Improved hybrid photovoltaic power prediction method based on sparse variational gaussian process

Abstract

This invention discloses a hybrid photovoltaic power prediction method based on a sparse variational Gaussian process, belonging to the field of photovoltaic power prediction technology. The method first acquires and processes meteorological parameters to construct a feature matrix, then constructs a prediction model integrating a convolutional neural network, a bidirectional long short-term memory network, an attention mechanism, and a sparse variational Gaussian process model. After the feature matrix is ​​processed by the deep learning module at the model's front end, preliminary point prediction results of photovoltaic power are obtained. Then, a feature vector is constructed by combining highly correlated meteorological parameters, time encoding, and this preliminary result, and input into the sparse variational Gaussian process model at the back end, finally outputting accurate point prediction results and quantifying their uncertainty. This method effectively integrates the feature extraction capabilities of deep learning with the uncertainty quantification advantages of probabilistic models, significantly reducing computational complexity through sparse variational inference, and achieving high-precision, high-efficiency photovoltaic power prediction with a reliable confidence interval.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power prediction technology, specifically to a hybrid photovoltaic power prediction method based on a sparse variational Gaussian process. Background Technology

[0002] As the energy structure shifts towards cleaner and lower-carbon energy, photovoltaic (PV) power generation, as a crucial component of renewable energy, has become a key support for the new power system due to its abundant resources and environmental friendliness. However, PV power output is significantly affected by meteorological conditions (such as irradiance, temperature, and wind speed), exhibiting strong randomness, volatility, and intermittency. Large-scale PV grid connection may cause problems such as grid voltage fluctuations and frequency deviations, posing challenges to the safe and stable operation of the power system. Therefore, improving the accuracy of PV power forecasting has become one of the key technologies for enhancing the absorption capacity of new energy sources and ensuring the safe operation of the power grid.

[0003] Currently, photovoltaic power prediction methods are mainly divided into physical model methods, statistical model methods, and machine learning methods. Physical models rely on equipment parameters and meteorological data, which are complex to model and have poor versatility; statistical models, such as time series regression, have high requirements for data quality and are difficult to adapt to complex scenarios; machine learning methods (such as support vector machines and random forests) perform well under small sample conditions, but their structures are fixed and they are prone to getting trapped in local optima, making it difficult to fully explore the complex nonlinear characteristics in multivariate time series data.

[0004] In recent years, deep learning methods such as Convolutional Neural Networks (CNNs), Long Short-Term Memory Networks (LSTMs), and their variants (such as Bidirectional LSTMs) have been widely applied to photovoltaic power prediction. CNNs excel at extracting local spatiotemporal features, BiLSTMs can capture bidirectional temporal dependencies, and attention mechanisms can adaptively allocate feature weights. Hybrid models combining these three technologies have shown good performance in multivariate time-series prediction. However, existing research mainly focuses on data preprocessing and model structure optimization, with relatively insufficient research on post-processing error correction of prediction results. Furthermore, most models can only output deterministic point predictions and lack the ability to quantify prediction uncertainty.

[0005] In probabilistic prediction, Gaussian process regression (GPR) is an effective probabilistic modeling method due to its nonparametric properties and natural posterior variance output. However, the computation time required for standard GPR prediction increases dramatically with the number of training samples, making it difficult to handle large-scale time series data. While methods such as quantile regression (QR) can provide probability intervals, they typically require pre-setting the error distribution, and photovoltaic prediction errors generally exhibit skewed and heavy-tailed characteristics, making them unsuitable for practical applications. Therefore, a post-processing method that can both correct point prediction bias and provide efficient probabilistic prediction capabilities is urgently needed.

[0006] In summary, existing technologies still have the following shortcomings in photovoltaic power prediction: Prediction accuracy is limited: a single model cannot fully describe the complex nonlinear relationship between meteorological factors and photovoltaic output; Lack of error correction mechanisms: Most methods focus on input feature optimization or model structure improvement, with insufficient research on post-processing error correction of prediction results; Weak ability to quantify uncertainty: Existing probabilistic prediction methods have high computational complexity or are sensitive to assumptions about error distribution, making it difficult to balance accuracy and efficiency. Low computational efficiency: Traditional methods such as Gaussian processes take too long to train on large-scale data, which limits their application in practical engineering. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a hybrid photovoltaic power prediction method based on a sparse variational Gaussian process, which aims to solve the problems in the background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a hybrid photovoltaic power prediction method based on a sparse variational Gaussian process, comprising the following steps: The NWP characteristics of different meteorological variables on the day to be measured are obtained, and the different meteorological variables and their NWP characteristics are used as meteorological parameters. The linear correlation between different meteorological parameters and photovoltaic power is quantified, and the meteorological parameter with the strongest correlation with photovoltaic power is selected. The meteorological parameters are normalized and preprocessed and then organized into a feature matrix. A CNN-BiLSTM-Attention prediction model based on SVGP was constructed. The prediction model includes a convolutional neural network, a bidirectional long short-term memory network, an attention mechanism module, and a sparse variational Gaussian process model. The feature matrix is ​​input into a CNN-BiLSTM-Attention prediction model based on SVGP. The feature matrix is ​​processed sequentially through a convolutional neural network, a bidirectional long short-term memory network, and an attention mechanism module to obtain the output of the attention mechanism module. The output of the attention mechanism module is flattened by a Flatten layer and then input into a fully connected layer to output a normalized initial prediction value of photovoltaic power. The normalized initial prediction value of photovoltaic power is then denormalized to obtain the preliminary point prediction result of photovoltaic power. A feature vector is constructed using the meteorological parameter with the strongest linear correlation with photovoltaic power, the preset time code, and the preliminary point prediction result of photovoltaic power. The feature vector is used as the input of a sparse variational Gaussian process model to output the final prediction result of photovoltaic power.

[0009] Furthermore, the Pearson correlation coefficient was used to analyze the correlation between photovoltaic power and meteorological parameters, quantifying the degree of linear correlation between different meteorological parameters and photovoltaic power.

[0010] Furthermore, the specific processing procedure of the convolutional neural network is as follows: the feature matrix is ​​input into the convolutional neural network, the local spatiotemporal features of the feature matrix are extracted through one-dimensional convolution operation, the feature matrix is ​​mapped to the hidden layer feature space through the ReLU activation function, and then the key information is extracted through the pooling layer to obtain the output of the convolutional neural network.

[0011] Furthermore, the specific processing procedure of the bidirectional long short-term memory network is as follows: the output of the convolutional neural network is used as the input of the bidirectional long short-term memory network, and the long-term dependencies of the time series data are captured bidirectionally through the forward and backward long short-term memory networks, and the bidirectional time series fusion features are output.

[0012] Furthermore, the specific processing procedure of the attention mechanism module is as follows: the output of the bidirectional long short-term memory network is used as the input of the attention mechanism module, and the output of the bidirectional long short-term memory network at each time step is adaptively weighted by the scaling dot product attention mechanism to focus on key features and alleviate feature redundancy, and output weighted fused features.

[0013] Furthermore, the sparse variational Gaussian process regression model introduces induced points to construct variational inference approximations; calculates the covariance matrix between the test point and the induced point, the covariance matrix between the test points, and the covariance matrix between the induced points; uses the a posteriori mean as the final photovoltaic power prediction result, quantifies the prediction uncertainty using the a posteriori variance, and constructs a confidence interval.

[0014] An electronic device includes a processor, a memory, and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program code, and the processor is used to call the program code stored in the memory to execute a hybrid photovoltaic power prediction method based on a sparse variational Gaussian process.

[0015] A non-volatile computer storage medium storing computer-executable instructions that execute a hybrid photovoltaic power prediction method based on a sparse variational Gaussian process.

[0016] Compared with existing technologies, the present invention has the following advantages: (1) This invention constructs a hybrid prediction framework combining a CNN-BiLSTM-Attention model and a Sparse Variational Gaussian Process (SVGP) model. The front-end deep learning model effectively extracts the deep nonlinear features between meteorological time-series data and photovoltaic power, achieving preliminary high-precision point prediction; the back-end SVGP model corrects errors and quantifies uncertainties in the preliminary prediction results. This innovative combination solves the comprehensive problem that a single model cannot simultaneously take into account the in-depth mining of time-series features, prediction error correction, and prediction uncertainty representation. At the same time, the SVGP module uses induced points to replace the full training samples, compressing the training information into a small number of fixed model parameters. In the prediction stage, only the fixed parameters and real-time test data need to be called for calculation, thereby significantly reducing the computational complexity of large-scale time-series data, achieving photovoltaic power prediction with both high precision and high real-time performance, and providing complete prediction information including confidence intervals for subsequent power dispatch decisions.

[0017] (2) This invention relies on the sparsity and variational inference core mechanism of the SVGP model. Its prediction process completely eliminates the real-time dependence on the original massive training data, and only needs to be calculated based on the induced points and related covariance matrices fixed in the training stage. This avoids the huge computational overhead caused by traditional Gaussian process regression and other methods that need to access all training samples during prediction, ensuring the simplicity and efficiency of the prediction process. It can better meet the stringent requirements of real-time response for multi-time-step, high-frequency power prediction services of photovoltaic power plants, and has excellent engineering application feasibility.

[0018] (3) This invention adopts an innovative architecture of multimodal feature fusion and hierarchical prediction, which integrates preliminary point prediction results, highly correlated meteorological parameter values, and time-coded information into a feature vector, which is then input into the SVGP module. This enables the model to comprehensively capture the complex patterns of photovoltaic power being affected by various factors such as changes in meteorological conditions and time-cycle effects, thereby improving the model's predictive adaptability and robustness under different meteorological conditions and time periods. In addition, the SVGP module directly outputs the posterior variance to construct the prediction confidence interval, effectively compensating for the shortcomings of traditional deterministic prediction models that can only provide single-point prediction values ​​and cannot assess the reliability of prediction results, thus providing more comprehensive data support for the power planning and operational risk assessment of photovoltaic power plants. Attached Figure Description

[0019] Figure 1 This is a heat map showing the correlation coefficients of photovoltaic meteorological elements.

[0020] Figure 2 This is a structural diagram of the CNN-BiLSTM-Attention prediction model based on SVGP.

[0021] Figure 3This is a comparison chart of the results of different forecasting methods under different weather conditions.

[0022] Figure 4 This is a scatter plot comparing CNN-BiLSTM-Attention and CNN-BiLSTM-Attention-SVGP under different weather conditions.

[0023] Figure 5 The graph shows the probability prediction results of CNN-BiLSTM-Attention-SVGP under different weather conditions.

[0024] Figure 6 This is a comparison chart of the evaluation results of different probability prediction methods. Detailed Implementation

[0025] Example 1

[0026] like Figure 1 As shown, the present invention provides a technical solution: a hybrid photovoltaic power prediction method based on a sparse variational Gaussian process, comprising the following steps: Step S1: Obtain the NWP characteristics of different meteorological variables (such as solar irradiance, ambient temperature, wind speed, etc.) for the day to be measured, and use the different meteorological variables and their NWP characteristics as meteorological parameters; quantify the linear correlation between different meteorological parameters and photovoltaic power, and select the meteorological parameter with the strongest correlation to photovoltaic power (in this embodiment, the NWP characteristics of irradiance are selected); normalize and preprocess the meteorological parameters and organize them into a feature matrix.

[0027] The prediction task of this invention is to predict the day-ahead photovoltaic power with a time resolution of 15 minutes and a prediction duration of 24 hours (i.e., 96 time steps). The NWP characteristics of different meteorological variables on the day to be measured can be directly obtained from the meteorological department. The NWP characteristics are numerical weather prediction characteristics.

[0028] The Pearson correlation coefficient was used to analyze the correlation between photovoltaic power and meteorological parameters, quantifying the degree of linear correlation between different meteorological parameters and photovoltaic power; the analysis results are shown in the following heat map. Figure 1 As shown. The calculation formula is as follows: ; In the formula, This represents the Pearson correlation coefficient, used to measure the degree of linear correlation between photovoltaic power and meteorological parameters; Indicates the total number of meteorological parameters; Indicates the first Meteorological parameters corresponding to each time step; express The mean; Indicates the first Measured photovoltaic power values ​​corresponding to each time step; express The mean; Correlation analysis revealed that among meteorological parameters, irradiance and its NWP characteristics were strongly positively correlated with photovoltaic power (correlation coefficients > 0.95), while temperature and wind speed and their NWP characteristics were weakly positively and weakly negatively correlated with photovoltaic power, respectively (absolute values ​​of correlation coefficients were between 0.3 and 0.4). Feature matrix Represented as: ; In the formula, Representation of the characteristic matrix The Middle line, number The elements of the column represent the first element. The first time step The values ​​of the meteorological parameters after normalization preprocessing.

[0029] The power output of renewable energy power plants is dominated by meteorological conditions. Compared to prediction methods that rely solely on historical power data, incorporating meteorological data can significantly improve prediction accuracy. This invention combines historical meteorological data with natural weather prediction (NWP) data to construct input features. Historical meteorological data depicts the evolution of meteorological elements, while NWP data forecasts future meteorological conditions based on historical data. This combination allows the model to learn the mapping relationship between meteorology and power output from historical data. Simultaneously, by leveraging the trend predictions of future meteorological changes using NWP data, the model dynamically captures the temporal impact of weather processes on power output, thereby improving prediction accuracy.

[0030] Step S2: Construct a CNN-BiLSTM-Attention prediction model based on SVGP; the prediction model includes a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), an attention mechanism module (Attention), and a sparse variational Gaussian process model (SVGP).

[0031] Step S3: Input the feature matrix into the CNN-BiLSTM-Attention prediction model based on SVGP. The feature matrix is ​​processed sequentially through a convolutional neural network, a bidirectional long short-term memory network, and an attention mechanism module to obtain the output of the attention mechanism module. After flattening the output of the attention mechanism module through a Flatten layer, it is input into a fully connected layer with 96 neurons to output the normalized initial prediction value of photovoltaic power for 96 time steps of the day to be measured. After inverse normalization of the normalized initial prediction value of photovoltaic power, the preliminary point prediction result of photovoltaic power is obtained. A feature vector is constructed using the meteorological parameter with the highest linear correlation with photovoltaic power, the preset time code, and the preliminary point prediction result of photovoltaic power. The feature vector is used as the input of the Sparse Variational Gaussian Process (SVGP) model to output the final prediction result of photovoltaic power.

[0032] The specific processing procedure of the convolutional neural network is as follows: the feature matrix X is input into the convolutional neural network, and the local spatiotemporal features of the feature matrix are extracted through one-dimensional convolution operations. These features are then mapped to the hidden layer feature space via the ReLU activation function, and finally, the key information is extracted through pooling layers to obtain the output of the convolutional neural network, represented as: ; ; ; ; In the formula, This represents the feature matrix of layer 0 (input layer), which is the initial input feature of the entire convolutional neural network; Indicates the first The output feature matrix of each convolutional layer; Represents the ReLU activation function; Indicates the first The output feature matrix of the layer, i.e., the first layer The input feature matrix of each convolutional layer; Indicates the first The convolution kernel of a convolutional layer; Indicates the first The bias parameters corresponding to each convolutional layer; This represents the input value of the ReLU activation function, i.e., the linear result of the convolution operation; Indicates the first The first layer of pooling Each output feature value; Indicates the index of the pooling window.

[0033] The specific processing procedure of the bidirectional long short-term memory network is as follows: the output of the convolutional neural network is used as the input of the bidirectional long short-term memory network. The long-term dependencies of time-series data are captured bidirectionally through both forward and backward long short-term memory networks, outputting bidirectional temporal fusion features, represented as: ; ; In the formula, This indicates that bidirectional long short-term memory networks are in The input features at each time step, i.e., the output of the convolutional neural network; The mapping function representing the standard Long Short-Term Memory network is common knowledge in the field of deep learning. It contains forget gate, input gate, output gate and gating computation of memory cells. This indicates that the positive long short-term memory network is in Output characteristics at any given time; This indicates that the positive long short-term memory network is in Output characteristics at any given time; This indicates that the inverse long short-term memory network is in Output characteristics at any given time; This indicates that the inverse long short-term memory network is in Output characteristics at any given time; Indicates feature concatenation operation; This indicates that bidirectional long short-term memory networks are in The output characteristics at each time step, i.e., the bidirectional temporal fusion characteristics; Traditional unidirectional LSTMs can only transmit information in a single direction along the time series, and their ability to characterize the overall pattern of the sequence is limited when processing sequence data with strong temporal dependencies and global correlations. BiLSTMs, by constructing forward and reverse sequences, more comprehensively capture the temporal correlations and change patterns within the data.

[0034] The specific processing procedure of the attention mechanism module is as follows: The output of the bidirectional long short-term memory network is used as the input to the attention mechanism module. An adaptive weight allocation is performed on the output of the bidirectional long short-term memory network at each time step using a scaled dot product attention mechanism. This focuses on key features and alleviates feature redundancy, outputting weighted fused features. Specifically: The complete output feature matrix of the bidirectional long short-term memory network is formed by stacking the outputs of the bidirectional long short-term memory network at all time steps in rows; The complete output feature matrix of the bidirectional long short-term memory network is obtained through three independent learnable linear transformation matrices. Mapped to query vectors respectively Key vector and value vector , is represented as: ; In the formula, , , They are query vectors Key vector and value vector Learnable linear transformation matrix; Calculate query vector With key vector The similarity score is obtained by scaling the dot product similarity, and is expressed as: ; In the formula, Represents the key vector Dimension size; Indicates the first Similarity score at each time step; Indicates the first The transpose of the key vector at each time step; The attention weights are obtained by normalizing the similarity scores using the Softmax function. , is represented as: ; In the formula, Indicates the first Attention weights for each time step; Indicates the first Similarity score at each time step; Represents the Softmax function; Indicates the natural index; Indicates the number of time steps; Value vector based on attention weights After performing a weighted summation, we obtain the weighted fusion feature, which is the output of the attention mechanism module, expressed as: ; In the formula, This represents the final output of the attention mechanism module; Indicates the first The value vector at each time step; The attention mechanism is a mechanism that mimics human visual attention. It can automatically focus on key information in the input data and effectively alleviate the gradient vanishing and long dependency problems of traditional neural networks in long sequence tasks.

[0035] The process of constructing a feature vector using meteorological parameters with the highest linear correlation to photovoltaic power, preset time codes, and preliminary point prediction results of photovoltaic power, and then using this feature vector as input to a sparse variational Gaussian process model (SVGP) to output the final prediction result of photovoltaic power is as follows: A feature vector is constructed using meteorological parameters with the highest linear correlation to photovoltaic power, preset time codes, and preliminary photovoltaic power prediction results. , This indicates the preliminary point prediction results of photovoltaic power. This represents the meteorological parameter with the highest degree of linear correlation to photovoltaic power. Represents the encoding at a preset time; for the feature vector Iterate through the entire prediction duration (i.e., 96 time steps) to obtain the input set for the prediction duration. ; The input set is calculated based on the Sparse Variational Gaussian Process (SVGP) model. The corresponding predicted posterior distribution: ; In the formula, The Sparse Variational Gaussian Process (SVGP) model is based on the input set. The posterior probability density function of the photovoltaic power random variable was calculated. Represents the random variable of photovoltaic power to be predicted; This represents the posterior mean; Indicates the posterior variance; This represents a normal distribution (Gaussian distribution). Calculate the posterior mean and posterior variance: ; ; In the formula, This represents the covariance matrix between the test points and the induced points; express The transpose of the matrix; Represents the covariance matrix between test points; This represents the covariance matrix between the induced points; This represents the mean vector of the variational distribution at the induced point; The covariance matrix representing the variational distribution of the induced point; Wherein, the test points are the input set. The individual elements in the model, where the inducing points are a fixed set of parameters built into the Gaussian process regression model, are used to calculate the similarity between each test point and each inducing point using the pre-defined squared exponential kernel function of the Gaussian process regression model, thus obtaining the covariance matrix between the test points and the inducing points. , represented as: ; In the formula, Indicates the first The test point and the first Covariance between the induced points; Represents the quadratic exponential kernel function; Indicates the first One test point; Indicates the first One inducement point; Calculate the similarity between each test point and all other test points to obtain the covariance matrix among the test points. , is represented as: ; In the formula, Indicates the first The test point and the first Covariance values ​​between test points; Indicates the first One test point; Indicates the first One test point; Calculate the similarity between each induced point and all other induced points to obtain the covariance matrix among the induced points. , is represented as: ; In the formula, Indicates the first The first induction point and the second Covariance between the induced points; Indicates the first One inducement point; Indicates the first One induction point.

[0036] posterior mean As the final point prediction result of photovoltaic power, the a posteriori variance Quantify the uncertainty of forecasting and construct a 95% confidence interval: Complete photovoltaic power prediction.

[0037] During the model training phase, a sliding window method can be used to simulate NWP characteristics based on historical meteorological data: For each meteorological variable in the historical meteorological data, a sliding window length of 4 time steps (corresponding to 1 hour) is set. The mean of the historical observations within the sliding window is calculated, and random noise based on historical difference data is superimposed. The noise follows a Gaussian distribution to obtain the NWP characteristics of the historical meteorological variables; the calculation formula is as follows: ; In the formula, Indicates the reference time , the future Simulated NWP features at each time step; Indicates the length of the sliding window; Indicates the first Each time step corresponds to the historical observed value of the meteorological variable; Indicates the reference time , the future Random noise term superimposed at each time step; This represents a normal distribution (Gaussian distribution). This represents the variance of the Gaussian distribution.

[0038] To improve prediction accuracy and characterize prediction uncertainty, a post-processing model based on Sparse Variational Gaussian Processes (SVGP) is constructed: using historical measured power values ​​of photovoltaic power plants as the target variable, training samples are built based on historical feature vectors and historical measured power values ​​of photovoltaic power plants; a training set for the Gaussian process regression model is then constructed based on all training samples. , Indicates the first Feature vectors of each historical time step Indicates the first Measured power values ​​of photovoltaic power plants at each historical time step. Indicates the total number of training samples; Introduction There are 1 induced points, and the set of induced points is . ,in , Indicates the number of induction points. Indicates the first There are induced points; the function values ​​corresponding to the induced points. Follows the prior distribution: ; In the formula, This represents the prior probability distribution of the function values ​​corresponding to the induced points; Indicates the first One inducement point; Indicates implicit function; Constructing a joint variational distribution: ; In the formula, Represents the variational distribution of the induced point. Represents the variational mean vector; Represents the variational covariance matrix; The implicit function represents the function value at a given induced point. Conditional distribution; With training set Using the input as input, and aiming to minimize the lower bound of evidence (ELBO), optimize the location of the induction point. Variational parameters and The trained SVGP model is obtained by combining the squared exponential kernel function parameters.

[0039] A second embodiment of the present invention also provides an electronic device, including a processor, a memory, and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program code, and the processor is used to call the program code stored in the memory to execute a hybrid photovoltaic power prediction method based on a sparse variational Gaussian process.

[0040] A third embodiment of the present invention also provides a non-volatile computer storage medium storing computer-executable instructions that execute a hybrid photovoltaic power prediction method based on a sparse variational Gaussian process.

[0041] To comprehensively evaluate the performance of the CNN-BiLSTM-Attention prediction model based on SVGP, two evaluation metrics were used: point prediction and probability prediction.

[0042] Point prediction uses mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), and coefficient of determination R. 2 Four evaluation indicators.

[0043] The probability prediction uses Prediction Interval Coverage Probability (PICP), Prediction Interval Normalized Average Width (PINAW), Continuous Ranked Probability Score (CRPS), and Training Time as evaluation metrics.

[0044] In this embodiment, for the CNN part, the number of convolutional kernels in the one-dimensional convolutional layer is set to 64, the kernel size is set to 1, and the activation function is the ReLU function. To prevent overfitting, the output of the convolutional layer is regularized by replacing the pooling layer with a Dropout layer, and the Dropout layer parameter is set to 0.3. The input is fed into a bidirectional LSTM layer, which adopts a single-layer bidirectional structure with 256 neurons each in the forward and backward directions. It is also regularized using a Dropout layer with a parameter set to 0.3. The attention part generates time-step attention weights through a Dense layer and a softmax activation function. Finally, after flattening the features through a Flatten layer, the input is fed into a fully connected layer with 96 neurons to output the prediction result. The model is trained for 250 epochs, and an early stopping strategy is used to prevent overfitting. The loss function during network compilation is MAE. The SVGP induction point M is set to 500.

[0045] To investigate the impact of this invention on the accuracy of photovoltaic power prediction, three typical weather conditions—sunny, cloudy, and rainy—were selected. Five prediction models—CNN, BiLSTM, CNN-BiLSTM, CNN-BiLSTM-Attention, and CNN-BiLSTM-Attention-SVGP—were used to predict the power sequence of the target photovoltaic power plant. The results of different prediction methods under different weather conditions were compared. Figure 3 As shown in the figure, the scatter plots comparing CNN-BiLSTM-Attention and CNN-BiLSTM-Attention-SVGP under different weather conditions are as follows: Figure 4 Table 1 shows a comparison of the error results of different point prediction methods; CNN-BiLSTM-Attention-SVGP is the CNN-BiLSTM-Attention prediction model based on SVGP improved in this invention.

[0046] Table 1 Comparison of error results for different point prediction methods

[0047] from Figure 3 It is evident that the performance of each model varies significantly under different weather conditions: in sunny weather, the predicted curves of each model closely match the actual values, with a relatively small overall error; in cloudy weather, irradiance fluctuations increase, but the models still track the overall trend well; while in rainy weather, power fluctuations are drastic, and the prediction errors of each model increase significantly. However, overall, the CNN-BiLSTM-Attention-SVGP model demonstrates superior prediction performance under all weather conditions.

[0048] Figure 4This is a scatter plot comparing the point prediction results of the CNN-BiLSTM-Attention model and the CNN-BiLSTM-Attention-SVGP model under different weather conditions. The horizontal axis represents the predicted value, the vertical axis represents the true value, and the red diagonal line represents the ideal prediction situation. The plot shows that the scatter points of both models are relatively concentrated and close to the ideal prediction line under sunny and cloudy weather conditions; while the power fluctuates drastically on rainy days, and the scatter point distribution is more dispersed. However, overall, the CNN-BiLSTM-Attention-SVGP model is closer to the ideal prediction line under all weather conditions, with a relatively smaller error.

[0049] To further evaluate the model's ability to characterize uncertainty, a probabilistic prediction experiment was conducted. Three typical weather conditions—sunny, cloudy, and rainy—were selected. Based on the CNN-BiLSTM-Attention model, SVGP was used for probabilistic modeling of photovoltaic power point prediction. The prediction results under different weather conditions are as follows: Figure 5 As shown in Table 2, five probability prediction methods—KDE, BLR, QR, GPR, and SVGP—were trained on the CNN-BiLSTM-Attention model and then compared and analyzed. The relevant evaluation metrics, PICP, PINAW, and CRPS, are shown in Table 3. In addition, the training time comparison of each probability model is shown in Table 3.

[0050] Table 2 Comparison of Evaluation Results of Different Probability Prediction Methods

[0051] Table 3 Comparison of training time results for different probability methods

[0052] from Figure 6 It can be seen that under different weather conditions, the model can effectively depict the trend of photovoltaic power variation and provide corresponding uncertainty intervals. Under sunny conditions, the prediction interval is relatively narrow and smooth, and the actual values ​​basically fall within the confidence interval. Under cloudy conditions, the interval width increases, which can reflect the uncertainty brought about by irradiance fluctuations. Under rainy conditions, due to the drastic power fluctuations, the coverage of the prediction interval decreases, and there are deviations in some periods, but overall it can still reflect the trend of uncertainty changes.

[0053] Tables 2 and 3 present the performance and training time comparisons of different probability prediction methods under three typical weather conditions: sunny, cloudy, and rainy. Overall, the GPR model achieved the best or near-best performance in most scenarios, showing significant advantages in interval width and overall error. The SVGP model's overall performance is close to GPR, but slightly lower in various indicators, mainly due to errors caused by variational approximation. Under cloudy conditions, although the KDE method has a high PICP, its PINAW is significantly larger, indicating that its prediction results are relatively conservative as it increases coverage by widening the interval width. Under rainy conditions, due to increased data volatility, the performance of all models declines, but GPR still maintains a superior CRPS, with SVGP approaching it. In terms of computational efficiency, GPR has the longest training time due to its cubic complexity calculations, while SVGP effectively reduces computational complexity by introducing induced points. While maintaining near-prediction accuracy, its training time is only 36.6% of GPR's, significantly improving computational efficiency. The KDE, BLR, and QR methods, due to their simple structure, all have training times of less than 1 second. In summary, GPR offers the best accuracy but incurs significant overhead during training, while SVGP strikes a good balance between accuracy and efficiency. Lightweight methods are suitable for scenarios with high real-time requirements. In practical applications, the appropriate method should be chosen flexibly based on the specific accuracy and efficiency needs.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A hybrid photovoltaic power prediction method based on a sparse variational Gaussian process, characterized in that, Includes the following steps: The NWP characteristics of different meteorological variables on the day to be measured are obtained, and the different meteorological variables and their NWP characteristics are used as meteorological parameters. The linear correlation between different meteorological parameters and photovoltaic power is quantified, and the meteorological parameter with the strongest correlation with photovoltaic power is selected. The meteorological parameters are normalized and preprocessed and then organized into a feature matrix. A CNN-BiLSTM-Attention prediction model based on SVGP was constructed. The prediction model includes a convolutional neural network, a bidirectional long short-term memory network, an attention mechanism module, and a sparse variational Gaussian process model. The feature matrix is ​​input into the CNN-BiLSTM-Attention prediction model based on SVGP. The feature matrix is ​​processed sequentially through a convolutional neural network, a bidirectional long short-term memory network, and an attention mechanism module to obtain the output of the attention mechanism module. The output of the attention mechanism module is flattened by the Flatten layer and then input into the fully connected layer to output the normalized initial prediction value of photovoltaic power. The normalized initial prediction value of photovoltaic power is then denormalized to obtain the preliminary point prediction result of photovoltaic power. A feature vector is constructed using the meteorological parameter with the strongest linear correlation with photovoltaic power, the preset time code, and the preliminary point prediction result of photovoltaic power. The feature vector is used as the input of the sparse variational Gaussian process model to output the final prediction result of photovoltaic power.

2. The hybrid photovoltaic power prediction method based on sparse variational Gaussian processes as described in claim 1, characterized in that: The Pearson correlation coefficient was used to analyze the correlation between photovoltaic power and meteorological parameters, quantifying the degree of linear correlation between different meteorological parameters and photovoltaic power.

3. The hybrid photovoltaic power prediction method based on sparse variational Gaussian processes as described in claim 2, characterized in that: The specific processing procedure of a convolutional neural network is as follows: the feature matrix is ​​input into the convolutional neural network, the local spatiotemporal features of the feature matrix are extracted through one-dimensional convolution operation, the feature matrix is ​​mapped to the hidden layer feature space through the ReLU activation function, and then the key information is extracted through the pooling layer to obtain the output of the convolutional neural network.

4. The hybrid photovoltaic power prediction method based on sparse variational Gaussian processes as described in claim 3, characterized in that: The specific processing procedure of the bidirectional long short-term memory network is as follows: the output of the convolutional neural network is used as the input of the bidirectional long short-term memory network. The long-term dependencies of time series data are captured bidirectionally through the forward and backward long short-term memory networks, and the bidirectional time series fusion features are output.

5. The hybrid photovoltaic power prediction method based on sparse variational Gaussian processes as described in claim 4, characterized in that: The specific processing procedure of the attention mechanism module is as follows: the output of the bidirectional long short-term memory network is used as the input of the attention mechanism module. The output of the bidirectional long short-term memory network at each time step is adaptively weighted by the scaling dot product attention mechanism to focus on key features and alleviate feature redundancy, and output weighted fused features.

6. The hybrid photovoltaic power prediction method based on sparse variational Gaussian processes as described in claim 5, characterized in that: The sparse variational Gaussian process regression model introduces induced points to construct a variational inference approximation; calculates the covariance matrix between the test point and the induced point, the covariance matrix between the test points, and the covariance matrix between the induced points; uses the posterior mean as the final photovoltaic power prediction result, quantifies the prediction uncertainty using the posterior variance, and constructs a confidence interval.

7. An electronic device, characterized in that, The system includes a processor, a memory, and a bus, wherein the processor and the memory are connected via the bus, the memory is used to store a set of program code, and the processor is used to call the program code stored in the memory to execute a hybrid photovoltaic power prediction method based on a sparse variational Gaussian process as described in any one of claims 1-6.

8. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer can execute instructions to perform a hybrid photovoltaic power prediction method based on a sparse variational Gaussian process, as described in any one of claims 1-6.

Images