Short-term photovoltaic power prediction method based on characteristic decomposition reconstruction

Abstract

The invention discloses a short-term photovoltaic power prediction method based on characteristic decomposition reconstruction, and the method comprises the steps: decomposing historical photovoltaic output power data into a plurality of intrinsic mode function sequences, screening out IMF components highly related to a prediction target through a Pearson's correlation coefficient method, and effectively eliminating noise interference; meanwhile, the screened IMF components are added and then denoised new features are added into the data set, the original data set is further enriched, higher-quality input is provided for the model, the capturing capacity of the model on the internal characteristics of the data in the training stage is remarkably enhanced, and then the prediction stability and accuracy are improved; besides, the Mamba model is adopted to predict the photovoltaic power generation power, input characteristics can be efficiently processed, intelligent screening and reservation of information are achieved, an accurate short-term photovoltaic power generation power prediction result is output, and powerful support is provided for operation and scheduling of a photovoltaic power generation system.

Description

Technical Field

[0001] The present invention belongs to the technical field of renewable energy power generation, and in particular relates to a short-term photovoltaic power prediction method based on characteristic decomposition and reconstruction. Background Art

[0002] Photovoltaic power generation, as a clean and renewable form of energy, is an indispensable aid to achieving the goals of carbon peak and carbon neutrality. By actively promoting photovoltaic power generation projects, we can not only effectively reduce greenhouse gas emissions, but also accelerate the transformation of energy structure and pave the way for sustainable development. However, photovoltaic power generation faces the impact of fluctuations in solar irradiance and other environmental factors, showing intermittent and unpredictable characteristics, which undoubtedly brings severe challenges to the planning and operation of power grids. Accurate photovoltaic power generation prediction technology can significantly improve the stability of the power grid and optimize energy scheduling strategies. Therefore, improving the accuracy of photovoltaic power generation prediction is crucial to fully tap the huge potential of solar energy in the energy field. Photovoltaic power generation prediction models can be roughly divided into three categories: physical models, statistical models, and artificial intelligence models, each of which has unique advantages and limitations. The physical model is based on the basic principles of photovoltaic power generation, combined with physical factors such as astronomical data, meteorological information, and the specific design parameters of the photovoltaic system to construct a photovoltaic power prediction model. The advantage of this type of model is that its prediction results have a clear physical basis and perform well when weather conditions are stable. However, its modeling process is extremely complex, and it is necessary to deeply consider the physical characteristics of the photovoltaic system and the changing weather conditions, resulting in relatively high implementation and maintenance costs. In contrast, statistical models mainly rely on historical data and use specific algorithms to establish a mapping relationship between input and output to achieve photovoltaic power prediction. However, photovoltaic power generation data presents complex time series characteristics, including dynamic and non-periodic patterns, which increases the difficulty of establishing a mapping relationship between historical data and prediction targets. Therefore, although statistical models have certain application value in some aspects, it is still a challenge to obtain accurate photovoltaic power prediction results through them.

[0003] With the rapid progress of artificial intelligence technology, deep learning has firmly occupied a core position in the field of photovoltaic power generation prediction. Researchers actively use a variety of recursive neural network models, such as RNN (recurrent neural network), LSTM (long short-term memory network) and GRU (gated recurrent unit), to accurately predict photovoltaic power generation. In addition, there are composite models such as CNN-LSTM, CNN-GRU and TCN-Transformer. These models cleverly stack and fuse different types of neural networks to allow each model to fully demonstrate its unique advantages and achieve complementary advantages, thereby further improving the prediction accuracy. At the same time, in order to seek better prediction performance and lower time cost, meta-heuristic algorithms have also been introduced into this field. Through the intelligent search of meta-heuristic algorithms, researchers can more effectively find the optimal solution of the model, thereby ensuring that the model achieves the best performance in photovoltaic power generation prediction. This series of innovative measures not only promotes the rapid development of photovoltaic power generation prediction technology, but also provides strong support for the efficient use of renewable energy. Although the various optimization methods implemented for the above single models have been widely verified to improve the accuracy of prediction to a certain extent, these basic models still encounter insurmountable inherent limitations in the optimization process. In addition, the characteristics of photovoltaic power generation include complex factors such as irregularity, randomness, and suddenness, which makes it difficult to fully and deeply capture all the information details in historical data through model training alone.

[0004] In order for the model to accurately predict photovoltaic power, it is necessary to fully capture the information characteristics of the data during the model training process and highlight the impact of key weather conditions. However, in the process of optimizing a single model, it is not possible to avoid the inherent defects of the original model. For example, CNN (convolutional neural network) is limited by the size of its convolution kernel and is unable to predict edge data. RNN (recurrent neural network), LSTM (long short-term memory network) and its derivative version XLSTM follow their inherent serial computing mode and are difficult to make breakthroughs in parallel processing, which leads to low computing efficiency and long time consumption. As for the Transformer model, when dealing with time series prediction tasks involving exogenous variables, its performance is often unsatisfactory and it is difficult to fully and deeply capture the complex and multivariate correlations within the data. The existence of these inherent defects undoubtedly sets up many obstacles for further leaps in model performance. Summary of the invention

[0005] The purpose of the present invention is to provide a short-term photovoltaic power prediction method based on eigendecomposition and reconstruction.

[0006] In a first aspect, the present invention provides a short-term photovoltaic power prediction method based on eigendecomposition and reconstruction, which comprises the following steps: Step 1: Construct a data set; the data set includes environmental characteristic factors of the measured area and their corresponding historical photovoltaic output power; Step 2: extract key characteristic factors from environmental characteristic factors, and reconstruct the data set with key characteristic factors and historical photovoltaic output power; Step 3: Perform signal decomposition on the historical photovoltaic output power in the data set to obtain all intrinsic mode components of each sample ; According to the eigenmode components The correlation with the historical photovoltaic output power is used to filter the intrinsic mode components from high to low, and the result of adding the filtered intrinsic mode components is added to the data set as a new denoising feature; Step 4: Use the data set obtained in step 3 to train a photovoltaic power prediction model, using key environmental factors, historical photovoltaic output power and denoised new features as inputs to the photovoltaic power prediction model; Step 5: Use the trained photovoltaic power prediction model to predict the power generation in the measured area.

[0007] Preferably, in step 3, the process of signal decomposition is: Step 3-1. Take the photovoltaic output power at different times as the original signal sequence ; To the original signal sequence Add N Gaussian white noise to obtain N preprocessed sequences , preprocessing sequence The expression is: in, is the standard deviation of the noise; For the The Gaussian white noise added is (0,1) distribution; ; N is the number of times white noise is added; Step 3-2. According to the preprocessing sequence Get the first eigenmode component , whose expression is: in, is the first-order modal component; Get the first eigenmode component removed The residual component after , whose expression is: Step 3-3. In the k-1th residual component Add N times Gaussian white noise to obtain a new sequence , obtain the kth eigenmode component according to the new sequence , whose expression is: in, To perform k times of EMD operation; Get the kth eigenmode component removed The residual component after , whose expression is: After K times of decomposition, the original signal sequence is broken down into: Where K is the number of times the empirical mode decomposition is performed.

[0008] Preferably, in the step 2, before extracting the key characteristic factors, outlier detection and data null value filling are performed on the environmental characteristic factors and the historical photovoltaic output power respectively.

[0009] Preferably, the method for outlier detection and data null value filling is as follows: The standard score method is used to detect outliers in the original data, and the standard score z of each data point is calculated: in, is the value of the data point; is the mean of the data set; is the standard deviation of the data set; If the standard score z of the acquired data point is greater than the preset maximum standard score, its value is replaced; The mean interpolation method is used to process the data null values ​​in the data set. The data null values ​​obtained by the mean interpolation method The calculation formula is as follows: in, is the responsiveness vector; is the value of each data point; n is the number of all data; is the number of valid data; .

[0010] Preferably, in step 2, the method for extracting key characteristic factors is: obtaining the Pearson correlation coefficient between all environmental characteristic factors and historical photovoltaic output power, and selecting environmental characteristic factors whose Pearson correlation coefficient is greater than a preset threshold as key characteristic factors.

[0011] Preferably, in step 3, the method for selecting some eigenmode components is: obtaining all eigenmode components With the original signal sequence The Pearson correlation coefficient between them is calculated, and the IMF components whose Pearson correlation coefficient is greater than the preset threshold are selected as the screened out eigenmode components.

[0012] Preferably, in step 4, the photovoltaic power prediction model adopts the Mamba model.

[0013] Preferably, in step one, the environmental characteristic factors include wind speed, temperature, humidity, global horizontal irradiance, diffuse horizontal irradiance, wind direction, and rainfall.

[0014] In a second aspect, the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores the computer program; and the processor executes the above-mentioned short-term photovoltaic power prediction method.

[0015] In a third aspect, the present invention provides a readable storage medium storing a computer program; when the computer program is executed by a processor, it is used to implement the above-mentioned short-term photovoltaic power prediction method.

[0016] The present invention has the following beneficial effects: 1. The present invention decomposes the historical photovoltaic output power data into multiple intrinsic mode function sequences, and uses the Pearson correlation coefficient method to screen out the IMF components that are highly correlated with the prediction target, thereby effectively eliminating noise interference; at the same time, the new denoising features are added to the data set after adding the screened IMF components, further enriching the original data set and providing higher quality input for the model; it not only deepens the mining and utilization of the intrinsic information of the data, but also significantly enhances the model's ability to capture the intrinsic characteristics of the data during the training stage, thereby improving the stability and accuracy of the prediction.

[0017] 2. The present invention adopts the Mamba model to predict photovoltaic power generation, which can efficiently process the input features, realize the intelligent screening and retention of information, and output accurate short-term photovoltaic power generation prediction results, providing strong support for the operation and scheduling of photovoltaic power generation systems; at the same time, the Mamba model makes full use of the parameterized SSM input mechanism of the model, so that the model can dynamically adjust its weight matrix during the testing phase, flexibly respond to changes in input data, thereby effectively filtering out data that is not related to the prediction task, accurately focusing on information that is critical to the current prediction task, and improving the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of the present invention.

[0019] Figure 2 This is a decomposition diagram of the sunny day data by the adaptive noise complete set empirical mode decomposition in the present invention.

[0020] Figure 3 It is the power waveform diagram of the present invention under different weather conditions.

[0021] Figure 4 The following are comparison charts of prediction results of different models under different weather conditions. (a) is a comparison chart of prediction results under sunny conditions; (b) is a comparison chart of prediction results under sunny to cloudy conditions; (c) is a comparison chart of prediction results under cloudy conditions; and (d) is a comparison chart of prediction results under rainy conditions.

[0022] Figure 5 Radar charts of evaluation indicators of different models under sunny conditions; among them, (a) is the determination coefficient radar chart; (b) is the mean square error radar chart; (c) is the mean absolute error radar chart; (d) is the root mean square error radar chart; and (e) is the computational efficiency radar chart. DETAILED DESCRIPTION

[0023] The present invention will be further described below in conjunction with the accompanying drawings.

[0024] like Figure 1 As shown, a short-term photovoltaic power prediction method based on eigendecomposition reconstruction includes the following steps: Step 1: Get the dataset Acquire multiple groups of historical data of the measured area (in this embodiment, the Alice Springs photovoltaic power station in central Australia), and construct a data set with each group of historical data as a sample; each group of historical data includes environmental characteristic factors and corresponding photovoltaic output power. Environmental characteristic factors include wind speed, temperature, humidity, global horizontal irradiance, diffuse horizontal irradiance, wind direction, and rainfall; the data time resolution of the sampling point is 5 minutes, and there are 288 data sampling points per day. According to the daily weather conditions and power fluctuations, the photovoltaic output power under four weather conditions of sunny, sunny to cloudy, cloudy, and rainy is predicted.

[0025] Step 2: Data preprocessing 2-1. Detecting outliers The Z-Score (standard score) method was used to detect outliers in the original data, and the standard score z of each data point was calculated: in, is the value of the data point; is the mean of the data set; is the standard deviation of the data set.

[0026] If the obtained standard score z is greater than the preset maximum standard score, its value will be replaced to ensure the availability of the data.

[0027] 2-2. Filling data with empty values The mean interpolation method is used to process the data null values ​​in the data set. The data null values ​​obtained by the mean interpolation method The calculation formula is as follows: in, is the responsiveness vector; is the value of each data point; n is the number of all data; is the number of valid data; .

[0028] 2-3. Select key characteristic factors The Pearson correlation coefficient method is used to select environmental characteristic factors that are strongly correlated with photovoltaic output power. The Pearson correlation coefficient formula is as follows: in, is the Pearson correlation coefficient, ; is the sample size; and The two variables are The value in the sample; and are the sample means of the two variables.

[0029] Environmental characteristic factors with a Pearson correlation coefficient greater than 0.4 were selected as key characteristic factors, and the key characteristic factors in the data set were retained to reconstruct the data set.

[0030] 2-4. Normalization Normalize the data in the dataset to reduce the data range of all features to a uniform scale, thereby reducing the adverse effects of the dimension differences between feature data measured in different units on the prediction results. The expression is:

[0031] in, is the original data in the dataset; is the maximum value in the original data; is the minimum value in the original data.

[0032] Step 3: Feature Decomposition and Reconstruction like Figure 2As shown, the historical photovoltaic output power is decomposed into signals. The specific steps are as follows: 3-1. Using photovoltaic output power at different times as the original signal sequence ; To the original signal sequence Add N Gaussian white noise to obtain N preprocessed sequences , preprocessing sequence The expression is: in, is the standard deviation of the noise; For the The Gaussian white noise added is (0,1) distribution; ; N is the number of times white noise is added.

[0033] 3-2. For all preprocessed sequences Perform EMD decomposition (empirical mode decomposition) to obtain the first-order modal component , according to the first-order modal component Get the first eigenmode component , whose expression is: Get the first eigenmode component removed The residual component after , whose expression is: 3-3. In the residual component Add N times Gaussian white noise to obtain a new sequence , perform EMD decomposition on the new sequence and take the mean to obtain the second eigenmode component , whose expression is: in, To perform k EMD decomposition operations.

[0034] Get the second eigenmode component removed The residual component after , whose expression is: 3-4. In the k-1th residual component Add N times Gaussian white noise to obtain a new sequence , perform EMD decomposition on the new sequence and take the mean to obtain the kth eigenmode component , whose expression is: Get the kth eigenmode component removed The residual component after , whose expression is: 3-5. After K times of decomposition, the original signal sequence is broken down into: Where K is the number of EMD decompositions performed.

[0035] Get all the eigenmode components after decomposition With the original signal sequence The Pearson correlation coefficient between them is used to select the IMF components with a Pearson correlation coefficient greater than 0.4, and they are added together as new denoising features to the dataset. The dataset is divided into training set, validation set, and test set according to the ratio of 7:2:1.

[0036] Step 4. Use key environmental factors, historical photovoltaic output power and denoised new features as the input of the Mamba model, and use real-time photovoltaic output power as the output of the Mamba model. Input the training set into the Mamba model, train the Mamba model, and use the validation set to verify the model parameter update results; the Mamba model includes a one-dimensional convolution layer, an SSM layer, and a linear layer. The one-dimensional convolution layer is used to extract local features of the input, and input the extracted local features into the SSM layer; the SSM layer converts the dimensions of the two weight matrices B and C of the linear layer into (M, L, N) respectively, and converts the dimension of the input stage retention (step length) into (M, L, D); where M is the training batch, L is the time series step length, and D is the input dimension; secondly, the discretization operation is used to obtain , ,make , It becomes input data dependent, realizing data-driven generation. The output of the SSM layer is combined with the residual connection, and the final prediction result is output through a linear layer.

[0037] In addition, by taking advantage of the small memory usage of SSM itself, the weight matrix and step size are directly loaded into the cache area (SRAM) of the graphics card, and a series of operations such as discretization are performed in the SRAM to obtain the output results in one step, and the final results are efficiently written back to the high-bandwidth memory (HBM) of the graphics card. This series of optimization measures greatly reduces the number of data read and writes, significantly speeds up the calculation process, and effectively shortens the time cost.

[0038] Step 5: Use the test set to test the present invention and the existing model respectively. The experimental results are as follows: Figure 3 and Figure 4 As shown in the figure, the coefficient of determination (R-Squared, R2), mean square error (Mean Squared Error, MSE), mean absolute error (Mean Absolute Error, MAE) and root mean square error (Root Mean Squard Error, RMSE) are used to evaluate the prediction results of the model respectively. The specific formulas are as follows: in, and They are The actual value and predicted value at the moment, is the average of the true values, is the total number of test samples; , , , They are the coefficient of determination, mean square error, mean absolute error and root mean square error.

[0039] The prediction results of different models under different weather conditions are as follows Figure 5 And as shown in Table 1.

[0040] Table 1 Comparison of evaluation indicators of different models under different weather conditions It can be seen from Table 1 that under all weather conditions, the model proposed in the present invention shows the best performance in terms of R2 index. Figure 4 and Figure 5It can be seen that the present invention presents the best fitting state in the stable period, rising period and even peak period of power output. Specifically, compared with other models, the R2 index of the model proposed by the present invention under four different weather conditions has increased by an average of 0.74%, 2.19%, 2.71% and 2.56%. It is particularly noteworthy that the improvement effect of the model is more significant under poor weather conditions. This outstanding performance is mainly attributed to the new denoising features added to the model. By carefully screening and integrating the IMF components with high correlation to form new feature inputs, the present invention can more accurately capture and learn the key information of the power peak period during the training process. Therefore, no matter at which stage, the present invention can maintain a high degree of fit. And with its unique data processing mechanism, it realizes intelligent screening and retention of information. Through the innovative parameterized matrix method, the model can still flexibly adjust its weight matrix during the test phase, so as to focus more on the core information in the input data and significantly enhance the accuracy of the prediction. Specifically, using the mean square error (MSE) as the measurement standard, the prediction error of the model under different weather conditions was reduced by an average of 2.55%, 7.76%, 6.98% and 0.57%; while in terms of the mean absolute error (MAE), the average reduction was 10.45%, 8.78%, 8.06% and 1.17% respectively. These data fully demonstrate the significant improvement in the prediction accuracy of the model, especially in the face of complex and changeable weather conditions, its prediction ability is still robust and reliable. It is worth noting that under rainy conditions, due to the relatively weak light intensity and irradiance, the power value is generally not high, which also explains to a certain extent the phenomenon of relatively low values ​​in the above indicators. The present invention not only achieves a significant improvement in prediction accuracy, but also makes a major breakthrough in training efficiency. As clearly shown in the data in Table 7, the model takes only 4.5 seconds at the lowest and no more than 7 seconds at the highest in a single training cycle (epoch). Compared with other models, this training speed is 3.9 to 5.9 times faster, showing a significant time advantage.

Claims

1. A short-term photovoltaic power prediction method based on eigendecomposition and reconstruction, characterized by: The following steps are involved: Step 1: Construct a data set; the data set includes environmental characteristic factors of the measured area and their corresponding historical photovoltaic output power; Step 2: extract key characteristic factors from environmental characteristic factors, and reconstruct the data set with key characteristic factors and historical photovoltaic output power; Step 3: Perform signal decomposition on the historical photovoltaic output power in the data set to obtain all intrinsic mode components of each sample ; According to the eigenmode components The correlation with the historical photovoltaic output power is used to filter the intrinsic mode components from high to low, and the result of adding the filtered intrinsic mode components is added to the data set as a new denoising feature; Step 4: Use the data set obtained in step 3 to train the photovoltaic power prediction model; Step 5: Use the trained photovoltaic power prediction model to predict the power generation in the measured area.

2. A short-term photovoltaic power prediction method based on eigendecomposition and reconstruction according to claim 1, characterized in that: In the step 3, the signal decomposition process is as follows: Step 3-1. Take the photovoltaic output power at different times as the original signal sequence ; To the original signal sequence Add N Gaussian white noise to obtain N preprocessed sequences , preprocessing sequence The expression is: in, is the standard deviation of the noise; For the The Gaussian white noise added is (0,1) distribution; ; N is the number of times white noise is added; Step 3-2. According to the preprocessing sequence Get the first eigenmode component , whose expression is: in, is the first-order modal component; Get the first eigenmode component removed The residual component after , whose expression is: Step 3-3. In the k-1th residual component Add N times Gaussian white noise to obtain a new sequence , obtain the kth eigenmode component according to the new sequence , whose expression is: in, To perform k times of EMD operation; Get the kth eigenmode component removed The residual component after , whose expression is: After K times of decomposition, the original signal sequence is broken down into: Where K is the number of times the empirical mode decomposition is performed.

3. The short-term photovoltaic power prediction method based on eigendecomposition and reconstruction according to claim 1 is characterized by: In the step 2, before extracting the key characteristic factors, outlier detection and data null value filling are performed on the environmental characteristic factors and the historical photovoltaic output power respectively.

4. The short-term photovoltaic power prediction method based on eigendecomposition and reconstruction according to claim 3 is characterized by: The method of outlier detection and data null value filling is as follows: The standard score method is used to detect outliers in the original data, and the standard score z of each data point is calculated: in, is the value of the data point; is the mean of the data set; is the standard deviation of the data set; If the standard score z of the acquired data point is greater than the preset maximum standard score, its value is replaced; The mean interpolation method is used to process the data null values ​​in the data set. The data null values ​​obtained by the mean interpolation method The calculation formula is as follows: in, is the responsiveness vector; is the value of each data point; n is the number of all data; is the number of valid data; .

5. The short-term photovoltaic power prediction method based on eigendecomposition and reconstruction according to claim 1 is characterized by: In the step 2, the method for extracting key characteristic factors is: obtaining the Pearson correlation coefficient between all environmental characteristic factors and the historical photovoltaic output power, and selecting environmental characteristic factors whose Pearson correlation coefficient is greater than a preset threshold as key characteristic factors.

6. The short-term photovoltaic power prediction method based on eigendecomposition and reconstruction according to claim 2 is characterized by: In the step 3, the method of selecting some eigenmode components is: obtaining all eigenmode components With the original signal sequence The Pearson correlation coefficient between them is calculated, and the IMF components whose Pearson correlation coefficient is greater than the preset threshold are selected as the screened out eigenmode components.

7. The short-term photovoltaic power prediction method based on eigendecomposition and reconstruction according to claim 1 is characterized by: In the step 4, the photovoltaic power prediction model adopts the Mamba model.

8. The short-term photovoltaic power prediction method based on eigendecomposition and reconstruction according to claim 1 is characterized by: In the step 1, the environmental characteristic factors include wind speed, temperature, humidity, global horizontal irradiance, diffuse horizontal irradiance, wind direction, and rainfall.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: The memory stores a computer program; the processor executes a short-term photovoltaic power prediction method based on eigendecomposition and reconstruction as described in any one of claims 1 to 8.

10. A readable storage medium storing a computer program; characterized in that: When the computer program is executed by a processor, it is used to implement a short-term photovoltaic power prediction method based on eigendecomposition and reconstruction as described in any one of claims 1 to 8.

Images