Multi-dimensional feature decoupling and time sequence attention fused photovoltaic power prediction method
By employing a multidimensional feature decoupling and temporal attention fusion method, the problems of insufficient data preprocessing and feature fusion in photovoltaic power prediction are solved, achieving high-precision and robust photovoltaic power prediction, especially in predictive capabilities under different meteorological conditions.
Patent Information
- Application Number
- CN202511668003.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-14
- Publication Date
- 2026-02-24
AI Technical Summary
Existing photovoltaic power prediction methods have shortcomings in data preprocessing, feature extraction, and feature fusion, which limits prediction accuracy and fails to effectively utilize the periodic information of photovoltaic power generation.
We employ a multidimensional feature decoupling and temporal attention fusion approach. We preprocess the data through variational mode decomposition and empirical wavelet transform, construct a dual-stream parallel deep learning network, utilize cross-attention mechanism for feature fusion, and combine static physical features for prediction.
It significantly improves the accuracy and robustness of photovoltaic power prediction, especially the prediction capability under different meteorological conditions, reduces the difficulty of model learning, and enhances physical interpretability.
Smart Images
Figure CN121566421A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of photovoltaic power prediction technology, specifically a photovoltaic power prediction method based on multidimensional feature decoupling and temporal attention fusion. Background Technology
[0002] Driven by the global energy transition and the goal of "carbon neutrality," renewable energy sources, represented by photovoltaic power generation, are developing at an unprecedented pace. However, the output power of photovoltaic power generation systems is strongly affected by various meteorological factors such as sunlight intensity, temperature, and humidity, exhibiting significant intermittency, volatility, and non-stationarity. This inherent uncertainty poses a significant challenge to the safe and stable operation of power systems, grid dispatch, and energy management.
[0003] Therefore, accurate prediction of the output power of photovoltaic power plants is a key technology for promoting large-scale grid connection and consumption of photovoltaic energy and ensuring the safe and economical operation of the power grid.
[0004] Current prediction methods are mainly divided into physical models, traditional statistical models, and artificial intelligence models. Among them, artificial intelligence models, represented by deep learning, have become a research hotspot due to their powerful nonlinear fitting capabilities. However, existing technologies still have the following shortcomings:
[0005] Insufficient data preprocessing: Photovoltaic power sequences have strong non-stationary and noise characteristics. When a single prediction model directly processes the raw data, it is difficult to effectively learn its inherent patterns, resulting in limited prediction accuracy.
[0006] Insufficient feature extraction: Existing models typically input historical power data (target sequence) and meteorological data (covariates) into the model indiscriminately, failing to use specialized network structures for feature extraction based on the different types of data, resulting in insufficient feature representation capabilities.
[0007] The feature fusion mechanism is simple: most models use a simple feature concatenation method to fuse multi-source information, which cannot dynamically capture the complex, time-varying interaction between historical power state and external meteorological factors, thus limiting the upper limit of the model's prediction.
[0008] Inadequate utilization of periodic information: Photovoltaic power generation has obvious daily and annual cycle characteristics. Most existing models rely on the network itself to learn implicitly from the data, failing to explicitly provide this prior knowledge to the model, resulting in low learning efficiency. Summary of the Invention
[0009] Based on this, the purpose of this invention is to provide a photovoltaic power prediction method that combines multidimensional feature decoupling and temporal attention fusion, so as to solve the technical problems in the background art mentioned above.
[0010] To achieve the above objectives, the present invention provides the following technical solution: a photovoltaic power prediction method based on multidimensional feature decoupling and temporal attention fusion, comprising the following steps: Step 1: Data Acquisition and Generation of Meteorological Variable Series; Step 2: Target power sequence decomposition and denoising; Step 3: Dual-stream parallel feature extraction: Construct a dual-stream parallel deep learning network structure to perform specialized feature extraction on the target subsequence group and the meteorological variable sequence group; Step 4: Cross-Attention Feature Fusion: Design a cross-attention fusion mechanism to intelligently fuse features extracted by the two-stream network; Step 5: Integration of physical properties; Step Six: Gated Nonlinear Prediction; Step 7: Reconstruct the prediction results.
[0011] Furthermore, in step one, historical power data of the target power plant is obtained as the target sequence, and various meteorological data from the same period are obtained as meteorological variables; based on the timestamp information of the historical data, time-based periodic features are created and merged with the original meteorological data to form a meteorological variable sequence group.
[0012] Furthermore, in step two, the variational mode decomposition (VMD) algorithm is used to decompose the normalized target power sequence into several variational mode components (VMF) and a residual term. Subsequently, empirical wavelet transform (EWT) is used to denoise the high-frequency mode components with the highest frequency and strongest randomness after decomposition in order to reduce noise interference, and finally a relatively stable target subsequence group is obtained.
[0013] Furthermore, the specific extraction steps in step three are as follows: (1) Target power sequence stream: Input the target subsequence group processed in step two into the iTransformer model, and use its powerful modeling capability for multivariate time series to extract the deep long-term time dependence features inside each power component. (2) Meteorological variable flow: The meteorological variable sequence group in step one is input into the Long Short-Term Memory (LSTM) network, and its gated recurrent mechanism is used to extract the complex dynamic evolution characteristics between multidimensional external factors.
[0014] Furthermore, the intelligent fusion in step four is as follows: the features output by the target sequence stream (iTransformer) are used as the query of the attention mechanism, and the features output by the meteorological variable stream (LSTM) are used as the key and value; by calculating the similarity between the query and the key, the model can dynamically assign weights to different meteorological variable features according to the internal state of the current power sequence, and generate a feature vector that deeply integrates internal and external causal relationships.
[0015] Furthermore, in step five, the static physical parameters of the photovoltaic power station, including the module model, installation tilt angle, and annual degradation rate, are encoded to form a static physical feature vector. Then, this static physical feature vector is concatenated with the dynamic feature vector generated in step four to form a comprehensive feature vector that combines time-varying dynamic and static physical constraints.
[0016] Furthermore, in step six, the integrated feature vector fused in step five is input into a gated linear unit (GLU) prediction head. The dynamic gating mechanism of the GLU is used to perform nonlinear mapping on the fused features in order to predict the output value of each target subsequence at future time.
[0017] Furthermore, in step seven, the predicted values of all target subsequences predicted in step six are linearly summed, reconstructed, and inversely normalized to obtain the final predicted result of the total output power of the photovoltaic power station.
[0018] In summary, the present invention has the following main advantages: based on signal decomposition and dual-stream attention fusion, it aims to significantly improve the accuracy and robustness of ultra-short-term photovoltaic power prediction through refined data preprocessing, professional dual-stream feature extraction, and intelligent attention fusion mechanism. Attached Figure Description
[0019] Figure 1 This is a diagram illustrating the method steps of the present invention. Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0021] The embodiments of the present invention will now be described.
[0022] like Figure 1 As shown, a photovoltaic power prediction method combining multidimensional feature decoupling and temporal attention fusion includes the following steps: Step 1: Data Acquisition and Generation of Meteorological Variable Series; Step 2: Target power sequence decomposition and denoising; Step 3: Dual-stream parallel feature extraction: Construct a dual-stream parallel deep learning network structure to perform specialized feature extraction on the target subsequence group and the meteorological variable sequence group; Step 4: Cross-Attention Feature Fusion: Design a cross-attention fusion mechanism to intelligently fuse features extracted by the two-stream network; Step 5: Integration of physical properties; Step Six: Gated Nonlinear Prediction; Step 7: Reconstruct the prediction results.
[0023] like Figure 1 As shown, in step one, historical power data of the target power station (taking Ningbo photovoltaic power station as an example) is obtained as the target sequence, along with various meteorological data of the same period (such as solar irradiance, temperature, humidity, etc.) as meteorological variables. Based on the timestamp information of the historical data, time-based periodic features are created and merged with the original meteorological data to form a meteorological variable sequence group. In this embodiment, by constructing periodic time features, the day-night and seasonal patterns of photovoltaic power generation are directly provided to the model as prior knowledge, which accelerates model convergence and improves the accuracy of prediction.
[0024] like Figure 1 As shown, in step two, the variational mode decomposition (VMD) algorithm is used to decompose the normalized target power sequence into several variational mode components (VMFs) and a residual term. Subsequently, empirical wavelet transform (EWT) is used to denoise the high-frequency mode components with the highest frequency and strongest randomness after decomposition to reduce noise interference, and finally a set of relatively stable target subsequences is obtained. In this embodiment, through the combined preprocessing of VMD-EWT, the complex original power sequence is decomposed into multiple simpler and more stable subsequences, and high-frequency noise is effectively filtered out, which significantly reduces the learning difficulty of the subsequent model and effectively reduces non-stationarity.
[0025] like Figure 1 As shown, the specific extraction steps in step three are as follows: (1) Target power sequence stream: Input the target subsequence group processed in step two into the iTransformer model, and use its powerful modeling capability for multivariate time series to extract the deep long-term time dependence features inside each power component. (2) Meteorological variable flow: The meteorological variable sequence group in step one is input into the Long Short-Term Memory (LSTM) network. Its gated loop mechanism is used to extract the complex dynamic evolution features between multidimensional external factors. In this embodiment, through the dual-stream architecture, the iTransformer is used to process the time pattern of the target power sequence, and the LSTM is used to process the dynamic influence of meteorological variables, realizing "dedicated model for dedicated use", and the feature extraction is more targeted and effective.
[0026] like Figure 1 As shown, the intelligent fusion in step four is as follows: the features output by the target sequence stream (iTransformer) are used as the query of the attention mechanism, and the features output by the meteorological variable stream (LSTM) are used as the key and value. By calculating the similarity between the query and the key, the model can dynamically assign weights to different meteorological variable features according to the internal state of the current power sequence, and generate a feature vector that deeply integrates internal and external causal relationships.
[0027] like Figure 1 As shown, in step five, the static physical parameters of the photovoltaic power station, including the module type (e.g., N-type bifacial monocrystalline silicon), installation tilt angle (e.g., 5°), and annual degradation rate (e.g., 1% in the first year and 0.4% annually thereafter), are encoded to form a static physical feature vector. Then, this static physical feature vector is concatenated with the dynamic feature vector generated in step four to form a comprehensive feature vector that combines time-varying dynamics and static physical constraints. In this embodiment, the core cross-attention fusion mechanism replaces simple feature concatenation, which can intelligently and dynamically model the complex relationship between power fluctuations and key meteorological factors, significantly improving prediction accuracy. Taking the Ningbo photovoltaic power station as an example, by incorporating the static physical parameters such as the photovoltaic module type, installation tilt angle, and annual degradation rate of the Ningbo power station into the model, the model is no longer a purely data-driven "black box." This provides physical constraints for the model, which can adaptively consider the performance degradation caused by equipment aging, thereby significantly improving the accuracy and physical interpretability of long-term predictions.
[0028] like Figure 1 As shown, in step six, the integrated feature vector fused in step five is input into a gated linear unit (GLU) prediction head. The dynamic gating mechanism of the GLU is used to perform nonlinear mapping on the fused features in order to predict the output value of each target subsequence at future time.
[0029] like Figure 1As shown, in step seven, the predicted values of all target subsequences predicted in step six are linearly summed, reconstructed, and denormalized to obtain the final prediction result of the total output power of the photovoltaic power station. In this embodiment, a variety of advanced deep learning components are combined with refined training strategies (such as learning rate scheduling and early stopping mechanism) to make the overall prediction accuracy of the model high and the robustness to power fluctuations under different meteorological conditions stronger.
[0030] Beneficial effects: Effective reduction of non-stationarity: Through the combined preprocessing of VMD-EWT, the complex original power sequence is decomposed into multiple simpler and more stable subsequences, and high-frequency noise is effectively filtered out, which significantly reduces the learning difficulty of subsequent models.
[0031] Specialized feature processing is achieved by using a dual-stream architecture, utilizing iTransformer to process the time pattern of the target power sequence, and LSTM to process the dynamic influence of meteorological variables, thus achieving "specialized model for specific purpose" and making feature extraction more targeted and effective.
[0032] Dynamically capturing internal and external causal relationships: The core cross-attention fusion mechanism replaces simple feature splicing, which can intelligently and dynamically model the complex relationship between power fluctuations and key meteorological factors, significantly improving prediction accuracy.
[0033] Explicitly utilizing periodic patterns: By constructing periodic time features, the day-night and seasonal patterns of photovoltaic power generation are directly provided to the model as prior knowledge, which accelerates model convergence and improves prediction accuracy.
[0034] Enhancing long-term accuracy and physical consistency: By incorporating static physical parameters such as the photovoltaic panel module model, installation tilt angle, and annual degradation rate of the Ningbo power station into the model, the model is no longer a purely data-driven "black box." This provides the model with physical constraints, enabling it to adaptively consider performance degradation caused by equipment aging, thereby significantly improving the accuracy and physical interpretability of long-term predictions.
[0035] Strong predictive ability and robustness: It combines a variety of advanced deep learning components and is supplemented by refined training strategies (such as learning rate scheduling and early stopping mechanism), which makes the overall prediction accuracy of the model high and more robust to power fluctuations under different meteorological conditions.
[0036] Although embodiments of the present invention have been shown and described, these specific embodiments are merely explanations of the invention and are not intended to limit it. The specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. After reading this specification, those skilled in the art may make modifications, substitutions, and variations to the embodiments as needed without departing from the principles and spirit of the invention, but such modifications, substitutions, and variations are protected by patent law as long as they are within the scope of the claims of the present invention.
Claims
1. A photovoltaic power prediction method based on multidimensional feature decoupling and temporal attention fusion, characterized in that, Includes the following steps: Step 1: Data Acquisition and Generation of Meteorological Variable Series; Step 2: Target power sequence decomposition and denoising; Step 3: Dual-stream parallel feature extraction: Construct a dual-stream parallel deep learning network structure to perform specialized feature extraction on the target subsequence group and the meteorological variable sequence group; Step 4: Cross-Attention Feature Fusion: Design a cross-attention fusion mechanism to intelligently fuse features extracted by the two-stream network; Step 5: Integration of physical properties; Step Six: Gated Nonlinear Prediction; Step 7: Reconstruct the prediction results.
2. The photovoltaic power prediction method based on multidimensional feature decoupling and temporal attention fusion according to claim 1, characterized in that: In step one, historical power data of the target power plant is obtained as the target sequence, along with various meteorological data from the same period (such as solar irradiance, temperature, humidity, etc.) as meteorological variables. Based on the timestamp information of the historical data, time-based periodic features are created and merged with the original meteorological data to form a meteorological variable sequence group.
3. The photovoltaic power prediction method based on multidimensional feature decoupling and temporal attention fusion according to claim 1, characterized in that: In step two, the variational mode decomposition (VMD) algorithm is used to decompose the normalized target power sequence into several variational mode components (VMF) and a residual term. Subsequently, empirical wavelet transform (EWT) is used to denoise the high-frequency mode components with the highest frequency and the strongest randomness after decomposition in order to reduce noise interference, and finally a relatively stable target subsequence group is obtained.
4. The photovoltaic power prediction method based on multidimensional feature decoupling and temporal attention fusion according to claim 1, characterized in that: The specific extraction steps in step three are as follows: (1) Target power sequence stream: Input the target subsequence group processed in step two into the iTransformer model, and use its powerful modeling capability for multivariate time series to extract the deep long-term time dependence features inside each power component. (2) Meteorological variable flow: The meteorological variable sequence group in step one is input into the Long Short-Term Memory (LSTM) network, and its gated recurrent mechanism is used to extract the complex dynamic evolution characteristics between multidimensional external factors.
5. The photovoltaic power prediction method based on multidimensional feature decoupling and temporal attention fusion according to claim 1, characterized in that: The intelligent fusion in step four is as follows: the features output by the target sequence stream (iTransformer) are used as the query of the attention mechanism, and the features output by the meteorological variable stream (LSTM) are used as the key and value. By calculating the similarity between the query and the key, the model can dynamically assign weights to different meteorological variable features according to the internal state of the current power sequence, and generate a feature vector that deeply integrates internal and external causal relationships.
6. The photovoltaic power prediction method based on multidimensional feature decoupling and temporal attention fusion according to claim 1, characterized in that: In step five, the static physical parameters of the photovoltaic power station, including the module model, installation tilt angle, and annual degradation rate, are encoded to form a static physical feature vector. Then, this static physical feature vector is concatenated with the dynamic feature vector generated in step four to form a comprehensive feature vector that combines time-varying dynamic and static physical constraints.
7. The photovoltaic power prediction method based on multidimensional feature decoupling and temporal attention fusion according to claim 1, characterized in that: In step six, the integrated feature vector fused in step five is input into a gated linear unit (GLU) prediction head. The dynamic gating mechanism of the GLU is used to perform nonlinear mapping on the fused features in order to predict the output value of each target subsequence at future time.
8. The photovoltaic power prediction method based on multidimensional feature decoupling and temporal attention fusion according to claim 1, characterized in that: In step seven, the predicted values of all target subsequences predicted in step six are linearly summed, reconstructed, and inversely normalized to obtain the final predicted result of the total output power of the photovoltaic power station.