Ultra-short-term photovoltaic power interval prediction method based on data optimization and deep learning
Through the method based on data optimization and deep learning, the photovoltaic power sequence is decomposed and the CNN-BiGRU model is constructed, which solves the accuracy and stability of the photovoltaic power prediction, and achieves more accurate and efficient photovoltaic power interval prediction to meet the scheduling needs of the power system. .
Patent Information
- Application Number
- CN202411291938.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-14
- Publication Date
- 2024-12-24
AI Technical Summary
When the existing photovoltaic power prediction model faces complex and changing weather conditions and system operating conditions, it is difficult to accurately predict the change in photovoltaic power generation power, and the interval prediction model has insufficient coverage and average width, which affects the prediction accuracy and stability.
Using ultra-short-term photovoltaic power interval prediction method based on data optimization and deep learning, the photovoltaic power sequence is decomposed through the ICEEMDAN algorithm, combined with the Pearson correlation coefficient to screen meteorological factors, construct a CNN-BiGRU model, and optimize hyperparameters using the improved sparrow search algorithm. Generate more accurate prediction intervals.
It significantly improves the accuracy and stability of photovoltaic power prediction, obtains a narrower prediction interval, meets the actual needs of power system scheduling, and improves prediction quality and efficiency.
Smart Images

Figure CN119180377A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of photovoltaic power prediction methods and relates to an ultra-short-term photovoltaic power interval prediction method based on data optimization and deep learning. Background Art
[0002] Photovoltaic power generation has strong randomness and uncertainty. Accurate power prediction is of great significance for the dispatch and operation of photovoltaic power generation in new power systems. At present, most studies focus on deterministic photovoltaic power prediction. However, in the actual operation of the power system, photovoltaic power is greatly affected by other factors and presents high uncertainty. It is difficult to provide sufficient photovoltaic information by relying solely on deterministic prediction. Interval prediction can provide the fluctuation range of future photovoltaic power, help power grid dispatchers evaluate uncertainty and risk factors, and make safer and more economical decisions. Therefore, accurate prediction of photovoltaic power interval is of great significance. In the research field of photovoltaic power interval prediction, many existing prediction models can generate prediction intervals, but often ignore the impact of the coverage and average width of the prediction interval on the model training process, which may not only affect the accuracy of the prediction model, but also reduce its stability in practical applications. In addition, in the face of complex and changeable weather conditions and system operation conditions, traditional interval quantization models are insufficient in adapting to the complexity of photovoltaic data and improving generalization performance. Especially in the non-stationary period of photovoltaic power generation, existing models often find it difficult to accurately predict its power changes. Therefore, further feature mining and deep learning of the data are needed to improve the accuracy of photovoltaic power prediction and meet the dispatch requirements in actual projects. Summary of the invention
[0003] The purpose of the present invention is to provide an ultra-short-term photovoltaic power interval prediction method based on data optimization and deep learning, which solves the problems of low accuracy in original data classification and poor prediction accuracy in the prior art.
[0004] The technical solution adopted by the present invention is an ultra-short-term photovoltaic power interval prediction method based on data optimization and deep learning, which is specifically implemented according to the following steps:
[0005] Step 1: preprocess the historical data of photovoltaic power generation;
[0006] Step 2: Use the ICEEMDAN algorithm to decompose the preprocessed photovoltaic power series, and obtain the trend component, periodic component and random component based on the obtained IMF component;
[0007] Step 3: Calculate the Pearson correlation coefficients between each meteorological factor and the trend component, periodic component and random component contained in the historical data, perform feature screening on each component according to the Pearson correlation coefficient, and construct an input data set;
[0008] Step 4: Use CNN, BiGRU and CQR to build an interval prediction model;
[0009] Step 5: Divide the component input data sets obtained in step 3 and input them into the interval prediction model. Use the ISSA algorithm to optimize the model hyperparameters and obtain the interval prediction results under the optimal hyperparameters.
[0010] The present invention is also characterized in that:
[0011] Step 1 is as follows:
[0012] The 3-sigma principle is used to identify outliers in the historical data of photovoltaic power generation, and the identified outliers are eliminated. Then the k-nearest neighbor algorithm is used to interpolate the missing values. Finally, the processed data is normalized to form a feature matrix.
[0013] Step 2 is as follows:
[0014] The ICEEMDAN algorithm is used to decompose the preprocessed photovoltaic power sequence to obtain several IMFs and a residual sequence. The complexity of each IMF is calculated using the permutation entropy, and the principal component analysis clustering algorithm is used to cluster the IMFs. The IMFs are divided into trend components, periodic components and random components.
[0015] The specific process of ICEEMDAN algorithm to decompose the photovoltaic power sequence is as follows: first, Gaussian white noise is adaptively added to the sequence in each iteration, then a set of IMFs is obtained through ICEEMDAN, and the average value of all decomposed IMFs of the same order is calculated to form the final IMFs. Finally, the obtained IMFs are subtracted from the original sequence to obtain the residual; this process is repeated until the entire sequence is decomposed, and finally several IMFs and a residual sequence are obtained.
[0016] In step 3, meteorological factors with a Pearson correlation coefficient ≥ 0.3 with each component are selected as input features.
[0017] In the interval prediction model in step 4, CNN is used to extract features of each component data, and BiGRU is used to extract data features with long-distance dependence from both the forward and backward directions, deeply explore the long-term time series relationship between photovoltaic power and its influencing meteorological factors, form the predicted value of photovoltaic power generation, and then use CQR to generate the prediction interval of each component.
[0018] Step 5 is as follows:
[0019] The input data sets of each component obtained in step 3 are divided into training set, validation set and test set in the ratio of 6:2:2. Then the sliding window method is used to input the data into the interval prediction model constructed in step 4. The ISSA is used to optimize the model hyperparameters. The interval prediction results of trend component, periodic component and random component are obtained under the optimal hyperparameters. Finally, the three interval prediction results are superimposed by weighted summation to obtain the final interval prediction result.
[0020] The ISSA algorithm in step 5 is an improvement on the SSA algorithm. The specific improvements include: using the reverse learning strategy to initialize the sparrow population; using the step factor dynamic adjustment strategy to optimize the step control parameters β and k in the SSA early warning detection sparrow position update formula; introducing Levy flight to update the scout position.
[0021] The beneficial effects of the present invention are:
[0022] The method of the present invention adopts the 3-sigma principle and the k-nearest neighbor algorithm to preprocess the abnormal values and missing values in the historical data respectively, which effectively improves the integrity and consistency of the data and provides high-quality data input for the subsequent prediction model; the ICEEMDAN algorithm is used to decompose the preprocessed photovoltaic power sequence, which reduces the non-stationarity of the photovoltaic power sequence, and considering that the photovoltaic power has strong volatility, a large number of IMF components will be generated after decomposition, which will seriously affect the calculation speed and accuracy. Therefore, the permutation entropy is used to measure the complexity of each IMF component, and the PCA clustering algorithm is used to divide them into trend components, periodic components and random components, which can obtain a more reasonable partition data set and capture the inherent laws of photovoltaic power changes in more detail, thereby significantly improving the prediction accuracy; the Pearson correlation coefficient is used to screen the features of each component, and the gas components with high correlation with each component are extracted. Using image variables as input features can build more representative and targeted input data sets, thereby improving the prediction performance of the model; using the improved sparrow search algorithm to optimize the neural network hyperparameters, the improved algorithm solves the problems of insufficient population diversity, easy to fall into local optimality and low solution accuracy in the original sparrow algorithm, significantly improving the optimization efficiency and accuracy of the algorithm; using CNN-BiGRU as the basic model, feature extraction and time series analysis are performed to predict and analyze photovoltaic data, and the long-term time series relationship between photovoltaic power and its influencing factors is deeply explored, further improving the prediction ability of the model, and using the CQR algorithm to convert the previous CNN-BiGRU prediction results into prediction intervals, quantifying the uncertainty of the prediction. The final interval prediction result can obtain a narrower prediction interval while meeting the preset confidence interval, thereby improving the accuracy of the prediction model.
[0023] The method of the present invention has fast prediction speed and high prediction accuracy, effectively improves the prediction quality of ultra-short-term photovoltaic power interval, and has important practical significance for the dispatching and operation of the entire power system. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 is a flow chart of the method of the present invention;
[0025] Figure 2 It is a flow chart of obtaining prediction results by using the interval prediction model in the method of the present invention;
[0026] Figure 3 is a decomposition result diagram obtained by using ICEEMDAN to decompose the photovoltaic power sequence in Example 1 of the present invention;
[0027] Figure 4 is a clustering result diagram of the permutation entropy in Example 1 of the present invention;
[0028] Figure 5 is a schematic diagram of components obtained by decomposition and reconstruction in Example 1 of the present invention;
[0029] Figure 6 is a schematic diagram of the relationship between the global horizontal radiation meteorological factors and the photovoltaic power sequence obtained by the feature screening method in Example 1 of the present invention;
[0030] Figure 7 Schematic diagram of the relationship between the total tilted radiation meteorological factor and the photovoltaic power sequence obtained by the feature screening method in Example 1 of the present invention;
[0031] Figure 8 Schematic diagram of the relationship between the diffuse horizontal radiation meteorological factor and the photovoltaic power sequence obtained by the feature screening method in Example 1 of the present invention;
[0032] Fig. 9 Schematic diagram of the relationship between the diffuse oblique radiation meteorological factor and the photovoltaic power sequence obtained by the feature screening method in Example 1 of the present invention;
[0033] Fig.10 Schematic diagram of the relationship between weather temperature meteorological factors and photovoltaic power sequence obtained by the feature screening method in Example 1 of the present invention;
[0034] Fig.11 is a graph of interval prediction results of each component obtained in Example 1 of the present invention;
[0035] Fig.12 It is a comparison chart of the interval prediction results of various models in Example 2 of the present invention. DETAILED DESCRIPTION
[0036] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0037] The present invention is based on the ultra-short-term photovoltaic power interval prediction method based on data optimization and deep learning, such as Figure 1 As shown, the specific implementation steps are as follows:
[0038] Step 1: Preprocess the historical data of photovoltaic power generation:
[0039] The 3-sigma principle is used to identify outliers in photovoltaic power data and remove the identified outliers. Then the k-nearest neighbor algorithm is used to interpolate the missing values in the historical data. Finally, the processed data is normalized to form a feature matrix.
[0040] Step 2: Use the “decomposition-reconstruction” framework to optimize the photovoltaic power sequence processed in step 1, and decompose the photovoltaic power sequence into three different intrinsic mode functions (IMF): trend component, periodic component and random component:
[0041] The improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) algorithm is used to decompose the preprocessed photovoltaic power sequence. The specific process is as follows: First, Gaussian white noise is adaptively added to the sequence in each iteration, and then a set of IMFs is obtained through ICEEMDAN. The average value of all decomposed IMFs of the same order is calculated to form the final IMFs; finally, the obtained IMFs are subtracted from the original sequence to obtain the residual. This process is repeated until the entire sequence is decomposed, and finally several IMFs and a residual sequence are obtained. The complexity of each IMF is calculated using the permutation entropy (PE), and the principal components analysis (PCA) clustering algorithm is used to cluster each IMF, and the IMF is divided into trend component, periodic component and random component to complete the data optimization and feature extraction process.
[0042] The improved ICEEMDAN adaptively adds Gaussian white noise to the decomposition and iteration process of the original CEEMDAN algorithm, which can suppress the problem of "modal aliasing" and decomposition of too many small-amplitude intrinsic mode functions (IMFs) in EMD. Compared with CEEMDAN, ICEEMDAN uses the IMF after the white noise is decomposed to determine each IMF component after the original sequence is decomposed, which can further reduce the problems of residual noise and modal aliasing in IMF.
[0043] Step 3: Calculate the Pearson correlation coefficient between each meteorological factor contained in the historical data and the trend component, periodic component and random component, perform feature screening on each component based on the Pearson correlation coefficient, select the meteorological factors that have a greater impact on each component as input features, and construct the corresponding input data set.
[0044] The Pearson correlation coefficient is used to measure the direct linear correlation between two variables (such as x and y). When the Pearson coefficient is in [0,1], it is a positive correlation. The closer the value is to 1, the stronger the correlation is. When the Pearson coefficient is in [-1,0], it is a negative correlation or no correlation. The formula is:
[0045]
[0046] Where cov is the variance of x and y, and σ is the standard deviation.
[0047] The size of the Pearson correlation coefficient is defined as follows: 1 is perfect correlation; 0.8 to 1 is very strong correlation; 0.4 to 0.8 is strong correlation; 0.3 to 0.4 is moderate correlation; 0 to 0.3 is weak correlation; 0 is completely uncorrelated. In addition, in step 3, meteorological factors with a Pearson correlation coefficient of ≥ 0.3 between each component are selected as input features.
[0048] Step 4: Use convolutional neural network (CNN), bidirectional gate recurrent unit neural network (BiGRU) and conformalized quantile regression (CQR) to build an interval prediction model.
[0049] The interval prediction model used in the method of the present invention is CQP-CNN-BiGRU, wherein CNN is used to extract features of each component data, and BiGRU simultaneously extracts data features with long-distance dependence from both the forward and backward directions, deeply explores the long-term time series relationship between photovoltaic power and its influencing meteorological factors, forms a predicted value of photovoltaic power generation, and then uses CQR to generate prediction intervals for each component.
[0050] Step 5: Figure 2As shown in FIG. 1 , the input data sets of each component obtained in step 3 are divided into training set, validation set and test set in a ratio of 6:2:2. This division ratio is to ensure that the model has sufficient data for learning during the training phase (60% as the training set), while retaining sufficient data for performance verification and testing of the model (20% as the validation set and 20% as the test set). Then, the sliding window method is used to input the data into the interval prediction model constructed in step 4, and the improved sparrow search algorithm (ISSA) is used to optimize the model hyperparameters. Under the optimal hyperparameters, the interval prediction results of the trend component, the periodic component and the random component are obtained. Finally, the three interval prediction results are superimposed by weighted summation to obtain the final interval prediction result.
[0051] Among them, in order to solve the problems of insufficient population diversity, easy to fall into local optimality and low solution accuracy in the Sparrow Search Algorithm (SSA), some improvements were made to the Sparrow Search Algorithm. The improved algorithm flow is as follows: Figure 2 As shown, including:
[0052] (1) The reverse learning strategy is used to initialize the sparrow population. This strategy compares the initial population with its reverse population and selects the best population as the next generation population. This method can select individuals that are closer to the optimal position as the initial individuals, thereby effectively improving the convergence speed of all individuals in the population and improving the optimization efficiency of the algorithm.
[0053] (2) The step-size factor dynamic adjustment strategy is used to optimize the step-size control parameters β and k in the position update formula of the early warning detection sparrow in SSA, improve the global optimization ability, avoid the local optimal situation, and improve the solution accuracy of the algorithm;
[0054] (3) In the optimization process of SSA, Levy flight is introduced to update the scout position. Levy flight is a non-Gaussian random gait with a step length that follows a heavy-tailed probability distribution. In the optimization calculation, it can take into account both local search and global search by controlling the step length, thus solving the problem that SSA may produce local optimality.
[0055] Example 1
[0056] Step 1: Select the data of Alice Springs PV power station from the public data set of the Australian Desert Knowledge Solar Center (DKASC) as the research object. The data set includes PV power station data for one year from September 21, 2015 to September 21, 2016. The data sampling period is 5 minutes, and the weather factors collected include: global horizontal radiation, total oblique radiation, diffuse horizontal radiation, diffuse oblique radiation, wind speed, weather temperature (degrees Celsius), received effective energy, humidity, wind direction and rainfall, a total of 9 influencing factors.
[0057] The 3-sigma principle is used to detect abnormal data of photovoltaic power generation, and the mean μ and standard deviation σ of photovoltaic power generation are obtained. If the data exceeds the interval (μ-3σ, μ+3σ), it is determined to be abnormal data and eliminated. Then the k-nearest neighbor algorithm is used to interpolate the missing values in the historical data. The interpolation is mainly based on the k photovoltaic samples closest to the missing values. The preset distance function is used to obtain the distance between these k samples and the missing value, and the estimated value of the missing value is obtained by weighted average. Finally, the processed data is normalized to form a feature matrix.
[0058] Step 2: Use the ICEEMDAN algorithm to decompose the preprocessed photovoltaic power sequence. The decomposition results are as follows: Figure 3 As shown; the complexity of each IMF is calculated using permutation entropy to evaluate the orderliness and complexity of its time series. The calculation results of permutation entropy are shown in Figure 4 As shown; then the permutation entropy value is input as a feature into the PCA clustering algorithm for feature dimensionality reduction to extract the main change characteristics of IMF. Finally, IMFs are divided into trend component, periodic component and random component to complete the data optimization and feature extraction process. The schematic diagram of the three components is shown in Figure 5 shown.
[0059] Step 3: Calculate the Pearson correlation coefficient between each meteorological factor and each component, and select the meteorological factors that have a greater impact on each component, namely: global horizontal radiation, total tilted radiation, diffuse horizontal radiation, diffuse tilted radiation and weather temperature. The relationship between the five meteorological factors and the photovoltaic power sequence is as follows: Figure 6-Figure 10 As shown, there is a strong correlation between global horizontal radiation, total tilted radiation and photovoltaic power generation (Pearson correlation coefficient is 0.8 to 1), a strong correlation between diffuse horizontal radiation, diffuse tilted radiation and photovoltaic power generation (Pearson correlation coefficient is 0.4 to 0.8), and a moderate correlation between weather temperature and photovoltaic power generation (Pearson correlation coefficient is 0.35). The above five meteorological factors are used as input features to construct the input data set.
[0060] Step 4: Build an interval prediction model.
[0061] Step 5: Input the data set in step 3 into the model constructed in step 4 to obtain the interval prediction results of trend component, cycle component and random component. The results are as follows: Fig.11 shown.
[0062] Example 2
[0063] In order to verify the superiority of the model of the present invention in interval prediction, the same data set is used to perform photovoltaic power interval prediction on QR-GRU, QR-BiGRU, CQR-CNN-BiGRU, ICEEMDAN-CNN-BiGRU models and the method of the present invention (CQR-CNN-BiGRU after data optimization) under the preset 90% optimal confidence interval. The above models all use network parameters optimized by the ISSA algorithm, and set the comprehensive index coverage width criterion (CWC) as the objective function. The specific formula is as follows:
[0064]
[0065] Wherein, η is a manually set parameter, which is mainly used to coordinate the contribution of PICP and PINAW; α is a quantile parameter, which generates a quantile value [α / 2,1-(α / 2)] by default; in this embodiment, it is set as follows: η=10, α=0.1, which generates a 90% confidence interval. It combines the considerations of the evaluation indicators prediction intervals coverage probability (PICP) and prediction interval normalized average width (PINAW), transforming the multi-objective problem into a single-objective problem, thereby ensuring the optimization of the model training process.
[0066] The prediction results of each model on the test set are as follows Fig.12 As shown, it can be seen that the CQR-CNN-BiGRU neural network model after data optimization in the present invention has significantly better accuracy in photovoltaic power prediction than other traditional models. The specific data comparison is shown in Table 1 below:
[0067] Table 1 Comparison of prediction results of various prediction models
[0068]
[0069]
[0070] As can be seen from Table 1, all models can meet the requirements of the confidence interval, and the coverage PICP is greater than the preset 90%. Compared with the QR-GRU, QR-BiGRU, CQR-CNN-BiGRU, ICEEMDAN-CNN-BiGRU and other models, the prediction results obtained by the CQP-CNN-BiGRU model after data optimization of the present invention have a smaller normalized average width PINAW value, optimal confidence interval deviation rate and higher CWC value than other models under the condition of meeting the PICP requirements, which verifies that the method of the present invention can obtain a higher quality prediction interval.
[0071] Example 3
[0072] The ultra-short-term photovoltaic power interval prediction method based on data optimization and deep learning is implemented in the following steps:
[0073] Step 1: preprocess the historical data of photovoltaic power generation;
[0074] Step 2: Use the ICEEMDAN algorithm to decompose the preprocessed photovoltaic power series, and obtain the trend component, periodic component and random component based on the obtained IMF component;
[0075] Step 3: Calculate the Pearson correlation coefficients between each meteorological factor and the trend component, periodic component and random component contained in the historical data, perform feature screening on each component according to the Pearson correlation coefficient, and construct an input data set;
[0076] Step 4: Use CNN, BiGRU and CQR to build an interval prediction model;
[0077] Step 5: Divide the component input data sets obtained in step 3 and input them into the interval prediction model. Use the ISSA algorithm to optimize the model hyperparameters and obtain the interval prediction results under the optimal hyperparameters.
Claims
1. An ultra-short-term photovoltaic power interval prediction method based on data optimization and deep learning, characterized in that: Follow the steps below to implement it: Step 1: preprocess the historical data of photovoltaic power generation; Step 2: Use the ICEEMDAN algorithm to decompose the preprocessed photovoltaic power series, and obtain the trend component, periodic component and random component based on the obtained IMF component; Step 3: Calculate the Pearson correlation coefficients between each meteorological factor and the trend component, periodic component and random component contained in the historical data, perform feature screening on each component according to the Pearson correlation coefficient, and construct an input data set; Step 4: Use CNN, BiGRU and CQR to build an interval prediction model; Step 5: Divide the component input data sets obtained in step 3 and input them into the interval prediction model. Use the ISSA algorithm to optimize the model hyperparameters and obtain the interval prediction results under the optimal hyperparameters.
2. The ultra-short-term photovoltaic power interval prediction method based on data optimization and deep learning according to claim 1 is characterized in that: Step 1 is as follows: The 3-sigma principle is used to identify outliers in the historical data of photovoltaic power generation, and the identified outliers are eliminated. Then the k-nearest neighbor algorithm is used to interpolate the missing values. Finally, the processed data is normalized to form a feature matrix.
3. The ultra-short-term photovoltaic power interval prediction method based on data optimization and deep learning according to claim 1 is characterized in that: Step 2 is as follows: The ICEEMDAN algorithm is used to decompose the preprocessed photovoltaic power sequence to obtain several IMFs and a residual sequence. The complexity of each IMF is calculated using the permutation entropy, and the principal component analysis clustering algorithm is used to cluster the IMFs. The IMFs are divided into trend components, periodic components and random components.
4. The ultra-short-term photovoltaic power interval prediction method based on data optimization and deep learning according to claim 3 is characterized in that: The specific process of ICEEMDAN algorithm to decompose the photovoltaic power sequence is as follows: first, Gaussian white noise is adaptively added to the sequence in each iteration, then a set of IMFs is obtained through ICEEMDAN, and the average value of all decomposed IMFs of the same order is calculated to form the final IMFs. Finally, the obtained IMFs are subtracted from the original sequence to obtain the residual; this process is repeated until the entire sequence is decomposed, and finally several IMFs and a residual sequence are obtained.
5. The ultra-short-term photovoltaic power interval prediction method based on data optimization and deep learning according to claim 1 is characterized in that: In step 3, meteorological factors with a Pearson correlation coefficient ≥ 0.3 with each component are selected as input features.
6. The ultra-short-term photovoltaic power interval prediction method based on data optimization and deep learning according to claim 1 is characterized in that: In the interval prediction model in step 4, CNN is used to extract features of each component data, and BiGRU is used to extract data features with long-distance dependence from both the forward and backward directions, deeply explore the long-term time series relationship between photovoltaic power and its influencing meteorological factors, form the predicted value of photovoltaic power generation, and then use CQR to generate the prediction interval of each component.
7. The ultra-short-term photovoltaic power interval prediction method based on data optimization and deep learning according to claim 1 is characterized in that: Step 5 is as follows: The input data sets of each component obtained in step 3 are divided into training set, validation set and test set in the ratio of 6:2:
2. Then the sliding window method is used to input the data into the interval prediction model constructed in step 4. The ISSA is used to optimize the model hyperparameters. The interval prediction results of trend component, periodic component and random component are obtained under the optimal hyperparameters. Finally, the three interval prediction results are superimposed by weighted summation to obtain the final interval prediction result.
8. The ultra-short-term photovoltaic power interval prediction method based on data optimization and deep learning according to claim 1 or 7, characterized in that: The ISSA algorithm in step 5 is an improvement on the SSA algorithm. The specific improvements include: using the reverse learning strategy to initialize the sparrow population; using the step factor dynamic adjustment strategy to optimize the step control parameters β and k in the SSA early warning detection sparrow position update formula; introducing Levy flight to update the scout position.
Citation Information
Cited By
Distributed photovoltaic power prediction method and system based on machine learning
CN119853029A
Short-term photovoltaic power prediction method for optimizing bidirectional gating circulation unit based on improved crown porcupine algorithm
CN120090163A
