Deep Learning-Based Ultra-Short-Term Photovoltaic Power Prediction Method
By preprocessing and clustering historical photovoltaic power generation data, and combining variational mode decomposition and parallel learning of CNN-GRU neural networks, the problem of insufficient accuracy in photovoltaic power generation prediction was solved, and high-precision photovoltaic power prediction was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIAN UNIV OF TECH
- Filing Date
- 2022-12-16
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies have poor accuracy in predicting photovoltaic power generation, especially under different weather conditions. Furthermore, traditional models are not effective at separating noise and nonlinear non-stationary signals.
By employing a deep learning-based approach, historical photovoltaic power generation data is preprocessed and clustered. The parameters are optimized using a variational mode decomposition algorithm, and parallel learning of CNN and GRU neural networks is combined to achieve high-precision prediction of photovoltaic power.
It improves the accuracy of ultra-short-term photovoltaic power prediction, especially under different weather conditions such as sunny, cloudy and rainy days, significantly improving prediction accuracy and outperforming traditional models.
Smart Images
Figure CN115879629B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of photovoltaic power technology and relates to a method for ultra-short-term photovoltaic power prediction based on deep learning. Background Technology
[0002] Solar energy and weather variables are random and unstable, resulting in high uncertainty in photovoltaic (PV) power generation. Accurate PV power prediction is crucial for the short-term dispatch and operation of power generation plans in new power systems. In recent years, the most widely used method is to first classify the raw data using clustering algorithms, and then use deep learning to predict PV power. However, single clustering methods are greatly affected by parameters and cannot effectively segment various datasets. Furthermore, PV power data resembles a noisy, non-stationary signal sequence; if noise and signal cannot be effectively separated, the prediction results will have significant errors. Additionally, due to the large differences within the cloudy / rainy day dataset, existing models have failed to predict it well. Therefore, further feature mining and deep learning are needed to improve the accuracy of PV power prediction and meet the dispatch requirements in practical engineering. Summary of the Invention
[0003] The purpose of this invention is to provide a deep learning-based method for predicting ultra-short-term photovoltaic power, which solves the problem of poor prediction accuracy in existing technologies.
[0004] The technical solution adopted in this invention is a deep learning-based ultra-short-term photovoltaic power prediction method, which is implemented according to the following steps:
[0005] Step 1: Preprocess the historical data of photovoltaic power generation to obtain preprocessed data;
[0006] Step 2: Cluster the preprocessed data into three different weather types: sunny, rainy, and cloudy.
[0007] Step 3: Use optimization algorithms to optimize the parameters in the variational mode decomposition algorithm to obtain the parameter-optimized signal decomposition model;
[0008] Step 4: Use the optimized signal decomposition model from Step 3 to perform signal decomposition on the photovoltaic power generation sequence in the dataset to obtain the sequence of each component of the photovoltaic power.
[0009] Step 5: Normalize the data and use it as an input variable for photovoltaic power generation prediction;
[0010] Step 6: Use parallel deep learning with CNN and GRU networks to predict photovoltaic power generation, and finally obtain the predicted value of total photovoltaic power generation.
[0011] The invention is further characterized in that,
[0012] Step 1 specifically involves using the 3-sigma principle to detect abnormal data in photovoltaic power generation, removing missing and abnormal data; using Pearson coefficient analysis to calculate the correlation coefficients between various weather factors and photovoltaic power in historical data, removing feature vectors with correlation coefficients of 1 and negative values, using the remaining feature vectors as the main meteorological factors affecting photovoltaic output, and normalizing the remaining data to obtain preprocessed data.
[0013] Step 2 specifically involves:
[0014] Step 2.1: Cluster the preprocessed data using Gaussian mixture model and hierarchical clustering respectively. Since the label assignment rules of these two different clustering methods are not consistent when clustering the preprocessed data, the label results of the two clustering methods are rearranged using the Hungarian algorithm to convert the labels of the same type of data into consistent labels.
[0015] Step 2.2: Test the clustering accuracy of the two clustering results, extract the consensus points and non-consensus points of the two clustering methods, use the consensus points as the training set and the non-consensus points as the test set, calculate the Euclidean distance between the test set and each sample point in the training set, sort all the distance values, select the k samples with the smallest distance as "voters", predict the classification of the sample to be tested based on the "voters", implement the K-nearest neighbor algorithm to predict the non-consensus points, obtain the clustering results of the non-consensus points, and the set of the prediction results and the consensus points is the final clustering result;
[0016] Step 2.3: Evaluate the three final clustering results based on the evaluation indicators of the Kalinsky-Harabass index, the Davidson-Bolding index, and the silhouette coefficient.
[0017] Step 3 specifically involves:
[0018] Step 3.1: Set the parameters and fitness function in the particle swarm optimization algorithm;
[0019] Step 3.2: Initialize particle positions and velocities, setting the penalty factor and mode number parameters from variational mode decomposition to the particle positions; input all initialized particle positions into the variational mode decomposition algorithm, and calculate the component entropy values of all intrinsic mode components for each particle position; compare the magnitudes of local minimum entropies, and update individual local minimums and the overall global minimum; update particle velocities and positions; optimize the penalty factor and mode number parameters in the variational mode decomposition algorithm using the particle swarm optimization algorithm, determine whether the maximum number of iterations is satisfied, calculate the minimum network envelope entropy of global particles and the component envelope entropy of each intrinsic mode component, thereby obtaining the optimal penalty factor and mode number.
[0020] Step 4 specifically involves setting the initial values for the variational mode decomposition algorithm based on the optimal penalty factor and mode number obtained from the particle swarm optimization algorithm in step 3, decomposing and calculating the photovoltaic power for the three weather types to obtain the time-domain graph of the decomposed signal; and then using the fast Fourier transform to obtain the frequency-domain graph.
[0021] Step 5 specifically involves processing the decomposition results obtained in Step 4 by normalizing and standardizing each intrinsic modal component; dividing the processed dataset into a training set (first 80%) and a test set (last 20%); extracting the weather factor sequences processed in Step 1 to form a meteorological feature matrix, which, along with the photovoltaic power component sequences processed in this step, serves as the input variables for the prediction model.
[0022] Step 6 specifically involves:
[0023] Step 6.1: Use a CNN neural network to identify and output the local features of the meteorological feature matrix in Step 5;
[0024] The GRU neural network is used to identify the temporal characteristics of each component sequence of photovoltaic power in step 5, which is then used as the input for the next stage.
[0025] Step 6.2: After the data has undergone parallel learning by the two networks in Step 6.1, the output values are concatenated into a high-dimensional vector and input into the CNN neural network model to perform deeper learning and training on the network under multiple feature inputs, so as to obtain the final prediction result.
[0026] The beneficial effects of this invention are:
[0027] This invention presents a deep learning-based method for ultra-short-term photovoltaic (PV) power prediction, effectively improving the accuracy of PV power prediction. First, historical data is preprocessed, and feature selection using Pearson correlation coefficient analysis eliminates irrelevant variables, reducing computation and accelerating model training. A clustering approach is employed, combining the advantages of three clustering algorithms to obtain a more reasonable dataset partitioning, thus improving prediction accuracy. Variational mode decomposition (VMD) algorithms require pre-setting two parameters: a penalty factor and the number of modes. However, accurately setting these values is often difficult in practice. Therefore, particle swarm optimization (PSO) is used to optimize the VMD parameters, avoiding errors caused by human experience and improving overall prediction accuracy. Parallel learning of CNN and GRU networks is used to identify local and temporal features of the data, enabling the network structure to fully mine the input data. Finally, a deeper learning process is achieved through CNN neural network concatenation and fusion, resulting in high-precision prediction. The PV power prediction model proposed in this invention demonstrates excellent performance, significantly outperforming other traditional models in predicting PV power under different weather conditions. Attached Figure Description
[0028] Figure 1 This is a flowchart of the ultra-short-term photovoltaic power prediction method based on deep learning according to the present invention;
[0029] Figure 2 This is a clustering result diagram of the clustering set used in the ultra-short-term photovoltaic power prediction method based on deep learning in this invention;
[0030] Figure 3 This is a flowchart illustrating the optimization of the VMD algorithm using particle swarm optimization in the deep learning-based ultra-short-term photovoltaic power prediction method of this invention.
[0031] Figure 4 shows the time-domain graph obtained by decomposing the clear day dataset using the PSO-VMD model in the ultra-short-term photovoltaic power prediction method based on deep learning in this invention. Figure 4(a) is the time domain graph and Figure 4(b) is the frequency domain graph.
[0032] Figure 5 is a time-frequency spectrum obtained by decomposing the multi-cloud dataset using the PSO-VMD model in the ultra-short-term photovoltaic power prediction method based on deep learning in this invention, where Figure 5(a) is the time domain graph and Figure 5(b) is the frequency domain graph.
[0033] Figure 6 is a time-domain graph obtained by decomposing the rainy day dataset using the PSO-VMD model in the ultra-short-term photovoltaic power prediction method based on deep learning in this invention, where Figure 6(a) is the time domain graph and Figure 6(b) is the frequency domain graph.
[0034] Figure 7 This is a diagram showing the clear-sky prediction results of the ultra-short-term photovoltaic power prediction method based on deep learning in this invention, using the CNN-GRU hybrid neural network model.
[0035] Figure 8 This is a comparison chart of the sunny day prediction results of various models in the ultra-short-term photovoltaic power prediction method based on deep learning of this invention;
[0036] Figure 9 This is a cloud prediction result diagram using the CNN-GRU hybrid neural network model in the ultra-short-term photovoltaic power prediction method based on deep learning in this invention;
[0037] Figure 10 This is a comparison chart of the multi-cloud prediction results of various models in the ultra-short-term photovoltaic power prediction method based on deep learning in this invention;
[0038] Figure 11 This is a rainy day prediction result diagram using the CNN-GRU hybrid neural network model in the ultra-short-term photovoltaic power prediction method based on deep learning in this invention;
[0039] Figure 12This is a comparison chart of the rainy day prediction results of various models in the ultra-short-term photovoltaic power prediction method based on deep learning of this invention. Detailed Implementation
[0040] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.
[0041] Example
[0042] This embodiment provides a deep learning-based method for ultra-short-term photovoltaic power prediction, such as... Figure 1 As shown, please follow these steps:
[0043] Step 1: Preprocess the historical data of photovoltaic power generation to obtain preprocessed data;
[0044] Data from a photovoltaic power plant owned by Kaneka Corporation was selected for predictive research. Photovoltaic data from May 2015 to October 2016 were used. The original data included multiple feature sequences such as solar irradiance, wind speed, and global irradiance. The data time step was 5 minutes, and data was sampled 24 hours a day. A total of 135,647 sets of data were collected from the photovoltaic power plant. Due to the periodicity and randomness of photovoltaic output, only the data during daytime power generation was selected, and nighttime data was removed as the original dataset. The original dataset contained multiple weather factors, such as power generation, wind speed, wind direction, weather temperature in Celsius, global irradiance, and rainfall, totaling 14 items.
[0045] The 3-sigma principle was used to detect abnormal data in photovoltaic power generation. First, the mean μ and standard deviation σ of photovoltaic power generation were obtained. If the data exceeded the range of (μ-3σ, μ+3σ), it was judged as abnormal data. Missing and abnormal data were removed, and 67,680 sets of valid data were retained. The correlation coefficients between various weather factors and photovoltaic power in the historical data were calculated using Pearson coefficient analysis. Eigenvectors with correlation coefficients of 1 and negative values were removed. The remaining eigenvectors were taken as the main meteorological factors affecting photovoltaic output. The remaining data were then normalized to obtain preprocessed data.
[0046] Step 2: Cluster the preprocessed data into three different weather types: sunny, rainy, and cloudy.
[0047] Step 2.1: Cluster the preprocessed data using Gaussian mixture model and hierarchical clustering respectively. Since the label assignment rules of these two different clustering methods are not consistent when clustering the preprocessed data, the label results of the two clustering methods are rearranged using the Hungarian algorithm (Kuhn-Munkres) to convert the labels of the same type of data into consistent labels.
[0048] Step 2.2: Test the clustering accuracy (ACC) of the two clustering results. Extract the consensus points and non-consensus points of the two clustering methods. Use the consensus points as the training set and the non-consensus points as the test set. Calculate the Euclidean distance between each sample point in the test set and the training set, and sort all the distance values. Select the k samples with the smallest distances as "voters". Predict the classification of the test sample based on the "voters". Implement the K-nearest neighbor algorithm to predict the non-consensus points, and obtain the clustering results of the non-consensus points. The set of the prediction results and the consensus points is the final clustering result. Figure 2 As shown;
[0049] Step 2.3: Evaluate the three final clustering results based on the evaluation indicators of the Calinski-Harabaz index (CH), the Davies-Bouldin index (DBI), and the silhouette coefficient.
[0050] Step 3, as follows Figure 3 As shown, the parameters in the variational mode decomposition (VMD) algorithm are optimized using an optimization algorithm to obtain the parameter-optimized signal decomposition model;
[0051] Step 3.1: Set the parameters (including inertia weight, learning factor, number of iterations, and number of particles) and fitness function in the Particle Swarm Optimization (PSO) algorithm;
[0052] Step 3.2: Initialize particle positions and velocities, setting the penalty factor and mode number parameters from Variational Mode Decomposition (VMD) to the particle positions; input all initialized particle positions into the VMD algorithm, and calculate the component entropy values of all Intrinsic Mode Functions (IMFs) for each particle position; compare the magnitudes of local minimum entropies, and update individual local minimums and the global minimum; update particle velocities and positions; optimize the penalty factor and mode number parameters in the VMD algorithm using the particle swarm optimization (PSO) algorithm, determine whether the maximum number of iterations is satisfied, calculate the minimum network envelope entropy of global particles and the envelope entropy of each IMF component, thereby obtaining the optimal penalty factor and mode number.
[0053] Step 4: Use the optimized signal decomposition model from Step 3 to perform signal decomposition on the photovoltaic power generation sequence in the dataset to obtain the sequence of each component of the photovoltaic power.
[0054] Based on the optimal penalty factor and mode number obtained by the particle swarm optimization algorithm in step 3, the initial value of the VMD algorithm is set, and the photovoltaic power of the three weather types is decomposed and calculated to obtain the decomposed signal time domain diagram; then, the fast Fourier transform is used to obtain the frequency domain diagram. The signal time domain diagram and frequency domain diagram for sunny days are shown in Figure 4, the signal time domain diagram and frequency domain diagram for cloudy days are shown in Figure 5, and the signal time domain diagram and frequency domain diagram for rainy days are shown in Figure 6.
[0055] Step 5: Normalize the data and use it as an input variable for photovoltaic power generation prediction;
[0056] The decomposition results obtained in step 4 are processed by normalization and standardization to handle each intrinsic modal component. The processed dataset is then divided, with the first 80% used as the training set and the last 20% as the test set. The weather factor sequences processed in step 1 are extracted to form a meteorological feature matrix, which, along with the photovoltaic power component sequences processed in this step, are used as input variables for the prediction model.
[0057] Step 6: Use parallel deep learning with CNN and GRU networks to predict photovoltaic power generation, and finally obtain the predicted value of total photovoltaic power generation.
[0058] Step 6.1: Use a CNN neural network to identify and output the local features of the meteorological feature matrix in Step 5;
[0059] The GRU neural network is used to identify the time-series characteristics of each component sequence of photovoltaic power after normalization in step 5, and this is used as the input for the next stage.
[0060] Step 6.3: After the data has undergone parallel learning by the two networks in Step 6.1, the output values are concatenated into a high-dimensional vector and input into the CNN neural network model to perform deeper learning and training on the network under multiple feature inputs, and obtain the final prediction result.
[0061] Comparative Example
[0062] Using the same original dataset, predictions were made using CNN neural networks, LSTM neural networks, SVM neural networks, and BP neural networks, respectively. The CNN-GRU hybrid neural network prediction results for sunny days in this invention are as follows: Figure 7 As shown, the sunny weather prediction results of each model are compared to... Figure 8 As shown, the multi-cloud prediction results of the CNN-GRU hybrid neural network of the present invention are as follows: Figure 9 As shown, the cloud prediction results of each model are compared to, for example... Figure 10 As shown, the rain prediction results of the CNN-GRU hybrid neural network of the present invention are as follows: Figure 11 As shown, the rain prediction results of each model are compared to, for example... Figure 12 As shown, the specific predicted data are as follows:
[0063] Table 1. Prediction data and mean square error corresponding to the prediction model:
[0064]
[0065] As shown in Table 1, the CNN-GRU hybrid neural network of the present invention extracts the data features of the meteorological feature matrix and the photovoltaic power sequence by initially parallelizing the CNN and GRU neural networks, and then uses the CNN model to predict the integrated results. Compared with models such as CNN, GRU, SVM, and BP, the obtained prediction results have extremely small mean square error (MSE) and high prediction accuracy, thus achieving high-precision photovoltaic prediction.
[0066] The comparison of the prediction results of the above models for three different weather conditions (sunny, cloudy, and rainy) shows that traditional models such as CNN, GRU, SVM, and BP have a significant decrease in prediction performance for cloudy and rainy days compared to sunny days, while the CNN-GRU hybrid neural network model can still maintain a good prediction performance for cloudy and rainy days.
[0067] As described above, this invention presents a deep learning-based ultra-short-term photovoltaic power prediction method. Its main advantages include: employing a clustering set method to cluster historical data, combining the advantages of three clustering algorithms to obtain a more reasonable dataset partitioning, thereby improving prediction accuracy; using the PSO algorithm to optimize the VMD algorithm to avoid errors caused by manually set parameters, improving overall prediction accuracy; utilizing CNN and GRU networks for parallel learning to identify local and temporal features of the data respectively, enabling the network structure to fully mine the input data; and further learning through CNN neural network splicing and fusion to achieve high-precision prediction. Furthermore, this prediction model can be used for photovoltaic power prediction under various weather conditions, maintaining high accuracy and holding significant importance for the entire scheduling system.
Claims
1. A deep learning-based ultra-short-term photovoltaic power prediction method, characterized in that, The specific steps are as follows: Step 1: Preprocess the historical data of photovoltaic power generation to obtain preprocessed data; Step 2: Cluster the preprocessed data into three different weather types: sunny, rainy, and cloudy. Step 2 specifically involves, Step 2.1: Cluster the preprocessed data using Gaussian mixture model and hierarchical clustering respectively. Since the label assignment rules of these two different clustering methods are not consistent when clustering the preprocessed data, the label results of the two clustering methods are rearranged using the Hungarian algorithm to convert the labels of the same type of data into consistent labels. Step 2.2: Test the clustering accuracy of the two clustering results, extract the consensus points and non-consensus points of the two clustering methods, use the consensus points as the training set and the non-consensus points as the test set, calculate the Euclidean distance between the test set and each sample point in the training set, sort all the distance values, select the k samples with the smallest distance as "voters", predict the classification of the test sample based on the "voters", implement the K-nearest neighbor algorithm to predict the non-consensus points, obtain the clustering results of the non-consensus points, and use the set of the predicted non-consensus point clustering results and the consensus points as the final clustering result; Step 2.3: Evaluate the final clustering results based on the evaluation indicators of the Kalinsky-Harabass index, the Davidson-Bolding index, and the silhouette coefficient. Step 3: Use optimization algorithms to optimize the parameters in the variational mode decomposition algorithm to obtain the parameter-optimized signal decomposition model; Step 4: Use the optimized signal decomposition model from Step 3 to perform signal decomposition on the photovoltaic power generation sequence in the dataset to obtain the sequence of each component of the photovoltaic power. Step 5: Normalize the data and use it as an input variable for photovoltaic power generation prediction; Step 6: Use parallel deep learning with CNN and GRU networks to predict photovoltaic power generation, and finally obtain the predicted value of total photovoltaic power generation.
2. The deep learning-based ultra-short-term photovoltaic power prediction method according to claim 1, characterized in that, Step 1 specifically involves using the 3-sigma principle to detect abnormal data in photovoltaic power generation, removing missing and abnormal data; using Pearson coefficient analysis to calculate the correlation coefficients between various weather factors and photovoltaic power in historical data, removing feature vectors with correlation coefficients of 1 and negative values, using the remaining feature vectors as the main meteorological factors affecting photovoltaic output, and normalizing the remaining data to obtain preprocessed data.
3. The deep learning-based ultra-short-term photovoltaic power prediction method according to claim 1, characterized in that, Step 3 specifically involves, Step 3.1: Set the parameters and fitness function in the particle swarm optimization algorithm; Step 3.2: Initialize particle positions and velocities, setting the penalty factor and mode number parameters from variational mode decomposition to the particle positions; input all initialized particle positions into the variational mode decomposition algorithm, and calculate the component entropy values of all intrinsic mode components for each particle position; compare the magnitudes of local minimum entropies, and update individual local minimums and the overall global minimum; update particle velocities and positions; optimize the penalty factor and mode number parameters in the variational mode decomposition algorithm using the particle swarm optimization algorithm, determine whether the maximum number of iterations is satisfied, calculate the minimum network envelope entropy of global particles and the component envelope entropy of each intrinsic mode component, thereby obtaining the optimal penalty factor and mode number.
4. The deep learning-based ultra-short-term photovoltaic power prediction method according to claim 3, characterized in that, Specifically, step 4 involves setting the initial value of the variational mode decomposition algorithm based on the optimal penalty factor and mode number obtained by the particle swarm algorithm in step 3, decomposing and calculating the photovoltaic power of the three weather types to obtain the time domain diagram of the decomposed signal; and then using the fast Fourier transform to obtain the frequency domain diagram.
5. The deep learning-based ultra-short-term photovoltaic power prediction method according to claim 1, characterized in that, Step 5 specifically involves processing the decomposition results obtained in step 4 by normalizing and standardizing each intrinsic modal component in the decomposition results; dividing the processed dataset into a training set (first 80%) and a test set (last 20%); extracting the weather factor sequences processed in step 1 to form a meteorological feature matrix, which, along with the photovoltaic power component sequences processed in this step, serves as the input variables for the prediction model.
6. The deep learning-based ultra-short-term photovoltaic power prediction method according to claim 5, characterized in that, Specifically, step 6 is as follows: Step 6.1: Use a CNN neural network to identify and output the local features of the meteorological feature matrix in Step 5; The GRU neural network is used to identify the time-series characteristics of each component sequence of photovoltaic power after normalization in step 5, and this is used as the input for the next stage. Step 6.2: After the data has undergone parallel learning by the two networks in Step 6.1, the output values are concatenated into a high-dimensional vector and input into the CNN neural network model to perform deeper learning and training on the network under multiple feature inputs, so as to obtain the final prediction result.