, is the power prediction value after denormalization.

[0033] Furthermore, the analysis of the correlation between the power and other data in S2 includes: Performing correlation analysis on each sequence based on the Spearman correlation coefficient method, which is used to measure the monotonic relationship between two variables, and its calculation formula is: , wherein, represents the Spearman correlation coefficient between sequence S and sequence K, represents the rank difference of each pair in the two sequences, and n represents the size of the data sample. The larger the value of, the stronger the correlation between sequence S and sequence K, and vice versa.

[0034] Furthermore, the specific processing of the CEEMDAN decomposition algorithm in S3 includes: S31: Initialization: Select appropriate white noise amplitude and number of iterations.

[0035] S32: By adding N Gaussian white noise sequences with a mean of 0 to the original power signal , constructing the sequence to be decomposed for a total of N experiments , and the formula is as follows: , wherein, is the Gaussian white noise weight coefficient; is the white noise sequence added for the th time.

[0036] S33: Applying the EMD algorithm to for decomposition to obtain the first mode component (IMF) and the first unique residual component , and the formula is as follows: , , S34: Adding noise to the residual component obtained after decomposition at the th stage and continuing to apply EMD for decomposition, and the formula is as follows: , , The original power signal sequence is decomposed into N mode components and a residual term Sum all the IMF components and the residual to obtain the decomposition expression of the original signal, as shown in the following formula: .

[0037] Furthermore, the Informer-BiGRU combined prediction model in S4 includes an Informer model and a BiGRU model. Among them, S41: The Informer consists of a sparse self-attention mechanism, self-attention distillation, and generative decoding.

[0038] The self-attention mechanism of the traditional Transformer model calculates the correlation between each element in the input sequence and other elements, and the computational complexity is , where n is the sequence length. When the sequence is very long, the computational amount is huge, resulting in slow training and inference processes.

[0039] To improve efficiency, the Informer introduces a sparse self-attention (ProbSparseAttention) mechanism. Through the Top-k strategy, only the most relevant attention scores are calculated, thus significantly reducing the computational complexity. Specifically, the Informer only calculates the Top-k most relevant attention scores of each position with other positions.

[0040] For the input power subsequence , first obtain the query matrix , key matrix and value matrix through linear transformation, and the calculation formula is: .

[0041] Then calculate the attention score matrix A, and the calculation formula is: , where is the key vector dimension, and A is an n×n matrix. To achieve sparsification, the Informer only calculates the Top-k maximum scores in each row, sets other scores to zero, and reduces the computational complexity from to .

[0042] By only focusing on the most relevant parts, the Informer improves the computational efficiency. The final output result is: , Based on the sparse self-attention calculation, the Informer model adopts self-attention distillation to reduce the dimension of each matrix, thereby further reducing the computational complexity.

[0043] The purpose of distillation is to reduce , , the redundant information in the matrix and improve the efficiency of model calculation through a low-dimensional approximate representation. In this way, the model can retain key attention information while reducing the computational burden. The calculation formula is: , where, , , is , , the compressed matrix, and the matrix will have a lower dimension, making the process of calculating attention more efficient.

[0044] Generative decoding combines the structures of the encoder and decoder. The encoder processes historical data to generate a context representation C, and then the decoder directly generates prediction results for all future time points. This can improve the prediction speed and reduce the error accumulation caused by step-by-step reasoning. Specifically, the core of generative decoding is: , where C is the context representation generated by the encoder and Y is all the prediction results output by the decoder.

[0045] Informer combines three techniques to optimize the calculation process: First, it uses sparse self-attention to reduce the complexity of attention calculation; second, it further compresses the computational volume and memory occupancy through self-attention distillation; finally, it uses generative decoding to improve the prediction speed and accuracy, avoiding interference from irrelevant information in high-frequency data while reducing complexity.

[0046] S42: Input the training samples into the BiGRU model. BiGRU consists of a forward GRU and a backward GRU, which can capture power information from both the past and the future simultaneously, while an ordinary unidirectional GRU can only utilize past information. The advantage of this bidirectional mechanism is to improve the short-term pattern recognition ability and more accurately extract local features of the sequence. In some cases, the information of future time steps has a certain impact on the prediction of the current time step, and BiGRU can utilize the information propagated backward to enhance the feature expression ability. The formula is: , , , , .

[0047] where, To update the activation value of the gate and determine the hidden state at the previous moment For the current moment Degree of influence; Is the input data Weight matrix of the update gate; Is the hidden state at the previous moment Weight matrix of the update gate; Is the bias term of the update gate; Is the activation value of the reset gate; Is the input data Weight matrix of the reset gate; Is the hidden state at the previous moment Weight matrix of the reset gate; Is the bias term of the reset gate; Is the candidate hidden state at the current moment; Is the input data Weight matrix of the candidate hidden state; Is the hidden state at the previous moment Weight matrix of the candidate hidden state; Is the bias term of the candidate hidden state; Is the final state at the current moment; Is the hidden state obtained by forward GRU calculation; Is the hidden state obtained by backward GRU calculation; Is the output of the bidirectional GRU at time t.

[0048] S43: To fully utilize the multi-scale feature extraction ability of CEEMDAN decomposition, the advantages of Informer in long-term dependence modeling, and the ability of BiGRU in short-term local feature modeling to improve the accuracy of predicting new energy power generation, the Informer model and the BiGRU model are combined, and the subsequences obtained by decomposition are combined with other features screened by Spearman to form multiple new data sets and input them into the model. The formula is as follows: , , , Among them, the first formula concatenates the output matrices of Informer and BiGRU; in the second and third formulas , and Are learning parameters, which are learned by model training. Through the feature concatenation of the first formula, the weighted fusion of the second formula, and the fully connected layer of the third formula, the final model prediction result is formed. Finally, each IMF and the residual term are reconstructed to obtain the final new energy power generation prediction result.

[0049] Furthermore, calculate the error between the power prediction value and the actual power value of the Informer-BiGRU model with CEEMDAN decomposition, and use evaluation indicators such as MSE, RMSE, and R 2 for performance evaluation.

[0050] Furthermore, the calculation of the model error and performance evaluation in S5 specifically includes: Use MAE, RMSE, R 2 Three evaluation indicators to evaluate the performance of the prediction model; among them, the power is expressed as input data: , and the output result is expressed as , and its evaluation criteria MAE, RMSE, R 2 The calculation formulas are as follows: , , , Among them, is the actual power test sample value; is the model power prediction result; is the power average; is the total number of power test samples.

[0051] The smaller the values of RMSE and MAE, the higher the accuracy of the predicted power of the model; the closer the value of R 2 is to 1, the closer the power prediction data curve of the model is to the real data curve.

[0052] Example 1: The following further details the implementation process of the present invention in combination with a specific example.

[0053] Step 1: Obtain relevant historical data sets. In this example, the photovoltaic power time series data is selected. Check the data set information and preprocess it. The processed power time series diagram is as Figure 2 shown: There may be data anomalies and data missing in the historical photovoltaic power time series data. Therefore, the median method is used to check and supplement it. The method is to calculate the median of a period of time series before the missing value and replace the missing value or abnormal value with it.

[0054] In addition, before using the data to train the model, the data can be divided into a training set and a test set according to a ratio of 7:2:1 to enhance the generalization ability of the model. At the same time, the data is processed by the normalization method to map the feature data into the interval of [0,1].

[0055] Step 2: Based on Step 1, perform a correlation analysis on the photovoltaic power data and other relevant data (irradiance, wind speed, wind direction, temperature, humidity, and pressure). The correlation heatmap is as shown in Figure 3 shown, and the correlation between each feature and the photovoltaic power is shown in Table 1: Table 1 .

[0056] According to the definition of the correlation coefficient, the closer its absolute value is to 1, the higher the correlation with the power. From the table, it can be seen that the power has a strong correlation with irradiance, temperature, and humidity, and a weak correlation with wind speed, wind direction, and pressure. Therefore, when making predictions, irradiance, temperature, and humidity are selected as influencing factors to highlight their influencing effects, and the influencing factors with weak correlations are removed to improve the performance and interpretability of the model.

[0057] Step 3: Based on the dataset processed in Step 1, use the CEEMDAN decomposition algorithm to decompose the photovoltaic power data sequence to make it stationary. The noise intensity is selected as 0.05, and the number of iterations is 100. The subsequences with different frequencies after decomposition are as shown in Figure 4 shown.

[0058] Due to the non-stationary characteristics of the photovoltaic power data, the algorithm preferentially decomposes the non-stationary characteristics of the photovoltaic power curve to form IMF0. In each iteration, CEEMDAN first extracts the highest-frequency component of the signal, then calculates the residual signal, and then extracts the lower-frequency IMF from the residual. This decomposition method ensures that each IMF mainly contains signals within a certain specific frequency range.

[0059] Step 4: Based on Steps 1, 2, and 3, input the decomposed data and correlation factors (irradiance, temperature, humidity) into the model for prediction.

[0060] The specific process of model training in Step 4 is as follows: In the experimental configuration of this case, the computer has a Corei5 CPU, 16GB of memory, and a Windows10 operating system. Based on the PyCharm platform, use the Python language to implement and build the model under the PyTorch framework to complete the prediction task.

[0061] In the constructed Informer-BiGRU prediction module, the number of Informer encoder blocks is 2, the number of decoder blocks is 1, the number of multi-head attention heads is 8, the number of neurons in the GRU layer is set to 64, the optimizer is set to Adam, the number of model training iterations is set to 100, the activation function is gelu, the learning rate is 0.001, Dropout is 0.1, and when batch_size is 64, the model achieves better prediction results.

[0062] Figure 5 It is the structure diagram of the informer - BiGRU combined model. The specific data steps in the model are as follows: Input the decomposed photovoltaic power subsequence data and feature data (irradiance, temperature, humidity) into the Informer - BiGRU model. Figure 6 , Figure 7 It is the structure diagram of the sparse self - attention mechanism and the Informer structure diagram. The Informer is mainly used for time - series prediction of long - term dependencies and introduces a sparse self - attention mechanism to improve computational efficiency. Figure 8 It is the BiGRU structure diagram. BiGRU is mainly used to extract local time - series features and enhance short - term dependencies. GRU (Gated Recurrent Unit) is a lightweight version of LSTM, which uses two gating mechanisms (update gate and reset gate) to optimize the learning process of time - series data. Finally, using the long - term dependence features of the Informer and the short - term dependence features of BiGRU, a fusion strategy is adopted to splice and weight - combine their outputs, and after passing through a fully - connected layer, the predicted value of the photovoltaic power at the next moment is output.

[0063] Step 5: Calculate the differences between the predicted values of multiple models and the actual true values to prove the effectiveness and accuracy of the model of the present invention. Figure 9 It is the prediction result diagram of the Informer - BiGRU combined with CEEMDAN decomposition of the present invention.

[0064] The specific processing process of the model prediction error test in step 5 is as follows: To verify the effectiveness of the proposed method, this study respectively uses LSTM, BiLSTM, BiGRU, Informer and their combined models for comparative experiments, and further combines CEEMDAN decomposition on this basis to explore its improvement effect on prediction performance. The experiments use RMSE (Root Mean Square Error), R 2 (Coefficient of Determination) and MAE (Mean Absolute Error) as evaluation indicators, and the experimental results are shown in the table.

[0065] The LSTM model shows certain ability in time - series modeling, but is limited by the unidirectional information flow. BiLSTM and BiGRU significantly improve the ability to capture short - term dependencies through bidirectional information propagation, and their performance is better than that of LSTM. The specific data is as follows: Compared with LSTM, BiLSTM reduces RMSE by 5.62%, increases R² by 3.36%, and reduces MAE by 7.94%. BiGRU performs better in terms of indicators such as RMSE (1.73), R² (80.59%), MAE (1.23), indicating that it has stronger feature extraction ability and generalization performance in power prediction.

[0066] The Informer significantly improves the modeling ability of long-term dependencies through the sparse self-attention mechanism, and its performance is better than that of the model without CEEMDAN. The Informer-BiGRU further optimizes the short-term time series dependency modeling ability, and the specific data is as follows: The RMSE of the Informer-BiGRU is 1.45, the R² is 86.43%, and the MAE is 1.00, showing a significant improvement compared with the Informer model.

[0067] To verify the role of decomposition, the above models are all combined with CEEMDAN decomposition. The experimental results prove that the performance of the models has been improved after decomposition. Especially for the CEEMDAN-Informer model, its RMSE has improved by 43.53% compared with the Informer model, proving the role of CEEMDAN decomposition in power prediction. As can be seen from the table, the proposed model CEEMDAN-Informer-BiGRU has significant advantages in modeling short-term dependencies, long-term trends, and reducing computational complexity, achieving the best prediction performance (RMSE is 0.90, R² is 94.75%, and MAE is 0.51).

[0068] Table 2 。

[0069] In summary, as the optimal model proposed in the present invention, CEEMDAN-Informer-BiGRU shows higher accuracy and stability in the photovoltaic power prediction task, and can provide a more reliable prediction tool for practical applications.

[0070] Example 2: This example proposes an electronic system, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method steps of the present invention.

[0071] It should be noted that the processing flow of the Example 2 corresponds to the specific steps of the method provided in the Embodiment 1 of the present invention, and has the corresponding functional modules and beneficial effects for executing the method. For technical details not described in detail in this example, reference can be made to the method provided in the Embodiment of the present invention.

[0072] The program code for implementing the methods of the present application can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing devices, such that when the program codes are executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program codes can be executed entirely on the machine, partially on the machine, executed partially on the machine as an independent software package and partially on a remote machine, or executed entirely on a remote machine or server.

[0073] In the context of the present application, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. The machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of the machine-readable storage medium would include electrical connections based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0074] Those of ordinary skill in the art should understand that the above embodiments are only exemplary descriptions of the present invention, aiming to illustrate the technical solutions of the present invention rather than limit the scope of the present invention. Without departing from the spirit of the present invention, the above embodiments and their different technical features can be appropriately combined, the relevant steps can be executed in different orders, and can also be adaptively adjusted according to actual needs. In addition, the concept of the present invention covers various variations, modifications, substitutions, and equivalent solutions that can be foreseen by those skilled in the art, including but not limited to the optimized design of materials, structures, shapes, processes, algorithms, and functions. Any changes made within the scope of the substantial content and basic principles of the present invention shall be regarded as falling within the protection scope of the present invention.

Claims

1. A power prediction method for a new energy power generation system combining CEEMDAN decomposition and Informer-BiGRU model, characterized in that, It includes the following steps: S1. Collect and preprocess historical new energy power generation data and related meteorological data, perform data cleaning and normalization to obtain the original dataset; S2. Conduct a correlation analysis on the predicted power data and its characteristic data, and screen out the highly correlated characteristic data; S3. Perform CEEMDAN decomposition on the dataset obtained in S1 to form subsequence data with different frequency characteristics and time scale features; S4. Construct a combined prediction model integrating Informer and BiGRU, input the decomposed subsequence data and the screened characteristic data into the combined prediction model, and obtain the prediction results of the subsequences through feature splicing and linear layer mapping; S5. Obtain the final predicted value by directly adding the prediction results of all components.

2. The method according to claim 1, characterized in that, Step S2 includes the following steps: S201. Calculate the Spearman correlation coefficient between new energy power generation and multiple influencing factors; S202. Visualize the correlation between variables through a correlation heatmap, screen out highly correlated features as model inputs according to the Spearman correlation coefficient, and remove redundant features at the same time.

3. The method according to claim 2, characterized in that, The calculation formula of the Spearman correlation coefficient is as follows: ， Among them, represents the Spearman correlation coefficient between sequence S and sequence K, represents the rank difference of each pair in the two sequences, n represents the size of the data sample, The larger the value of, the stronger the correlation between sequence S and sequence K, and vice versa, the weaker the correlation.

4. The method according to claim 1, characterized in that In step S3, perform CEEMDAN decomposition on the dataset obtained in S1, which specifically includes the following steps: (1) Initialization: Select appropriate white noise amplitude and number of iterations; (2)Add white noise: Add Gaussian white noise to the original power signal to construct multiple preprocessing sequences; (3) EMD decomposition: For each preprocessed sequence perform EMD decomposition to obtain the first-order IMF component , and calculate its mean as the first-order IMF, and at the same time calculate the residual; (4) Iterative processing: Input the residual into new white noise, repeat the EMD decomposition to obtain the next-order IMF, and calculate the new residual; (5) Repeat iteration: Repeat the above process until the stop condition is met, and finally obtain all IMF components; Add all IMF components and the residual to obtain the decomposition expression of the original power signal.

5. The method according to any one of claims 1 or 4, characterized in that The calculation formula of CEEMDAN decomposition is as follows: By adding a Gaussian white noise sequence with a mean of 0 N times to the original power signal a sequence to be decomposed for a total of N experiments is constructed : ， Among them, is the Gaussian white noise weight coefficient; is the th added white noise sequence, ; Pair Decompose using the EMD algorithm to obtain the first mode component IMF and the first unique residual component : ， ， Add noise to the residual components obtained after decomposition in the j phase and continue to decompose them using EMD: ， ， Among them, j = 2, 3,..., N； Repeat the decomposition until the termination condition is met. The termination criterion is that the number of extreme points of the residual signal does not exceed 2 at most; Finally, the original power signal sequence is decomposed into N modal components and a residual term : 。 6. The method according to claim 1, wherein Step S4 constructs a combined prediction model integrating Informer and BiGRU, including: (1) Combine the subsequences obtained by decomposition with other features screened by Spearman to form multiple new datasets and input them into the model; (2) Adopt the Informer structure for global time series modeling; (3) Adopt the BiGRU structure for short-time series dependence relationship modeling to obtain local power sequence features; (4) By fusing the long-term dependence features of Informer and the short-term dependence features of the output of BiGRU, splice and weight-combine the outputs of the two, and train different weight ratios to achieve the final prediction of the new energy power generation subsequence.

7. The method according to claim 6, characterized in that The formula of the Informer-BiGRU combined prediction model is as follows: (1) Adopt the Informer structure for global time series modeling: For the input sequence , first obtain the query matrix , the key matrix and the value matrix through a linear transformation: ， ， ， Sparse self-attention output results and pass them to the self-attention distillation module: ， ， Among them, is the dimension of the key vector, A is an n×n matrix, that is, the attention score matrix; Compress by low-dimensional approximate representation , , Matrix: ， ， ， Among them, , , is , , the compressed matrix; The core formula of generative decoding: ， The encoder processes historical data to generate a fixed-dimensional representation, and the decoder directly generates all future prediction results. Among them, C is the context representation generated by the encoder, and Y is all prediction results output by the decoder; (2) Adopt the BiGRU structure for short-time series dependence relationship modeling: ， ， ， ， ； Among them, is the activation value of the update gate, which determines the hidden state at the previous moment for the current moment influence degree; is the input data weight matrix of the update gate; is the hidden state at the previous moment weight matrix of the update gate; is the bias term of the update gate; is the activation value of the reset gate; is the input data weight matrix of the reset gate; is the hidden state at the previous moment weight matrix of the reset gate; is the bias term of the reset gate; is the candidate hidden state at the current moment; is the input data weight matrix of the candidate hidden state; is the hidden state at the previous moment weight matrix of the candidate hidden state; is the bias term of the candidate hidden state; is the final state at the current moment; is the hidden state obtained by forward GRU calculation; is the hidden state obtained by backward GRU calculation; is the output of the bidirectional GRU at time t (3) Feature concatenation and weighted fusion are performed on the output matrices of Informer and BiGRU: ， ， The final power prediction result is formed through a fully connected layer: ， Among them, , , are learning parameters, which are obtained by self-learning through model training.

8. The method according to claim 1, characterized in that, The training process of the Informer-BiGRU combined prediction model includes: (1) Use the PyTorch framework for construction and training; (2) Divide the training set, validation set, and test set in a ratio of 7:2:1 to enhance the generalization ability of the model; (3) Use the trial-and-error method to adjust the hyperparameters to improve the prediction accuracy of the model; (4)Adopt the mean absolute error MAE, root mean square error RMSE, and coefficient of determination Calculate the percentage of the power prediction error in the actual value to measure the relative error.

9. The method according to claim 8, characterized in that, The calculation formula for the evaluation criteria of the prediction model is as follows: ， ， ， Among them, is the actual power test sample value; is the model power prediction result; is the power average; is the total number of power test samples.

10. An electronic system, comprising: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method steps described in any one of claims 1-9.

Images