Multi-time scale photovoltaic power generation power prediction method based on multi-mode decomposition

Abstract

The invention discloses a multi-time-scale photovoltaic power generation power prediction method based on multi-modal decomposition, and the method comprises the steps: carrying out the signal decomposition of historical power data through employing different signal processing methods, obtaining two sub-sequence sets, and adding the two sub-sequence sets into a data sample set as new features, the problem that a single decomposition method may omit some modal subsequences easy to aliasing is effectively avoided, and the problems that the modal decomposition subsequences may have less information amount and insufficient feature information are solved. Meanwhile, the two obtained sub-sequence sets are further subjected to signal decomposition, a signal decomposition result serves as a new feature to be added into a data sample set, and the sub-sequences obtained through one-step decomposition and two-step decomposition, original photovoltaic power generation power data and weather features are integrated, so that diversified and information-clear features are effectively extracted; and the quality of a decomposition result is improved, so that the prediction performance is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power generation prediction, and in particular relates to a multi-time scale photovoltaic power generation prediction method based on multiple modal decompositions. Background Art

[0002] With the advancement of globalization and industrialization, energy demand continues to rise, and the consumption of fossil fuels has led to an increase in greenhouse gas emissions, exacerbating global warming and affecting biodiversity and the human environment. In this context, the development of renewable energy is particularly important. Photovoltaic power generation has become the key to the development of a low-carbon economy with its advantages of being clean and pollution-free. However, photovoltaic power generation is affected by the intermittent solar radiation and environmental fluctuations, which will cause the problem of unstable output power, which brings many challenges to the operation, scheduling and planning of the power system. For this reason, photovoltaic power generation power prediction has become a research focus.

[0003] Photovoltaic power generation prediction can be divided into ultra-short-term (several seconds to 30 minutes), short-term (30 to 360 minutes), medium-term (6 to 24 hours) and long-term (more than 24 hours) predictions, which are respectively applied to real-time scheduling, economic load scheduling, system maintenance and long-term planning. Mainstream prediction methods include physical models, statistical models and artificial intelligence models. Physical models are based on solar radiation characteristics and photovoltaic system parameters, and have strong theoretical characteristics but high computing resource requirements; statistical methods such as autoregression and fuzzy theory are suitable for short-term predictions, but have limited ability to handle nonlinear relationships; artificial intelligence models (especially deep learning) have become a research hotspot in recent years due to their powerful nonlinear modeling capabilities and big data training advantages. Deep learning models such as recursive neural networks (RNNs) and their improved versions, long short-term memory networks (LSTMs) and gated recurrent units (GRUs) have outstanding performance in photovoltaic power generation prediction. LSTM handles long-term dependencies through a gating mechanism, and GRU further simplifies the structure, but the accuracy will still decline in medium- and long-term predictions. The Transformer model based on the self-attention mechanism and its variant Informer optimize the long-term series prediction performance by improving the attention mechanism, but the effect is limited in short-term prediction.

[0004] To improve model accuracy, some researchers have extracted features through modal decomposition technology and combined it with deep learning models to build a hybrid method. However, the features provided by a single decomposition method are limited, so the improvement in model performance is limited, and some high-frequency subsequences decomposed may show irregularity, which makes it difficult to identify signal features. Summary of the invention

[0005] The purpose of the present invention is to provide a multi-time scale photovoltaic power generation prediction method based on multiple modal decompositions.

[0006] In a first aspect, the present invention provides a multi-time scale photovoltaic power generation prediction method based on multiple modal decompositions, which comprises the following steps: Step 1: Obtain and extract historical power data and meteorological data of photovoltaic power stations to construct a data sample set; Step 2: Use different signal processing methods to decompose the historical power data, obtain subsequence set A and subsequence set B, and add them to the data sample set as new features; Step 3: Perform signal decomposition on the two subsequence sets obtained in step 2 respectively to obtain subsequence set C and subsequence set D, and add them into the data sample set as new features; Step 4: Build a photovoltaic power generation prediction model, and use the data sample set obtained in step 3 to train the photovoltaic power generation prediction model; Step 5: Use the trained photovoltaic power prediction model to predict the power generation of the photovoltaic power station.

[0007] Preferably, each subsequence in the subsequence set is added to the data sample set as a feature.

[0008] Preferably, in step three, the first subsequences in subsequence set A and subsequence set B are respectively selected for signal decomposition.

[0009] Preferably, the signal processing methods in step 2 are singular spectrum analysis and fully adaptive noise ensemble empirical mode decomposition; the singular spectrum analysis method is as follows: Step 2-1-1. Set the window length L and the number of sliding windows K to construct the trajectory matrix corresponding to the historical power data ; Step 2-1-2. Trajectory matrix Perform eigenvalue decomposition and sort the eigenvectors according to the eigenvalues; Step 2-1-3. Calculate the trajectory matrix Projection on the eigenvector to obtain the time-evolving weight ; Step 2-1-4. Superimpose subsequences with the same signal and weight them according to the time evolution Obtain multiple subsequences processed by singular spectrum analysis, and construct a subsequence set A; The fully adaptive noise ensemble empirical mode decomposition method is as follows: Step 2-2-1. Add white noise to the historical power data to obtain the noise signal Step 2-2-2. Noise signal Perform empirical mode decomposition to obtain group I results, and average the results to obtain the eigenmode components and the first-order residual ; Step 2-2-3. First-order residual Add noise and perform empirical mode decomposition to obtain the eigenmode components and the second-order residual ; Step 2-2-4. Repeat the above steps until the decomposition of the residual is completed.

[0010] Preferably, in the step three, a variational mode decomposition method is used to perform signal decomposition on the first subsequence.

[0011] Preferably, in step 1, the meteorological data include columnar cloud water content, columnar cloud ice content, ground pressure, relative humidity, total cloud cover, longitude wind, latitudinal wind component, temperature, surface solar radiation, surface thermal radiation, net solar radiation, and total precipitation.

[0012] Preferably, in step 4, the photovoltaic power generation power prediction model adopts an LSTM-Informer model.

[0013] Preferably, in the step 1, before constructing the data sample set, outlier detection and data filling are performed on the historical power data and meteorological data respectively, and the data sample set is constructed with the filled data.

[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 multi-time-scale 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 multi-time-scale photovoltaic power generation prediction method.

[0016] The present invention has the following beneficial effects: 1. The present invention adopts a variety of modal decomposition methods, and uses singular spectrum analysis and variational modal decomposition to decompose the historical photovoltaic power data, which can produce different modal decomposition results, forming a more comprehensive description of the complex characteristics of the signal, thereby generating heterogeneous modal subsequences that can better reflect the inherent laws of complex signals, effectively avoiding the problem that a single decomposition method may miss some modal subsequences that are easy to alias, and solving the problem that the modal decomposition subsequence may have little information and insufficient feature information.

[0017] 2. In the subsequences generated by the first modal decomposition, the present invention further uses variational modal decomposition to process the high-frequency first subsequence with high irregularity. By integrating the subsequences obtained by one-step decomposition and two-step decomposition, the original photovoltaic power data and weather characteristics, it effectively extracts diverse and clear features, improves the quality of the decomposition results, and thus enhances the prediction performance.

[0018] 3. The present invention solves the problem of poor accuracy of Informer in short-term photovoltaic power prediction by combining the long short-term memory network and the photovoltaic power prediction model constructed by Informer, and meets the application of different scenarios such as real-time scheduling, economic load scheduling, system maintenance and long-term planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the present invention.

[0020] Figure 2 It is a structural diagram of the photovoltaic power generation prediction model in the present invention.

[0021] Figure 3 This is a subsequence diagram obtained by the present invention using the singular spectrum analysis method.

[0022] Figure 4 This is a subsequence diagram obtained by the fully adaptive noise set empirical mode decomposition method of the present invention.

[0023] Figure 5 This is a diagram of the architecture of the long short-term memory network model in the present invention.

[0024] Figure 6 This is the architecture diagram of the Informer model in the present invention.

[0025] Figure 7 The figure is a schematic diagram comparing the prediction effects of the present invention and the existing model.

[0026] Figure 8 Schematic diagram for comparing the error indicators of the present invention and the existing model; wherein, (a) is a schematic diagram for comparing the mean square error; (b) is a schematic diagram for comparing the mean absolute error; (c) is a schematic diagram for comparing the root mean square error; and (d) is a schematic diagram for comparing the coefficient of determination. DETAILED DESCRIPTION

[0027] The present invention will be further described below in conjunction with the accompanying drawings.

[0028] like Figure 1 and Figure 2 As shown, a multi-time scale photovoltaic power generation prediction method based on multiple modal decomposition includes the following steps: Step 1: Extract the historical power data and meteorological data of the photovoltaic power station. The meteorological data includes multiple meteorological features, namely column cloud water content, column cloud ice content, surface air pressure, relative humidity, total cloud cover, U-wind component at 10 m height (wind in the longitude direction), V-wind component at 10 m height (wind in the latitude direction), temperature at 2 m height, surface solar radiation, surface thermal radiation, net solar radiation at the top of the atmosphere, and total precipitation. Identify the abnormal and missing data in the historical power data and meteorological data, and use the data at the previous and next moments or mesoscale data to fill the abnormal and missing data. Construct a data sample set with the filled data.

[0029] Step 2: Use the historical power data as the original signal sequence ; Process the original signal sequence using singular spectrum analysis (SSA) and complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) respectively.

[0030] 2-1. Singular spectrum analysis In some embodiments, the specific process of singular spectrum analysis is as follows: 2-1-1. Set the window length L (L < N / 2); Let the number of sliding windows be K = N - L + 1, and construct the trajectory matrix as: where N is the length of the original signal sequence .

[0031] 2-1-2. Construct the covariance matrix according to the trajectory matrix ; Perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalues ( ) and the corresponding eigenvectors ; Arrange all the eigenvalues in descending order to form the eigenvalue set , and all the eigenvectors form the eigenvector set ; where ; L is the number of eigenvalues. Use the eigenvalue set as the singular spectrum of the time series . During the process of singular value decomposition, there is ; where is the basic matrix, ; is the right singular vector of the m-th eigenvalue, ; The trajectory matrix X is expressed as:

[0032] 2-1-3. Calculate the trajectory matrix In the feature vector Projection on to obtain the time evolution weight , whose expression is: in, is the lth element of the eigenvector corresponding to the mth eigenvalue; ; .

[0033] 2-1-4. Superimpose the subsequences with the same signal to obtain the reconstructed signal (modal component IMF) of the original sequence. The mth eigenvalue corresponds to the i-th element of the reconstructed signal. The expression is: Use singular spectrum analysis to analyze the original signal sequence The reconstructed signal obtained after processing is a set of subsequences , the reconstructed signal corresponding to the first eigenvalue as the first subsequence.

[0034] In other embodiments, other existing methods may be used to perform singular spectrum analysis.

[0035] 2-2. Fully Adaptive Noise Ensemble Empirical Mode Decomposition 2-2-1. To the original signal sequence Add white noise to form multiple different noise signal sets, add white noise I times in total, and obtain the noise signal for: in, The original signal sequence The original signal at each moment in ; is the noise amplitude; is white noise.

[0036] 2-2-2. For each noise signal Perform EMD decomposition to obtain group I results, and average the results to obtain the eigenmode components and the first-order residual : in, is the first-order modal component.

[0037] 2-2-3. First-order residual Continue to add noise and perform EMD decomposition to obtain the intrinsic mode components and the second-order residual : IM F 2 = 1 I ∑ i=1 I E 2 [re s 1 (t)+ ε 2 E 2 ( W i t )] in, represents EMD decomposition; is the amplitude of the added Gaussian white noise; is white noise.

[0038] 2-2-4. Repeat the above steps until the decomposition of the residual is completed, and the decomposition formula of the original signal is obtained: in, ; N is the number of EMD decompositions performed.

[0039] The original signal sequence is decomposed using fully adaptive noise ensemble empirical mode decomposition. The intrinsic mode components obtained after processing are taken as a subsequence set .

[0040] 2-3. The subsequence sets A and B after SSA and CEEMDAN processing are as follows Figure 3 and Figure 4 As shown, each subsequence in the subsequence set A and the subsequence set B is added to the data sample set as a new feature.

[0041] Step 3: Use variational mode decomposition (VMD) to decompose the first subsequence in the subsequence set A and the subsequence set B respectively.

[0042] In some embodiments, the specific process of variational mode decomposition is as follows: The original signal of the variational mode decomposition is taken as the first subsequence to be decomposed ; Use variational mode decomposition to transform the original signal Decompose and minimize the sum of the estimated bandwidth of each mode after VMD decomposition by minimizing the following objective function: in, is the partial derivative with respect to t; is the modal component after VMD decomposition; is the number of modal components after VMD decomposition; represents the center frequency of the kth modal component of the signal; represents Dirac distribution; j is the imaginary unit; .

[0043] To solve the above problem, we need to introduce the Lagrangian multiplier λ and the penalty factor α to construct the Lagrangian function : in, is the inner product of two functions; is the Lagrange multiplier with respect to time function.

[0044] Use the alternating direction method of multipliers (ADMM) to solve the components and center frequency, that is, update alternately in the following way , and λ: in, is the representation of the kth IMF in the frequency domain at the n+1th iteration; for Fourier transform of is the frequency.

[0045] in, is the center frequency of the kth IMF in the n+1th iteration.

[0046] in, is the representation of the Lagrange multiplier of the n+1th iteration in the frequency domain; τ is the update step size of the Lagrange multiplier λ.

[0047] For a given judgment accuracy e>0, the iteration stops when the following equation is satisfied: After the iteration is finished, a subsequence set C and a subsequence set D are obtained after VMD decomposition processes the first subsequences in the subsequence set A and the subsequence set B, respectively.

[0048] In other embodiments, other existing methods may be used to perform variational mode decomposition to obtain subsequence set C and subsequence set D.

[0049] Each subsequence in the subsequence set C and the subsequence set D is added to the data sample set as a new feature; after standardizing the data in the data sample set, the data sample set is divided into a training sample set, a verification sample set and a test sample set.

[0050] Step 4: Construct a photovoltaic power generation prediction model; the photovoltaic power generation prediction model includes a long short-term memory network (LSTM) and a residual attention transformer (Informer). The structure of LSTM is as follows Figure 7 As shown in the figure, the core components of LSTM are memory cells and three gates: input gate, forget gate, and output gate, which work together to determine the memory and forgetting of information. The principle of LSTM is as follows:

[0051] 1. The input gate is used to decide how much new information to update. The output of the input gate for: in, is the Sigmoid activation function; is the weight parameter; is the current input value; is the previous state value; is the bias parameter.

[0052] 2. The forget gate is used to control how much of the past memory element is retained. The output of the forget gate for: in, is the weight parameter; is the bias parameter.

[0053] 3. The output gate is used to determine how much information to output. The output of the output gate for: in, is the weight parameter; is the bias parameter.

[0054] 4. Memory is used to store historical information. The output of memory for: in, is the weight parameter; is the bias parameter; is the memory cell state at the current time step.

[0055] 5. Status Update Calculate the current state based on the input gate, forget gate and updated memory element.

[0056] in, is the hidden state at the current time step.

[0057] Based on Transformer, Informer significantly improves the efficiency and performance of the model by introducing sparse self-attention mechanism and self-attention distillation mechanism. In addition, Informer also adopts a one-time output long time step strategy to avoid the error accumulation problem in traditional autoregressive methods and further optimize the effect of time series prediction. The model architecture of Informer is shown in the figure below: Figure 8 Informer has the following features:

[0058] 1. Sparse Self-Attention Mechanism The core of the Transformer model is the self-attention mechanism, whose computational complexity is (L is the length of the sequence). For long time series, the amount of computation will increase with the square of the sequence length, resulting in extremely high consumption of memory and computing resources. To solve this problem, Informer proposed a sparse self-attention mechanism, which takes advantage of the sparse distribution of information in time series and reduces the computational complexity. Informer achieves sparsity through the following methods:

[0059] Informer assumes that most of the information in the time series is driven by a few key time steps, so it does not need to Specifically, Informer uses the sparsity of the point distribution between query keys to select the k keys with the highest attention scores at each query position. In this way, only these high-weighted attention scores need to be calculated, which greatly reduces the amount of calculation.

[0060] 2. Self-Attention Distillation Mechanism In long time series, the longer the sequence is, the more redundant information there is, which not only increases the complexity of the model, but may also lead to overfitting and decreased generalization ability of the model. To this end, Informer proposes a self-attention distillation mechanism to extract multi-scale time series features by compressing the sequence length layer by layer. The specific process is as follows:

[0061] (1) Sequence compression: In the output of each layer of Transformer, the sequence is downsampled through the pooling operation. For example, a sequence of length L is compressed into a subsequence of length L / 2 or shorter. This process can effectively remove redundant information while retaining the main patterns and trends of the time series.

[0062] (2) Inter-layer feature transfer: The compressed sequence of each layer is used as the input of the next layer, and higher-level time series features are extracted layer by layer. This layer-by-layer compression method ensures that the model can capture global information while significantly reducing computational overhead.

[0063] 3. Output long time steps at one time Traditional time series forecasting methods usually adopt a multi-step rolling forecast strategy, that is, based on the forecast results of the previous step, the forecast values ​​of the subsequent time steps are gradually generated. Although this method is intuitive, it will introduce cumulative errors: the forecast error of each step will be passed to the next step, resulting in a decrease in the accuracy of long-term forecasts. Informer generates the entire forecast sequence at once, rather than rolling forecasts step by step.

[0064] Step 6: Use the training sample set to train the photovoltaic power prediction model, and use the validation sample set to verify the model performance of the training process. Use the test sample set to test the trained photovoltaic power prediction model, and after the test results meet the requirements, perform photovoltaic power prediction on multiple time scales for the photovoltaic power station.

[0065] Step 7: In order to verify the effectiveness and accuracy of the present invention, the present invention (TSM-LSTM-Informer) and the existing models (Autoformer, Reformer, Transformer) are used to predict photovoltaic power generation. The photovoltaic power generation and meteorological data provided by the Global Energy Forecasting Competition 2014 (GEFCom2014) are used. The data from April 1, 2012 to July 1, 2014 in region 1 are selected for training. The data collection interval is 1 hour. The curve graph of the prediction result is shown in the figure below: Figure 6 shown.

[0066] Use MSE (mean square error), MAE (mean absolute error), RMSE (root mean square error) and R 2(Coefficient of Determination) is used as an evaluation index to evaluate the prediction results of different models. MAE is the mean absolute value of the difference between the predicted value and the true value, which measures the average deviation of the model prediction. MSE measures the prediction performance of the model by calculating the average of the squared errors between the predicted value and the true value. RMSE is the average square root of the difference between the predicted value and the true value. Compared with MAE, RMSE is more sensitive because it imposes a higher penalty on large errors. The smaller the values ​​of MAE, MSE and RMSE, the higher the prediction accuracy of the model. 2 is a statistic used to measure the degree of fit of the regression model to the observed data. 2 The closer the value is to 1, the better the model fits the data. These indicators can be expressed by the following formula:

[0067] in, are the actual value, predicted value and average value respectively; is the sample size.

[0068] The multi-step evaluation results of different models at 1h, 2h, 4h, 12h, 24h, 48h and 72h are shown in Table 1 and Figure 5 shown.

[0069] Table 1 Prediction results of different models

[0070] The Count in Table 1 represents the number of optimal results obtained by the present invention compared with other models. It can be seen from Table 1 that the prediction effect of the present invention at all time steps is better than that of other models, especially at short time steps (1h, 2h, 4h). Compared with the Reformer with the best comprehensive performance among the compared models, the MAE of the present invention at 1h was reduced by 58.26%, and the RMSE was reduced by 65.62%. At the same time, at short time steps (1h, 2h, 4h), the error growth rate of the present invention is also significantly lower than that of other models. Compared with 1h, the MAE of the present invention at 4h increased by 29.85%, and the R 2 The MAE of Reformer increased by 33.33%, while that of R 2The present invention has 4 indicators slightly inferior to Reformer at 72 hours, but the MAE and RMSE at all other time steps are significantly lower than those of other models, and the R² indicator is always maintained at a high level, especially at short time steps (1 hour, 2 hours, 4 hours).

[0071] After comparative analysis, it can be concluded that the present invention performs best in multi-step prediction tasks and has the following advantages: 1. High accuracy under short time steps, which can effectively extract short-term features of time series and significantly reduce prediction errors. 2. Stability under long time steps. Even under medium and long time steps (12h-72h), the error of the present invention still has good performance, indicating that it can effectively extract long-term features of time series. Therefore, the present invention can better predict future photovoltaic power generation.

Claims

1. A multi-time scale photovoltaic power generation prediction method based on multiple modal decompositions, characterized by: The following steps are involved: Step 1: Obtain and extract historical power data and meteorological data of photovoltaic power stations to construct a data sample set; Step 2: Use different signal processing methods to decompose the historical power data, obtain subsequence set A and subsequence set B, and add them to the data sample set as new features; Step 3: Perform signal decomposition on the two subsequence sets obtained in step 2 respectively to obtain subsequence set C and subsequence set D, and add them into the data sample set as new features; Step 4: Build a photovoltaic power generation prediction model, and use the data sample set obtained in step 3 to train the photovoltaic power generation prediction model; Step 5: Use the trained photovoltaic power prediction model to predict the power generation of the photovoltaic power station.

2. The multi-time-scale photovoltaic power generation prediction method based on multiple modal decompositions according to claim 1 is characterized in that: Each subsequence in the subsequence set is added to the data sample set as a feature.

3. The multi-time-scale photovoltaic power generation prediction method based on multiple modal decomposition according to claim 1 is characterized by: In the step three, the first subsequences in the subsequence set A and the subsequence set B are respectively selected for signal decomposition.

4. The multi-time-scale photovoltaic power generation prediction method based on multiple modal decompositions according to claim 1 is characterized in that: The signal processing methods in step 2 are singular spectrum analysis and fully adaptive noise ensemble empirical mode decomposition respectively; the singular spectrum analysis method is as follows: Step 2-1-1. Set the window length L and the number of sliding windows K to construct the trajectory matrix corresponding to the historical power data ; Step 2-1-2. Trajectory matrix Perform eigenvalue decomposition and sort the eigenvectors according to the eigenvalues; Step 2-1-3. Calculate the trajectory matrix Projection on the eigenvector to obtain the time-evolving weight ; Step 2-1-4. Superimpose subsequences with the same signal and weight them according to the time evolution Obtain multiple subsequences processed by singular spectrum analysis, and construct a subsequence set A; The fully adaptive noise ensemble empirical mode decomposition method is as follows: Step 2-2-1. Add white noise to the historical power data to obtain the noise signal Step 2-2-2. Noise signal Perform empirical mode decomposition to obtain group I results, and average the results to obtain the eigenmode components and the first-order residual ; Step 2-2-3. First-order residual Add noise and perform empirical mode decomposition to obtain the eigenmode components and the second-order residual ; Step 2-2-4. Repeat the above steps until the decomposition of the residual is completed.

5. The multi-time scale photovoltaic power generation prediction method based on multiple modal decomposition according to claim 1 is characterized by: In the step three, the variational mode decomposition method is used to perform signal decomposition on the first subsequence.

6. The multi-time-scale photovoltaic power generation prediction method based on multiple modal decompositions according to claim 1 is characterized in that: In the step 1, the meteorological data include columnar cloud water content, columnar cloud ice content, ground pressure, relative humidity, total cloud cover, longitude wind, latitudinal wind component, temperature, surface solar radiation, surface thermal radiation, net solar radiation, and total precipitation.

7. The multi-time scale photovoltaic power generation prediction method based on multiple modal decompositions according to claim 1 is characterized by: In the step 4, the photovoltaic power generation power prediction model adopts the LSTM-Informer model.

8. The multi-time-scale photovoltaic power generation prediction method based on multiple modal decompositions according to claim 1 is characterized by: In the step 1, before constructing the data sample set, outlier detection and data filling are performed on the historical power data and meteorological data respectively, and the data sample set is constructed with the filled data.

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 the multi-time-scale photovoltaic power prediction method based on multiple modal decomposition as described in any one of claims 1-8.

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 the multi-time-scale photovoltaic power prediction method based on multiple modal decomposition as described in any one of claims 1 to 8.

Images