Photovoltaic ultra-short-term probabilistic forecasting method based on data decomposition and reconstruction
By combining the ACEEMDAN algorithm and the improved EQN loss function, high-frequency and low-frequency features of photovoltaic power generation are extracted, solving the problems of computational complexity and prediction accuracy of existing methods. This improves the point prediction accuracy and uncertainty quantification of photovoltaic power generation, and is suitable for grid dispatch and distributed solar power grid connection.
Patent Information
- Application Number
- CN202610528948.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-21
- Publication Date
- 2026-07-21
AI Technical Summary
Existing photovoltaic power generation probability prediction methods suffer from problems such as high computational complexity, modal aliasing, frequency redundancy, inconsistent number of IMFs, difficulty in capturing high-frequency mutation features and low-frequency fluctuations with a single neural network, and overly conservative prediction results due to loss functions in 5-minute ultra-short-term predictions, which cannot meet the needs of engineering applications.
The ACEEMDAN algorithm is used to decompose the data into high-frequency and low-frequency components. Features are extracted through CNN, iTransformer, and BiLSTM networks. The prediction is performed by combining multi-head attention mechanism and improved EQN loss function, so as to realize multi-scale feature extraction and narrow and calibrated prediction range of photovoltaic power signal.
It significantly improves the accuracy of point prediction and uncertainty quantification of photovoltaic power generation. The model maintains stability and efficient computation on different datasets, adapts to ultra-short-term photovoltaic probability prediction at the 5-minute scale, and supports grid dispatch and grid connection of distributed solar power generation.
Smart Images
Figure CN122436952A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of photovoltaic power generation technology, specifically relating to a photovoltaic ultra-short-term probability prediction method based on data decomposition and reconstruction. Background Technology
[0002] Photovoltaic power generation is significantly affected by natural factors such as solar irradiance and atmospheric changes, exhibiting marked intermittency, volatility, and uncertainty. This characteristic poses a significant challenge to the safe and stable dispatch of the power grid and the refined management of distributed energy resources. Among these challenges, ultra-short-term probability prediction on a 5-minute timescale can simultaneously provide point prediction results for photovoltaic power generation and a quantitative range of uncertainty. It is a core technical means to support real-time power balance of the power grid and the absorption of new energy sources, and plays a crucial role in promoting the large-scale grid connection of distributed solar power generation.
[0003] Currently, photovoltaic (PV) power generation probabilistic prediction methods are mainly based on statistical models, Bayesian methods, and deep learning models. In recent years, ensemble methods have emerged that attempt to improve prediction accuracy by integrating the complementary advantages of multiple models to obtain precise probabilistic prediction results. However, existing methods still suffer from several intractable technical shortcomings: First, traditional signal decomposition methods such as CEEMDAN suffer from high computational complexity, inconsistent numbers of intrinsic mode functions (IMFs), frequency redundancy, and mode aliasing, making them unsuitable for the non-stationary and nonlinear characteristics of PV power. Second, single neural network models struggle to simultaneously and accurately capture both high-frequency abrupt changes and low-frequency fluctuations in PV power, and their insufficient exploration of the spatiotemporal dependencies of meteorological data limits point prediction accuracy. Third, existing EQN networks' loss functions only combine quantile loss and evidence loss, which can easily lead to models arbitrarily widening the prediction interval to increase coverage, resulting in overly conservative predictions and extremely low practical value for uncertainty quantification. Furthermore, most models perform poorly in mining intrinsic data relationships, cross-dataset generalization performance, and computational efficiency for engineering deployment, failing to meet the engineering application requirements for 5-minute ultra-short-term PV probabilistic prediction. Summary of the Invention
[0004] The purpose of this invention is to provide a photovoltaic ultra-short-term probability prediction method based on data decomposition and reconstruction, which solves the technical problem that existing photovoltaic ultra-short-term probability prediction methods cannot meet the engineering requirements of 5-minute ultra-short-term photovoltaic probability prediction.
[0005] The technical solution adopted in this invention is a photovoltaic ultra-short-term probability prediction method based on data decomposition and reconstruction, comprising the following steps:
[0006] S1: Collect photovoltaic power and meteorological data, and use linear interpolation to fill in missing values to complete data preprocessing; S2: Using the ACEEMDAN algorithm, the photovoltaic power data is decomposed into fixed high-frequency and low-frequency dual components through four stages: decomposition preprocessing, local decomposition, pseudo-mode merging, and IMF binary aggregation. The dual components are then regularized with meteorological data. S3: CNN, iTransformer, and BiLSTM are used to extract the corresponding features of high-frequency, low-frequency, and meteorological data, respectively. S4: The unified feature representation is obtained by fusing features through a multi-head attention mechanism and input into the EQN network to output the probability prediction result; S5: Train the EQN network until convergence using an improved EQN loss function with a width penalty term to obtain the prediction model.
[0007] The invention is further characterized by: The specific decomposition preprocessing in S2 is as follows: Calculate photovoltaic power data at a given time point Local standard deviation :
[0008] in, For a sliding window of radius w, The number of data points within the window. For in-window indexing, This represents the local average value of the data within the window. N=100 sets of standardized white noise conforming to the standard normal distribution are generated. The white noise is pre-decomposed using the EMD algorithm to obtain the noise intrinsic mode function components.
[0009] The local decomposition in S2 is specifically as follows: Perform adaptive noise modulation:
[0010] in, After modulation Moment Decomposition stage noise, Indicates the first Each decomposition stage Time residuals for Local annotation difference at time, Noise figure for The first moment of white noise One EMD component, Represents the current calculation of the number Group noise; The noise-modulated signal is decomposed using EMD and then ensemble-averaged to obtain the first... One intrinsic modulus function:
[0011] in, This represents the EMD decomposition process. For the first The intrinsic modulus functions obtained from the decomposition. Representing collective ownership Group noise; Update the residuals and repeat the above steps until the iteration termination condition is met, to obtain... The iteration terminates when the number of residual extrema is less than 3, the residuals show a monotonically increasing / decreasing trend, the residual energy is less than 1% of the total energy of the original signal, or the residual standard deviation is less than 1%. Any of the following conditions:
[0012] in, Representing the The residuals obtained from the second decomposition To The first time the photovoltaic power generation data at time 1 The result of the decomposition.
[0013] The pseudo-pattern merging in S2 is specifically as follows: Calculate the orthogonality index of adjacent intrinsic moduli:
[0014] Among them, the operator The dot product of vectors, operator Represents the Euclidean norm, operator Represents absolute value operation. Representative mode The orthogonality index between them; If adjacent intrinsic moduli functions If a pair is identified as a pseudo-pattern pair, merge the pseudo-pattern pairs:
[0015] in, For the index set of pseudo-pattern pairs, For the first index within the set One IMF; Repeat the above steps for all adjacent intrinsic moduli. The optimized independent intrinsic modulus functions are obtained.
[0016] The specific details of IMF binary aggregation in S2 are as follows: The power spectral density of each optimized intrinsic mode function is calculated using Fast Fourier Transform to determine the representative frequency.
[0017] in, For the first The representative frequency of each IMF For the first The main frequencies of the IMF For the first The center frequency of each IMF; Based on the prediction of specific frequency thresholds on a 5-minute scale, the intrinsic mode functions are divided into high-frequency and low-frequency groups:
[0018] in, For high frequency groups, For low frequency groups, For the i-th IMF, High-frequency threshold; Summing them separately yields the final high-frequency and low-frequency components:
[0019] in, For high-frequency components, Low-frequency components, For the i-th IMF, It is the residual vector; RobustScaler was used to perform regularization on high-frequency components, low-frequency components, and meteorological data. RobustScaler was also used to standardize the photovoltaic power target variable separately.
[0020] Specifically, S3 consists of: a CNN network taking regularized high-frequency components as input, capturing local mutation features of high-frequency components through convolutional layers, and outputting fixed-dimensional high-frequency features after dimensionality reduction by pooling layers; an iTransformer network taking regularized low-frequency components as input, embedding time-series variables as discrete tokens, capturing correlations and long-term trend features between variables through an attention mechanism, and outputting fixed-dimensional low-frequency features; and a BiLSTM network taking regularized meteorological data as input, fusing forward and backward LSTMs to extract the spatiotemporal dependency features of meteorological data, and outputting fixed-dimensional meteorological features through linear projection.
[0021] S4 specifically refers to the fusion process of the multi-head attention mechanism: high-frequency features, low-frequency features, and meteorological features are projected into Query, Key, and Value vectors, respectively. The associated weights are calculated in parallel by multiple independent attention heads. Then, the fused features of each attention head are concatenated and subjected to linear transformation to obtain a unified feature representation that matches the input dimension of the EQN network.
[0022] In S4, the EQN network includes a shared feature extraction layer, a quantile prediction head, and an evidence prediction head connected in sequence. The shared feature extraction layer performs deep feature extraction on the unified feature representation, the quantile prediction head outputs the multi-quantile prediction results of photovoltaic power generation, and the evidence prediction head outputs the evidence value corresponding to each quantile, which is used to quantify the prediction confidence.
[0023] In S5, the improved EQN loss function includes three terms: quantile loss, evidence loss, and width penalty loss, specifically: Quantile loss, expressed as:
[0024] in, For quantile loss, For batch size, The number of quantiles, For the first The true value of photovoltaic power for each sample. For the first The first sample Quantile predicted values, For the first quantile level, For indicator functions; The loss of evidence is:
[0025] in, For the loss of evidence, For the first In the nth sample The evidence value of each quantile, Batch size; The width penalty loss is:
[0026] in, For width loss, The weights for each confidence interval are... The maximum width threshold for each confidence interval. The normalization constant is For the first The first sample The width of each confidence interval.
[0027] The formula for calculating the total loss of the improved EQN loss function is as follows:
[0028] in, Let be the total loss function of the EQN network. The weighting coefficients for quantile prediction accuracy are... For evidence regularization weighting coefficients, This is the width penalty coefficient.
[0029] The beneficial effects of this invention are: This invention proposes a photovoltaic ultra-short-term probabilistic prediction method (DRDPN) based on data decomposition and reconstruction. The proposed ACEEMDAN algorithm achieves fixed two-component decomposition of photovoltaic power signals through four stages: decomposition preprocessing, local decomposition, pseudo-mode merging, and IMF binary aggregation. This solves the problems of mode mixing, inconsistent IMF numbers, and frequency redundancy in the traditional CEEMDAN algorithm, reducing computational complexity and improving the model's cross-dataset compatibility and computational efficiency. A CNN-iTransformer-BiLSTM heterogeneous neural network is used for multi-scale feature extraction, matching a dedicated network architecture to the feature characteristics of different components. This fully exploits the high-frequency mutation and low-frequency fluctuation characteristics of photovoltaic power data, as well as the spatiotemporal dependence of meteorological data, significantly improving the point prediction accuracy of photovoltaic power generation. The improved EQN loss function introduces a width penalty term, effectively reducing the interval width while ensuring prediction interval coverage, generating a narrow and well-calibrated prediction interval, enhancing the information value of uncertainty quantification, and solving the problem of overly conservative prediction results in traditional loss functions. The overall DRDPN model performs well in deterministic evaluation metrics (normalized mean absolute error nMAE, normalized root mean square error nRMSE, coefficient of determination R). 2 The method demonstrates excellent performance on both the continuous hierarchical probability score (CRPS), average coverage error (ACE), and Winkler score (WS). It maintains stable prediction performance across different datasets and exhibits strong cross-dataset stability and generalization ability. The method is suitable for ultra-short-term photovoltaic probability prediction at the 5-minute scale, and can fully explore the inherent spatiotemporal dependencies between photovoltaic data and meteorological data, providing strong support for grid connection of distributed solar power generation. It can be directly applied to grid dispatch and operation management of photovoltaic power generation systems. Attached Figure Description
[0030] Figure 1 This is a flowchart illustrating the prediction method of the present invention; Figure 2 This is an algorithm block diagram of ACEEMDAN in the prediction method of this invention; Figure 3 This is a schematic diagram showing the decomposition results of photovoltaic power generation data using the traditional CEEMDAN algorithm. Figure 4 This is a schematic diagram showing the decomposition results of photovoltaic power generation data using the ACEEMDAN algorithm of this invention. Figure 5 A comparison chart of prediction results for the power generation of a photovoltaic station using different methods; Figure 6 A radar chart comparing prediction results using different methods on different datasets. Detailed Implementation
[0031] The overall implementation conditions for the following embodiments are as follows: the prediction scale is 5 minutes; the model hardware environment used is: CPU Intel Core i9-13900K, GPU NVIDIA GeForce 4080 Super, memory 64GB; the software environment is: Python 3.8.20; the basic model training parameters are: batch size 128, number of iterations 200, learning rate 0.0001, sliding window radius w=5, noise figure β=2, number of white noise groups N=100, orthogonality threshold 3. Figure 1 As shown, after collecting photovoltaic power generation data, the data is first cleaned, and then ACEEMDAN is used to decompose the data into a low-frequency signal and a high-frequency signal, which are then regularized together with the meteorological data. After regularization, the high-frequency data is fed into a convolutional neural network, the low-frequency data into an iTransformer network, and the meteorological data into a bidirectional long short-term memory network for feature extraction. Then, multi-head attention is used for feature fusion, and finally, an evidence quantile network is used for prediction.
[0032] Example 1 The photovoltaic ultra-short-term probability prediction method disclosed in this embodiment, based on data decomposition and reconstruction, includes the following steps: S1: Real-time monitoring data of photovoltaic power is collected from the photovoltaic power station, and meteorological data such as irradiance, ambient temperature, relative humidity, wind speed, wind direction, and atmospheric pressure are collected from the supporting meteorological station. The sampling interval of all data is consistent with the prediction scale, which is 5 minutes. Linear interpolation is used to complete the missing values in the photovoltaic power data and meteorological data, and obvious outliers are removed to ensure the integrity and validity of the data, resulting in the preprocessed photovoltaic power sequence. and meteorological data sequences; S2: Using the ACEEMDAN algorithm to process photovoltaic power sequences Decomposed into high-frequency components and low frequency components The ACEEMDAN algorithm is divided into four stages (decomposition preprocessing, local decomposition, pseudo-pattern merging, and IMF binary aggregation) to eliminate the influence of dimensionality. S3: CNN, iTransformer, and BiLSTM are used to extract the corresponding features of high-frequency, low-frequency, and meteorological data respectively. S4: Features are fused through a multi-head attention mechanism to obtain a unified feature representation, which is then input into the EQN network to output the probability prediction result. S5: The improved EQN loss function with a width penalty term is used to train the model until convergence.
[0033] Example 2 This embodiment, based on Embodiment 1, provides a detailed explanation of the four stages in S2: decomposition preprocessing, local decomposition, pseudo-pattern merging, and IMF binary aggregation. The decomposition preprocessing is specifically as follows: Calculate photovoltaic power sequence At the point of time Local standard deviation Characterizing the spatiotemporal fluctuations of the signal, it serves as a scaling factor for adaptive noise modulation:
[0034] in, For a sliding window of radius w, The number of data points within the window. For in-window indexing, This represents the local average value of the data within the window. Generate N=100 groups that conform to the standard normal distribution Standardized white noise The EMD algorithm is used to pre-decompose white noise to obtain the intrinsic mode function components of the noise. This prepares for subsequent noise modulation.
[0035] Local decomposition specifically refers to: like Figure 2 As shown, ACEEMDAN is divided into four stages. Stage I calculates the local standard deviation of photovoltaic power generation data and generates and decomposes noise accordingly. Stage II uses the calculation results of Stage I to perform EMD decomposition. Stage III performs iterative cyclic merging based on the correlation between the decomposition results to eliminate pseudo-modes. Stage IV then merges the merged results of Stage III into a low-frequency component and a high-frequency component based on the frequency characteristics of the merged results of Stage III.
[0036] 1. Initialize residuals Decomposition level k=0, maximum decomposition level ; 2. When If the iteration termination condition is not met, perform the following operation: increment k by 1, starting from the noise intrinsic mode function components. Extract the noise from the k-th layer; Adaptive noise modulation is performed, dynamically adjusting the noise amplitude based on the local fluctuation characteristics of the signal. This suppresses pseudo-mode decomposition during the signal stability period and ensures decomposition fidelity during periods of high fluctuation.
[0037] in, After modulation Moment Decomposition stage noise, Indicates the first Each decomposition stage Time residuals for Local annotation difference at time, Noise figure for The first moment of white noise One EMD component, Represents the current calculation of the number Group noise; Signal modulated by noise EMD decomposition is performed to extract the first intrinsic mode function. Ensemble averaging is then applied to reduce mode aliasing, resulting in the... One intrinsic modulus function:
[0038] in, This represents the EMD decomposition process. For the first The intrinsic modulus functions obtained from the decomposition. Representing collective ownership Group noise; Update the residuals and repeat the above steps until the iteration termination condition is met, to obtain... Each intrinsic modulus function and monotonic residual:
[0039] in, Representing the The residuals obtained from the second decomposition To The first time the photovoltaic power generation data at time 1 The result of the decomposition.
[0040] The iteration termination conditions are: the number of residual extreme points is less than 3, the residual shows a monotonically increasing / decreasing trend, the residual energy is less than 1% of the total energy of the original signal, or the residual standard deviation is less than 1%. When any of the conditions is met, the iteration stops, and K eigenmode functions are obtained. and 1 monotonic residual .
[0041] To address the decomposed mode aliasing problem caused by the non-stationarity and nonlinearity of photovoltaic power signals, a pseudo-mode pair with frequency redundancy is detected and merged using an orthogonality index. The pseudo-mode merging process specifically involves: Calculate the orthogonality index of adjacent intrinsic moduli functions Characterizing the linear correlation between adjacent components:
[0042] Among them, the operator The dot product of vectors, operator Represents the Euclidean norm, operator Represents absolute value operation. Representative mode The orthogonality index between them; If adjacent intrinsic moduli functions If the adjacent intrinsic mode functions are determined to be a pseudo-mode pair, then their coefficient of determination is... There is a strong linear correlation. Detected pseudo-pattern pairs are merged, and frequency redundancy components are integrated.
[0043] in, For the index set of pseudo-pattern pairs, For the first index within the set One IMF; Repeat the above steps for all adjacent intrinsic moduli. , to obtain the optimized Each independent intrinsic mode function preserves the key time patterns of the signal.
[0044] The specific IMF binary aggregation is as follows: To address the issue of inconsistent eigenmode function numbers in traditional decomposition methods, a frequency-domain binary aggregation method is employed to achieve dimensionality invariance, outputting fixed high-frequency and low-frequency components. For each optimized eigenmode function, its power spectral density is calculated using Fast Fourier Transform (FFT) to obtain the frequency center. (Energy-weighted spectral distribution) and dominant frequency (Energy Concentration Peak); The power spectral density of each optimized intrinsic mode function is calculated using Fast Fourier Transform to obtain the frequency center. (Energy-weighted spectral distribution) and dominant frequency (Energy concentration peak), adaptively select representative frequency The overall frequency characteristics of the components are reflected:
[0045] in, For the first The representative frequency of each IMF For the first The main frequencies of the IMF For the first The center frequency of each IMF; Predicting specific frequency thresholds on a 5-minute scale The intrinsic mode functions are divided into high-frequency groups ( ) and low-frequency group ( The criteria for judgment are:
[0046] in, For high frequency groups, For low frequency groups, For the i-th IMF, High-frequency threshold; Summing the components of the two groups separately yields the final high-frequency component. and low frequency components To achieve two-component decomposition of photovoltaic power signals:
[0047] in, For high-frequency components, Low-frequency components, For the i-th IMF, It is the residual vector; RobustScaler was used to regularize the high-frequency components, low-frequency components, and meteorological data sequences. Meanwhile, RobustScaler was used to train and standardize the target variable (photovoltaic power) separately.
[0048] like Figure 3 As shown in the figure, the IMF represents the original decomposition result. It can be seen that without the detection and merging of pseudo-modals, it should contain 11 decomposition results. Figure Four As shown, after the detection and merging of pseudo-modals, the output of ACEEMDAN contains only two components. Figure 3 and Figure 4 It can be concluded that the ACEEMDAN algorithm can decompose a complete photovoltaic power generation data into a high-frequency component and a low-frequency component.
[0049] Example 3 Based on Example 2, this example provides a detailed explanation of S3: To address the different characteristics of high-frequency components, low-frequency components, and meteorological data, heterogeneous neural networks such as CNN, iTransformer, and BiLSTM are employed for feature extraction, achieving accurate capture of multi-scale and multi-type features. The input and implementation details for each network are as follows: 1. CNN network: The input is a regularized high-frequency component related sequence (input sequence length 12), with 3 convolutional layers (convolutional kernel size 3) and 3 pooling layers (pooling kernel size 2), and the output feature dimension is 64. The convolutional layers capture the local abrupt features of the high-frequency components (such as sudden power changes caused by cloud occlusion), and the pooling layers reduce the dimension and retain key information. Finally, the output is a high-frequency feature vector with a fixed dimension of 64. 2. iTransformer network: The input is the regularized low-frequency component. (Number of variables 1, backtracking length 48), embedding dimension 128, set 4 attention layers (6 heads), embed the low-frequency component variables as discrete tokens, use the attention mechanism to capture the correlation between variables in the variable dimension, extract the long-term trend features of the low-frequency components (such as the daytime variation trend of photovoltaic power) in the time dimension, and after output compression processing, obtain a low-frequency feature vector with a fixed dimension of 64. 3. BiLSTM network: In practice, a bidirectional LSTM (3 layers, 128 hidden layer dimensions) is used. The input is a regularized meteorological data sequence (the input dimension is the original input dimension minus 1). The forward LSTM and backward LSTM are fused to extract the spatiotemporal dependency features of the meteorological data. The output is processed by LSTMFeatureProcessor (including attention pooling, linear projection and dropout=1) to aggregate the output of multiple time steps into a meteorological feature vector with a fixed dimension of 64.
[0050] Example 4 Based on Example 3, this example provides a detailed explanation of S4: High-frequency features, low-frequency features, and meteorological feature vectors output from CNN, iTransformer, and BiLSTM are input into a multi-head attention mechanism to achieve deep fusion of heterogeneous features: 1. Feature projection: Project the three feature vectors into three vectors, Query, Key, and Value, respectively, with each projection dimension being 256; 2. Parallel computation of multiple attention heads: Set up 8 independent attention heads, each of which calculates the association weight between the query and the key. After normalization by the softmax function, the weight is multiplied by the value to obtain the fused feature of each attention head. 3. Feature concatenation and linear transformation: The fused features of the 8 attention heads are concatenated into a high-dimensional feature vector, and a linear transformation is performed through a fully connected layer to obtain a unified feature representation with a dimension of 64 (matching the dimension of the network input EQN). This feature integrates the high-frequency mutation, low-frequency trend and spatiotemporal dependence information of photovoltaic power and meteorological data.
[0051] The fused unified feature representation is input into the EQN network, which combines evidence-based deep learning and quantile regression to achieve probability quantile prediction and uncertainty quantification of photovoltaic power generation. The network consists of three core components, and the key implementation points are as follows: 1. Shared feature extraction layer: Composed of fully connected layers adapted to the input dimension, it further abstracts and extracts the unified feature representation with dimension 64 to obtain more representative deep features, adapting to the input requirements of the EQN network; 2. Quantile Prediction Header: Consists of a single fully connected layer, outputting multi-quantile prediction results. In this embodiment, the number of quantiles is set to 11 according to the network code, corresponding to different confidence intervals, outputting quantile prediction values of photovoltaic power generation. ; 3. Evidence Prediction Head: Consists of a single fully connected layer, outputting the evidence value corresponding to each quantile. It is used to quantify the confidence level of predictions and effectively distinguish between cognitive uncertainty and accidental uncertainty.
[0052] Example 5 Based on Example 4, this example provides a detailed explanation of the improved EQN loss function in S5: This embodiment employs an improved EQN loss function that incorporates a width penalty term to train the overall DRDPN model, addressing the issues of excessively wide prediction intervals and imbalance between coverage and accuracy inherent in traditional loss functions. The weighting coefficients are set accordingly. , , The calculation formulas for each loss item and the total loss, as well as the key points for implementation, are as follows: 1. Quantile loss guides the model to converge to the true quantile. It employs an asymmetric penalty function, expressed as:
[0053] in, For quantile loss, For batch size, The number of quantiles, For the first The true value of photovoltaic power for each sample. For the first The first sample Quantile predicted values, For the first quantile level, For indicator functions; 2. Evidence Loss: Regularize the evidence values to prevent the model from becoming overconfident.
[0054] in, For the loss of evidence, For the first In the nth sample The evidence value of each quantile, Batch size; 3. Width Penalty Loss: Penalizes prediction intervals that exceed the maximum width threshold, guiding the model to generate compact and well-calibrated prediction intervals.
[0055] in, For width loss, The weights for each confidence interval are (all are set to 1 in this embodiment). The maximum width threshold for each confidence interval (set according to the actual photovoltaic power fluctuation range). The normalization constant is For the first The first sample The width of each confidence interval.
[0056] Example 6 Based on Example 5, this example provides a detailed explanation of the total loss of the improved EQN loss function: The total loss function combines quantile loss, evidence loss, and width penalty loss, and is calculated using the following formula:
[0057] in, Let be the total loss function of the EQN network. The weighting coefficients for quantile prediction accuracy are... For evidence regularization weighting coefficients, This is the width penalty coefficient.
[0058] The Adam optimizer is used to calculate the model's loss value according to the total loss function mentioned above. All parameters of the model are updated through backpropagation. The model is iteratively trained for 2000 rounds until the loss value converges and tends to stabilize, thus completing the construction of the DRDPN model. This model can be used to achieve ultra-short-term photovoltaic probability prediction on a 5-minute scale.
[0059] To verify the effectiveness of the method of this invention, actual operating data (photovoltaic power + meteorological data) from a photovoltaic station at City University of Hong Kong in 2022 was used as the experimental dataset. This method was compared with six mainstream photovoltaic prediction methods: LSTM-QR, XGBOOST, TFT, BDNN, and MDN. A comprehensive evaluation was conducted using deterministic evaluation metrics (nMAE, nRMSE, R²) and probabilistic evaluation metrics (CRPS, ACE, WS). The experimental results are as follows: 1. Prediction results for a single dataset ( Figure 4 ):Figure 4 The results of the DRDPN method and the baseline method on dataset #1 are presented. The median value predicted by the DRDPN method closely matches the actual photovoltaic power value, outperforming all other models, while its prediction interval is also significantly narrower than other methods. This advantage mainly stems from two core mechanisms: First, the method decomposes photovoltaic power into specific frequency components, using a CNN network for high-frequency components and an iTransformer network for low-frequency components for feature extraction. This strategy effectively captures multi-scale time patterns and improves point prediction accuracy. Second, the interval width penalty term introduced in the EQN network effectively reduces the interval width while ensuring prediction interval coverage, achieving more informative uncertainty quantification. In contrast, the baseline method directly processes the raw photovoltaic data without frequency decomposition or explicit interval width constraints, resulting in poor point prediction performance and overly broad confidence intervals. As shown in Table 1, DRDPN achieves the best performance on dataset #1 with nMAE of 0.0507 and nRMSE of 0.0906, R² of 0.9918, CRPS of 0.0446 and WS of 1.0671, further validating its superiority in single-dataset prediction.
[0060] 2. Cross-dataset prediction results ( Figure 5 , Figure 6 ): Figure 5 The performance of each method is demonstrated on four different photovoltaic (PV) station datasets (datasets #1-#4). DRDPN significantly outperforms other models because its predicted median closely matches the observed values. Furthermore, its prediction interval is significantly narrower than that of competing methods. These improvements are driven by two main mechanisms. First, the method decomposes PV power into frequency-specific components, applying a convolutional neural network to the high-frequency components and an iTransformer to the low-frequency components. This strategy effectively captures multi-scale temporal patterns and improves the accuracy of point predictions. Second, the interval width penalty in the EQN network contributes to the narrowing of the prediction interval, maintaining adequate coverage while reducing interval width. These properties provide more informative uncertainty quantification. Figure 5 It can be concluded that DRDPN has achieved good results in both accurate point prediction and probability prediction.
[0061] Figure 6 It integrates all evaluation metrics (deterministic metrics: nMAE, nRMSE, R²; probabilistic metrics: CRPS, ACE, WS) for easy and intuitive comparison. Figure 6As can be seen, DRDPN maintains the most compact polygonal structure across all datasets, demonstrating excellent stability and achieving a balanced performance in both deterministic and probabilistic metrics. Specifically, XGBoost exhibits high accuracy in deterministic predictions, but its CRPS and WS values increase significantly on datasets #2 and #3, revealing limitations in its probabilistic prediction capabilities. LSTM-QR, BDNN, and MDN show greater variability in their metrics, with the most significant increases in CRPS, WS, and ACE, reflecting their poor generalization ability across different dataset parameters. TFT maintains moderate stability and a relatively balanced distribution of metrics, but its overall performance is inferior to DRDPN. Figure 6 It can be concluded that DRDPN consistently maintains the most compact and stable polygonal structure across all four datasets, confirming its superior robustness and generalization ability, while other benchmark methods exhibit higher sensitivity to specific dataset features. Further validation using the data in Table 1 shows that DRDPN achieves the best R² across all datasets, and also reaches the best CRPS and WS values on most datasets, fully demonstrating its outstanding cross-dataset stability and generalization ability.
[0062] To more comprehensively and accurately verify the superiority of the DRDPN method, Table 1 summarizes the performance of DRDPN and five benchmark methods (LSTM-QR, TFT, XGBoost, BDNN, MDN) on four photovoltaic equipment datasets (datasets #1-#4), covering deterministic evaluation metrics (nMAE, nRMSE, R²) and probabilistic evaluation metrics (CRPS, ACE, WS).
[0063] Table 1
[0064] As shown in Table 1, DRDPN achieves the best R² across all datasets and also performs best in CRPS and WS metrics on most datasets, fully demonstrating its superior generalization ability and cross-dataset stability. In contrast, each benchmark method has significant shortcomings: LSTM-QR, BDNN, and MDN have poor overall performance and high volatility; XGBoost performs well in deterministic metrics on some datasets, but its probabilistic prediction metrics fluctuate wildly; TFT performs moderately and does not show a clear advantage. In conclusion, DRDPN exhibits the best overall performance in photovoltaic ultra-short-term probabilistic prediction and is better suited to meet practical engineering needs.
[0065] The experimental results above demonstrate that the method proposed in this invention achieves higher prediction accuracy, better prediction interval calibration, stronger cross-dataset stability, and high computational efficiency in ultra-short-term probability prediction of photovoltaics. It can fully explore the spatiotemporal dependence of photovoltaic data and meteorological data, and provide strong support for grid connection of distributed and fluctuating solar power generation.
Claims
1. A photovoltaic ultra-short-term probability prediction method based on data decomposition and reconstruction, characterized in that, Includes the following steps: S1: Collect photovoltaic power and meteorological data, and use linear interpolation to fill in missing values to complete data preprocessing; S2: Using the ACEEMDAN algorithm, the photovoltaic power data is decomposed into fixed high-frequency and low-frequency dual components through four stages: decomposition preprocessing, local decomposition, pseudo-mode merging, and IMF binary aggregation. The dual components are then regularized with meteorological data. S3: CNN, iTransformer, and BiLSTM are used to extract the corresponding features of high-frequency, low-frequency, and meteorological data, respectively. S4: The unified feature representation is obtained by fusing features through a multi-head attention mechanism and input into the EQN network to output the probability prediction result; S5: Train the EQN network until convergence using an improved EQN loss function with a width penalty term to obtain the prediction model.
2. The photovoltaic ultra-short-term probability prediction method based on data decomposition and reconstruction according to claim 1, characterized in that, The specific decomposition preprocessing in S2 is as follows: Calculate photovoltaic power data at a given time point Local standard deviation : in, For a sliding window of radius w, The number of data points within the window. For in-window indexing, This represents the local average value of the data within the window. N=100 sets of standardized white noise conforming to the standard normal distribution are generated. The white noise is pre-decomposed using the EMD algorithm to obtain the noise intrinsic mode function components.
3. The photovoltaic ultra-short-term probability prediction method based on data decomposition and reconstruction according to claim 2, characterized in that, The local decomposition in S2 is specifically as follows: Perform adaptive noise modulation: in, After modulation Moment Decomposition stage noise, Indicates the first Each decomposition stage Time residuals for Local annotation difference at time, Noise figure for The first moment of white noise One EMD component, Represents the current calculation of the number Group noise; The noise-modulated signal is decomposed using EMD and then ensemble-averaged to obtain the first... One intrinsic modulus function: in, This represents the EMD decomposition process. For the first The intrinsic modulus functions obtained from the decomposition. Representing collective ownership Group noise; Update the residuals and repeat the above steps until the iteration termination condition is met, to obtain... The iteration terminates when the number of residual extreme points is less than 3, the residual shows a monotonically increasing / decreasing trend, the residual energy is less than 1% of the total energy of the original signal, or the residual standard deviation is less than 1%. Any of the following conditions: in, Representing the The residuals obtained from the second decomposition To The first time the photovoltaic power generation data at time 1 The result of the decomposition.
4. The photovoltaic ultra-short-term probability prediction method based on data decomposition and reconstruction according to claim 3, characterized in that, The pseudo-pattern merging in S2 is specifically as follows: Calculate the orthogonality index of adjacent intrinsic moduli: Among them, the operator The dot product of vectors, operator Represents the Euclidean norm, operator Represents absolute value operation. Representative mode The orthogonality index between them; If adjacent intrinsic moduli functions If a pair is identified as a pseudo-pattern pair, merge the pseudo-pattern pairs: in, For the index set of pseudo-pattern pairs, For the first index within the set One IMF; Repeat the above steps for all adjacent intrinsic moduli. The optimized independent intrinsic modulus functions are obtained.
5. The photovoltaic ultra-short-term probability prediction method based on data decomposition and reconstruction according to claim 4, characterized in that, The specific details of IMF binary aggregation in S2 are as follows: The power spectral density of each optimized intrinsic mode function is calculated using Fast Fourier Transform to determine the representative frequency. in, For the first The representative frequency of each IMF For the first The main frequencies of the IMF For the first The center frequency of each IMF; Based on the prediction of specific frequency thresholds on a 5-minute scale, the intrinsic mode functions are divided into high-frequency and low-frequency groups: in, For high frequency groups, For low frequency groups, For the i-th IMF, High-frequency threshold; Summing them separately yields the final high-frequency and low-frequency components: in, For high-frequency components, Low-frequency components, For the i-th IMF, It is the residual vector; RobustScaler was used to perform regularization on high-frequency components, low-frequency components, and meteorological data. RobustScaler was also used to standardize the photovoltaic power target variable separately.
6. The photovoltaic ultra-short-term probability prediction method based on data decomposition and reconstruction according to claim 5, characterized in that: S3 specifically refers to: the input of the CNN network is the regularized high-frequency component, the local mutation features of the high-frequency component are captured through the convolutional layer, and the high-frequency features with fixed dimensions are output after dimensionality reduction by the pooling layer; the input of the iTransformer network is the regularized low-frequency component, the time series variables are embedded as discrete tokens, the correlation and long-term trend features between variables are captured through the attention mechanism, and the low-frequency features with fixed dimensions are output. The input to the BiLSTM network is regularized meteorological data. The spatiotemporal dependency features of the meteorological data are extracted by fusing forward and backward LSTMs, and the fixed-dimensional meteorological features are output through linear projection.
7. The photovoltaic ultra-short-term probability prediction method based on data decomposition and reconstruction according to claim 6, characterized in that: S4 specifically refers to the fusion process of the multi-head attention mechanism as follows: high-frequency features, low-frequency features, and meteorological features are projected into Query, Key, and Value vectors, respectively. The associated weights are calculated in parallel by multiple independent attention heads. Then, the fused features of each attention head are concatenated and subjected to linear transformation to obtain a unified feature representation that matches the input dimension of the EQN network.
8. The photovoltaic ultra-short-term probability prediction method based on data decomposition and reconstruction according to claim 7, characterized in that: In S4, the EQN network includes a shared feature extraction layer, a quantile prediction head, and an evidence prediction head connected in sequence. The shared feature extraction layer performs deep feature extraction on the unified feature representation. The quantile prediction head outputs the multi-quantile prediction results of photovoltaic power generation. The evidence prediction head outputs the evidence value corresponding to each quantile, which is used to quantify the prediction confidence.
9. The photovoltaic ultra-short-term probability prediction method based on data decomposition and reconstruction according to claim 8, characterized in that: In S5, the improved EQN loss function includes three terms: quantile loss, evidence loss, and width penalty loss, specifically: Quantile loss, expressed as: in, For quantile loss, For batch size, The number of quantiles, For the first The true value of photovoltaic power for each sample. For the first The first sample Quantile predicted values, For the first quantile level, For indicator functions; The loss of evidence is: in, For the loss of evidence, For the first In the nth sample The evidence value of each quantile, Batch size; The width penalty loss is: in, For width loss, The weights for each confidence interval are... The maximum width threshold for each confidence interval. The normalization constant is For the first The first sample The width of each confidence interval.
10. The photovoltaic ultra-short-term probability prediction method based on data decomposition and reconstruction according to claim 9, characterized in that: The total loss calculation formula for the improved EQN loss function is as follows: in, Let be the total loss function of the EQN network. The weighting coefficients for quantile prediction accuracy are... For evidence regularization weighting coefficients, This is the width penalty coefficient.