Photovoltaic power generation power prediction method, system, equipment and medium
The photovoltaic power prediction method, which employs adaptive mode decomposition, global optimization, and error compensation, addresses the issue of low accuracy in existing models under extreme weather conditions, achieving high-precision and stable predictions during periods of dramatic climate change.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LANZHOU JIAOTONG UNIV
- Filing Date
- 2026-01-26
- Publication Date
- 2026-05-08
AI Technical Summary
Existing photovoltaic power prediction models have low accuracy when extreme weather events occur alternately. They lack adaptive matching and global generalization capabilities for multiple weather physical mechanisms during climate transition periods and neglect prediction error correction, resulting in unstable model performance during drastic climate changes.
Adaptive mode decomposition technology is used to decompose photovoltaic power time series data, an enhanced feature set is constructed, and the hyperparameters of the neural network are optimized by the tornado optimization algorithm with Coriolis force. The prediction is then performed in combination with the Transformer architecture. Subsequently, the residual sequence is decomposed and error is compensated, and the error is compensated by the gated recurrent unit neural network.
Maintaining stable predictive performance under extreme weather conditions improves model accuracy and robustness, reduces prediction errors, and enhances the reliability and practicality of the model during rapid climate transitions.
Smart Images
Figure CN122000880A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power generation prediction technology, specifically relating to a photovoltaic power generation prediction method, system, equipment, and medium. Background Technology
[0002] The output of photovoltaic (PV) power plants is influenced by factors such as solar radiation, temperature, cloud cover, dust storms, and wind speed, exhibiting typical strong non-stationary, multi-scale, multi-modal, and coupled time-series characteristics. Against the backdrop of intensifying global climate change, periodic climate phase transitions triggered by large-scale climate modes such as El Niño-Southern Oscillation (ENSO) can lead to drastic shifts in regional weather patterns within a short period, ranging from sunny skies to torrential rains and dust storms. In such complex climate scenarios, achieving high-precision prediction of PV output is crucial for enhancing the power system's capacity to absorb renewable energy and its operational resilience.
[0003] Currently, to improve prediction accuracy, researchers typically employ a hybrid model framework that combines signal decomposition techniques with deep learning models. Much research on hybrid models still focuses primarily on optimizing the model structure itself to improve data fitting accuracy, lacking in-depth research on how meteorological physical mechanisms systematically affect model performance under different weather conditions. This deficiency is particularly pronounced under variable weather models driven by large-scale climate modes such as ENSO. Accurately modeling photovoltaic output under drastically different extreme weather conditions is challenging for current technologies due to the lack of robust modeling methods for climate change. This results in unstable performance of existing models during climate transition periods, making it difficult to guarantee reliability and robustness in engineering applications. Furthermore, most studies focus on point predictions, neglecting the crucial step of prediction error correction. Uncorrected systematic biases in predicted values also affect practicality.
[0004] It is evident that current photovoltaic power prediction technology has significant shortcomings in the following aspects: existing hybrid prediction models lack the ability to adaptively match and globally generalize to multiple weather physical mechanisms during the climate transition period, resulting in low prediction accuracy when extreme weather occurs alternately. Summary of the Invention
[0005] To address the issue of low accuracy in photovoltaic power prediction models during alternating extreme weather events, this invention provides a photovoltaic power prediction method, system, equipment, and medium.
[0006] A first aspect of this invention provides a method for predicting photovoltaic power generation, comprising the following steps: Acquire historical photovoltaic power time series data of photovoltaic power plants and their corresponding historical meteorological data; Adaptive mode decomposition is performed on the historical photovoltaic power time series data to obtain multiple eigenmode function subsequences with different time scale characteristics and a residual subsequence; Based on the intrinsic mode function subsequences, residual subsequences, and corresponding historical meteorological data, an enhanced feature set is constructed; A training dataset including all weather types is constructed. With the goal of minimizing the prediction error of the training dataset, a tornado optimization algorithm with Coriolis force is used to globally optimize the hyperparameters of the prediction model based on the neural network architecture to obtain the globally optimal model hyperparameters. The enhanced feature set is input into a prediction model configured with globally optimal hyperparameters to obtain a preliminary prediction sequence of photovoltaic power; the residual sequence between the preliminary prediction value sequence and the actual value is calculated; the residual sequence is decomposed in a second step to obtain multiple error mode components; a gated recurrent unit neural network is used to learn and predict the temporal patterns of the error mode components to obtain error compensation values. The preliminary prediction sequence is added to the error compensation value to obtain the final photovoltaic power prediction sequence.
[0007] Furthermore, adaptive mode decomposition is performed on the historical photovoltaic power time series data; wherein the adaptive mode decomposition adopts the fully ensemble empirical mode decomposition algorithm with added noise.
[0008] Furthermore, the hyperparameters of the prediction model based on the neural network architecture are globally optimized, wherein the global optimization is performed using a tornado optimization algorithm with Coriolis force.
[0009] Furthermore, the prediction model of the neural network architecture is a neural network model based on the Transformer architecture. The prediction model learns historical data through encoder and decoder structures to obtain the mapping of future power output. The prediction model, based on the globally optimal hyperparameter configuration, processes the input enhanced feature set and outputs a preliminary prediction sequence of photovoltaic power. Specifically, the enhanced feature set is input into the encoder of the prediction model to extract spatiotemporal dependent features; the spatiotemporal dependent features are input into the decoder of the prediction model, and the power values of multiple future time steps are mapped through recursive or forward computation to form the preliminary prediction sequence.
[0010] Furthermore, the residual sequence is further decomposed in a secondary manner, specifically by using the ensemble empirical mode decomposition algorithm to perform mode decomposition on the error sequence.
[0011] Furthermore, the step of using a gated recurrent unit neural network to perform time-series prediction on the multiple error mode components to obtain the error compensation value specifically involves predicting each error mode component separately through a preset neural network, and then summing the prediction results of each preset neural network to obtain the total error compensation value.
[0012] Furthermore, the multiple error mode components are time-series predicted using a gated recurrent unit neural network. Specifically, the multiple error mode components are treated as multivariate time series and input into a gated recurrent unit neural network for prediction to obtain the total error compensation value sequence.
[0013] A second aspect of the present invention provides a photovoltaic power generation prediction system, comprising: The acquisition module is used to acquire historical photovoltaic power time series data of photovoltaic power plants and their corresponding historical meteorological data; The decomposition module is used to perform adaptive mode decomposition on the historical photovoltaic power time series data to obtain multiple eigenmode function subsequences with different time scale characteristics and a residual subsequence. The construction module is used to construct an enhanced feature set based on the intrinsic mode function subsequences, residual subsequences, and corresponding historical meteorological data; The optimization module is used to construct a training dataset that includes all weather types. With the goal of minimizing the prediction error of the training dataset, the hyperparameters of the prediction model based on the neural network architecture are globally optimized using the tornado optimization algorithm with Coriolis force to obtain the globally optimal model hyperparameters. The prediction module is used to input the enhanced feature set into a prediction model configured with globally optimal hyperparameters to obtain a preliminary prediction sequence of photovoltaic power; calculate the residual sequence between the preliminary prediction value sequence and the actual value; perform a secondary decomposition on the residual sequence to obtain multiple error mode components; and use a gated recurrent unit neural network to learn and predict the temporal patterns of the error mode components to obtain error compensation values. The results output module is used to add the preliminary prediction sequence to the error compensation value to obtain the final photovoltaic power prediction sequence.
[0014] A third aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.
[0015] A fourth aspect of the present invention provides a readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.
[0016] The photovoltaic power prediction method provided by this invention has the following beneficial effects: This invention decomposes the original power sequence into physically meaningful sub-modes and integrates them with meteorological driving factors, enabling the model to simultaneously understand both "fluctuation phenomena" and "physical causes." More importantly, a global optimization step, targeting the overall data error across multiple extreme weather events, forces the model to learn a unified mapping relationship adapting to different physical mechanisms, endowing it with "climate resilience." This allows the model to maintain stable performance during periods of rapid climate transition, avoiding drastic fluctuations in prediction performance caused by climate model switching. Specifically, signal decomposition and feature concatenation provide the model with information-rich input, laying the foundation for high-precision learning; global optimization ensures that the model parameters are in a globally optimal state adaptable to multiple scenarios; and the error compensation mechanism specifically captures and corrects residual complex physical effects that the main model has not fully learned. The synergistic effect of these three elements allows the final prediction results to maintain stable prediction performance without adjusting the model structure or parameters when facing drastic shifts in weather patterns such as sunny, rainy, and dust storms caused by ENSO phase transitions, thus improving the model's prediction accuracy during alternating extreme weather events. Attached Figure Description
[0017] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a schematic flowchart of a photovoltaic power generation prediction method provided by the present invention according to an exemplary embodiment.
[0019] Figure 2 This is a schematic diagram of the prediction algorithm framework in an embodiment of the present invention.
[0020] Figure 3 This is a flowchart illustrating the principle of CEEMDAN in an embodiment of the present invention.
[0021] Figure 4 This is a schematic diagram of the IMF components and residual curves in an embodiment of the present invention.
[0022] Figure 5 This is a schematic diagram of the actual power generation curves under three weather conditions during the ENSO climate transition period in an embodiment of the present invention.
[0023] Figure 6 This is a schematic diagram of the power generation prediction curve for sunny days under three types of weather conditions in an embodiment of the present invention.
[0024] Figure 7This is a schematic diagram of the predicted power generation curves for rainy days under three different weather conditions in an embodiment of the present invention.
[0025] Figure 8 This is a schematic diagram of the predicted power generation curves for sandstorm days under three different weather conditions in an embodiment of the present invention. Detailed Implementation
[0026] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.
[0027] Traditional physical models, based on radiative transfer equations or cloud image analysis, offer high accuracy and clear physical interpretability under clear-sky conditions; however, they heavily rely on the accuracy of meteorological inputs, and their performance deteriorates sharply when the weather is complex and variable. Statistical and machine learning models employ data-driven strategies, avoiding complex physical modeling, but require tedious feature engineering and have limited ability to characterize nonlinear relationships.
[0028] ENSO (El Niño-Southern Oscillation) events trigger global climate anomalies. For example, strong El Niño phases often bring drought and sunny weather with high radiation, while subsequent La Niña phases may be accompanied by drastically different environments such as cloudy and rainy weather with high humidity. Existing research shows that accurately modeling photovoltaic (PV) output under drastically different extreme weather conditions is a significant challenge for traditional methods. Due to the lack of modeling methods inherently robust to such dramatic climate changes, existing models are unstable during climate transitions, making it difficult to guarantee reliability and robustness in engineering applications. Furthermore, most studies focus on point predictions, neglecting the crucial step of prediction error correction; uncorrected systematic biases in predictions also affect practicality. Developing accurate PV prediction models capable of operating under rapidly changing weather conditions is essential in real-world climate scenarios. Therefore, a structured, phased, and physics-based prediction framework is needed to enhance cross-weather adaptability.
[0029] Current photovoltaic power prediction technologies suffer from several significant shortcomings: First, they lack adaptability to rapid weather changes. Traditional models mostly employ fixed structures and parameters, lacking mechanisms for adaptive adjustment to different weather patterns. When meteorological conditions change drastically during the ENSO phase transition (e.g., from sunny to rainy or dusty weather), fixed-parameter models struggle to capture the new input-output mapping in a timely and efficient manner, leading to uneven performance across different scenarios. Second, they lack effective processing of non-stationary, multimodal signals. Photovoltaic power sequences under extreme weather conditions often exhibit fluctuations at multiple scales, such as smooth diurnal cycles in sunny weather, rapid random fluctuations in rainy weather, and multi-factor coupled oscillations in dusty weather. If a single model is directly used to forcibly fit the mixed multimodal signals, the model struggles to simultaneously account for high and low frequency components, often resulting in insufficient extraction of key patterns and large prediction errors. Third, the models lack inherent climate robustness design. Most current research focuses on improving prediction accuracy by refining deep learning network structures, but rarely explores in depth how to adjust the model for different weather physics mechanisms. When faced with climate anomalies such as ENSO, this deficiency causes the model to exhibit significant instability and reduced reliability. Finally, neglecting error compensation is also a major drawback of existing technologies. Most forecasting systems only provide one-time predictions without modeling and correcting for systematic residuals. Under complex climate conditions, this means that any model bias will accumulate and affect the final accuracy of the prediction.
[0030] Based on the shortcomings of the existing technologies, this invention aims to improve the robustness and accuracy of photovoltaic power prediction models under conditions of drastic climate change. To achieve this goal, this invention proposes a modular multi-stage prediction framework, which decomposes the complex photovoltaic power prediction problem into four sequentially connected functional modules. Each module performs its own function and specifically addresses a key technical challenge, thereby significantly improving the model's adaptability to meteorological changes as a whole. (1) By preprocessing the signal decomposition, the fluctuation characteristics of different scales in the photovoltaic power sequence are extracted to solve the problems of non-stationarity and mixed modes in the original signal, making it easier for the model to learn stable modes; (2) The global optimization algorithm is used to optimize the model parameter configuration for different weather scenarios, ensuring that the model has near-optimal mapping capabilities in multiple climate scenarios and improving the generalization performance across scenarios; (3) A deep learning module based on Transformer is introduced to fully learn the complex spatiotemporal dependence between photovoltaic power and multiple meteorological factors, improving the modeling ability for nonlinear and multivariate coupled processes; (4) A residual error compensation link is added to perform secondary learning and correction on subtle physical effects that are difficult to capture in the initial prediction, thereby eliminating system bias to the greatest extent. Through the aforementioned multi-pronged improvements, this invention enables a photovoltaic power prediction method highly adaptable to climate change, maintaining stable and excellent prediction accuracy under various weather conditions such as sunny days, rainy days, and sandstorms. Its ultimate goal is to significantly reduce prediction errors and improve indicators such as R² in extreme scenarios such as the complete ENSO climate transition year, thereby greatly enhancing the reliability and practical value of the model in real-world engineering applications.
[0031] like Figure 1 As shown, the method specifically includes the following steps: S1. Obtain historical photovoltaic power time series data.
[0032] S2. Decomposition Step: Perform adaptive mode decomposition on the historical photovoltaic power time series data to obtain multiple subsequences at different time scales.
[0033] S3. Optimization steps: Construct a prediction model based on a deep neural network architecture, use the multiple subsequences at different time scales as training feature data, train the prediction model on a training set containing multiple weather types, and use the overall prediction error of the training set as the objective function to globally optimize the hyperparameters of the prediction model to obtain the optimized prediction model.
[0034] Preferably, in this invention, the Transformer is used as the master prediction model to learn the mapping from historical power and meteorological elements to future power. Compared to traditional LSTM / GRU, the Transformer is more suitable for parallel computation and has stronger representation capabilities on large datasets, enabling it to handle the unified modeling task of distinctly different patterns under sunny and extreme weather conditions. Furthermore, the use of CNN-BiLSTM models or other prediction models in time series forecasting is not elaborated upon in this invention.
[0035] The Transformer model is a deep neural network architecture based on a self-attention mechanism, renowned for its outstanding performance in sequence modeling tasks. The Transformer abandons the recurrent structure of traditional RNNs, using a multi-head attention mechanism to process each element in the sequence in parallel, enabling the model to learn dependencies between any positions in the sequence. Its encoder-decoder structure was first applied in tasks such as machine translation and has since been widely adopted in time series forecasting. The Transformer excels at capturing long-distance dependencies and handling multivariate inputs, making it highly suitable for photovoltaic power forecasting with complex time-series characteristics. In this invention, the Transformer is used as the master forecasting model to learn the mapping from historical power and meteorological elements to future power. Compared to traditional LSTM / GRU, the Transformer is more suitable for parallel computation and has stronger representation capabilities on large datasets, enabling it to handle the unified modeling task of drastically different patterns under sunny and extreme weather conditions.
[0036] S4. Prediction Steps: Using the optimized prediction model, process the feature data containing the subsequences corresponding to the time period to be predicted, and output a preliminary prediction sequence of photovoltaic power.
[0037] S5. Error Compensation Steps: Calculate the error sequence between the preliminary predicted value sequence and the actual power value sequence; decompose and predict the error sequence to obtain the error compensation value sequence.
[0038] S6. Output the final prediction result: Combine the preliminary prediction value sequence with the error compensation value sequence to obtain the final photovoltaic power prediction result.
[0039] The specific execution process is as follows: Based on the decomposition module, the historical photovoltaic power time series data is decomposed using a fully ensemble empirical mode decomposition algorithm with added noise, resulting in multiple intrinsic mode functions (IMFs) representing physical fluctuations at different time scales and a residual subsequence. These IMFs, the residual subsequence, and the corresponding historical meteorological feature data are then input into a prediction model that has undergone hyperparameter global optimization for training and inference. This model learns the complex spatiotemporal dependence between photovoltaic power and multiple source factors, outputting a preliminary photovoltaic power prediction sequence for future periods. The error sequence between this preliminary prediction sequence and the corresponding actual power value sequence is calculated. The error sequence is then decomposed using an ensemble empirical mode decomposition algorithm to obtain multiple error mode components. Subsequently, a gated recurrent unit neural network is used to predict these error mode components, obtaining an error compensation value sequence. Finally, the preliminary prediction sequence and the error compensation value sequence are added together to generate the final photovoltaic power prediction sequence.
[0040] Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN): An improved signal decomposition algorithm, it extends the classic EMD (Empirical Mode Decomposition) by repeatedly decomposing and averaging white noise of varying amplitudes to overcome the mode aliasing problem of EMD. Furthermore, adaptive noise is added in the steps to stabilize the decomposition results. CEEMDAN can decompose complex non-stationary signals into a series of intrinsic mode functions (IMFs) and residuals, where each IMF represents a component of the signal at different time scales. Its applications include processing time series with multi-scale fluctuations. In this invention, it is used to separate high-frequency disturbances and low-frequency trends in photovoltaic power, thereby reducing modeling difficulty.
[0041] Ensemble Empirical Mode Decomposition (EEMD): A precursor to CEEMDAN, EEMD reduces instability by adding finite-amplitude white noise to the original signal, performing multiple EMD decompositions, and then averaging the results. While it lacks the "complete" and "adaptive noise" improvements of CEEMDAN, the principle is similar. In this invention, EEMD is used to decompose residual errors, extracting the dominant modes from the error for use in the error prediction model. Because the error sequence is weaker and shorter than the power sequence, EEMD is sufficient for this task with lower computational cost.
[0042] Gated Recurrent Unit (GRU): A variant of Recurrent Neural Network (RNN), GRU controls the memorization and forgetting of information through "update gates" and "reset gates." Compared to LSTM, it is simpler, has fewer parameters, but performs similarly. GRU is suitable for modeling short-term dependencies in time series and converges faster during training. In this invention, GRU is used to model the predicted residual sequence. Because the residual sequence is relatively simple and the amount of data is limited, a lightweight GRU can fully learn its pattern without overfitting. As the learner of the error compensation module, GRU plays a role in "fine-tuning" the prediction results.
[0043] Based on the above inventive concept, the present invention proposes an embodiment, such as... Figure 2 As shown, the prediction framework in this embodiment consists of four steps connected in series: the CEEMDAN decomposition step (fluctuation mode decoupling stage), the TOC optimization step (mapping relationship optimization stage), the Transformer modeling step (spatiotemporal dependency learning stage), and the EEMD-GRU error compensation step (residual error compensation stage). The input-output relationship and working principle of the four steps are as follows:
[0044] S1, CEEMDAN decomposition steps (wave pattern decoupling).
[0045] The input to this step is historical photovoltaic power time series data (which can be combined with basic meteorological observation data, such as irradiance, for synchronous processing), and the output is several subsequences with different scale characteristics. The original photovoltaic power sequence is decomposed using the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm. CEEMDAN achieves adaptive decomposition of complex non-stationary signals by adding adaptive noise to the original signal and performing multiple EMD decompositions, decomposing the original power signal X(t) into a series of intrinsic mode functions and a residual.
[0046] The working mechanism of the CEEMDAN decomposition algorithm is as follows: Figure 3 As shown, the decomposition steps are as follows: ; x : Original signal, : No. k The average of the intrinsic mode functions (IMFs), k The number of IMFs obtained from the decomposition. r : The final residual term.
[0047] Input: Original photovoltaic power signal X(t) + adaptive white noise; Iteration process: Perform EMD (Empirical Mode Decomposition) on the signal after adding noise multiple times, and output one IMF (Intrinsic Mode Function) in each iteration; Aggregation and update: Take the average of the IMFs of the same order obtained from multiple iterations to obtain stable IMF components; Residual update: After deducting the decomposed IMF components, update the residual r, and repeat the above steps until the residual can no longer be decomposed, and finally obtain all components (IMF1~IMF8 and residual).
[0048] These IMFs represent different frequency components in the raw power: high-frequency IMFs capture transient disturbances (such as output fluctuations caused by rapidly passing clouds), mid-frequency IMFs reflect mesoscale weather evolution (such as power changes caused by the movement of rain bands), and low-frequency IMFs and residuals correspond to deterministic trend components (such as diurnal variations in sunshine).
[0049] The data visualization results after CEEMDAN decomposition are as follows: Figure 4 As shown, Figure 4 Figures (a) to (h) show that the original photovoltaic power signal was decomposed into multiple IMF components. Figure 4 (i) is the residual curve.
[0050] Specifically, Figure 4 (a) and (b) represent high-frequency IMF: the curve fluctuates violently and has a high frequency, corresponding to "transient disturbances" (such as power fluctuations caused by the rapid movement of cloud clusters); Figure 4 (c), (d), (e), and (f) represent the intermediate frequency IMF: moderate fluctuation amplitude and moderate period, corresponding to "mesoscale weather evolution" (such as power changes caused by the movement of rain belts); Figure 4 (g) and (h) represent low-frequency IMF: the curves are smooth and the trends are obvious, corresponding to "deterministic trends" (such as the power base trend caused by the diurnal periodic changes in sunshine).
[0051] By decoupling at the frequency band level, the originally mixed multi-scale fluctuations are broken down into sub-signals with clearer physical meanings, thus significantly reducing the difficulty of subsequent unified modeling. Downstream prediction models can learn the patterns of each modality's subsequence separately, without having to consider both the strong trend and weak details of the entire signal simultaneously. This stage is essentially equivalent to "physical structure mining" of the photovoltaic power sequence—separating various meteorological factors affecting power fluctuations (diurnal cycle, cloud movement, dust disturbance, etc.), reducing the non-stationarity and complexity of the sequence at the data level.
[0052] S2, TOC optimization steps (mapping relationship optimization).
[0053] This step is performed during model training, using the Tornado Optimizer with Coriolis force (TOC) to solve for the model's globally optimal hyperparameters. Its input is a training dataset containing all weather types, resulting in a set of optimized model hyperparameter configurations.
[0054] Tornado Optimizer with Coriolis Force (TOC): A recently proposed metaheuristic global optimization algorithm. Inspired by the formation and development of tornadoes in nature, TOC hypothesizes that there are multiple "storm cells" (i.e., a set of solutions) in the solution space. It updates the solutions by simulating the behavior of airflow rotation and uplift (corresponding to local search) and storm movement (corresponding to global search), and introduces the Coriolis force to simulate the deflection of airflow paths by the Earth's rotation, thereby increasing the diversity and scope of the solution space exploration. The advantages of the TOC algorithm are its strong global optimization capability and good adaptability to complex multi-peak problems, making it suitable for hyperparameter tuning, engineering optimization design, and other fields. In this invention, TOC is used to automatically find the optimal hyperparameters of the prediction model under various weather combinations, effectively improving the overall performance of the model under different climatic conditions.
[0055] Given that the mapping relationship between input features and power output dynamically changes with weather mechanisms, a fixed-parameter model cannot maintain optimal performance in all scenarios. Therefore, this invention introduces an independent global optimization step after signal decomposition to automatically adjust model parameters to adapt to all weather scenarios. In this invention, the TOC optimizer uses the prediction performance across the entire training set as the objective function (e.g., the overall mean squared error (MSE) as the optimization objective) to globally optimize the model's key hyperparameters. These hyperparameters include, but are not limited to, the initial learning rate and regularization coefficient of the Transformer model, which significantly affect the model's generalization performance. TOC optimization updates the position of the hyperparameter group by simulating the generation and evolution of tornado cyclones, thereby gradually approaching the global minimum of the error. The final set of optimal hyperparameters enables the prediction model to achieve performance balance under various climatic conditions. That is, the optimized model can achieve near-optimal error performance in various scenarios such as radiation-dominated environments and multi-factor coupled environments. This stage essentially optimizes the physical mapping relationship of the model's input-output relationship, ensuring that the model can adapt to multimodal meteorological inputs and that the performance of another scenario is not weakened by the selection of hyperparameters for a particular scenario. Through the TOC step, the globally optimal hyperparameter configuration is automatically applied to the next step of model training, laying a solid foundation for the training of subsequent deep learning models.
[0056] S3, Modeling Steps (Spatiotemporal Dependency Learning).
[0057] A prediction model based on a deep neural network with a globally optimal hyperparameter configuration is constructed. The input includes the IMF subsequences obtained from the aforementioned decomposition, the residual sequence, and an enhanced feature set composed of relevant meteorological features (such as irradiance, temperature, and wind speed). The output is a preliminary predicted sequence of photovoltaic power. This invention uses a deep neural network with a Transformer architecture as the main learner to leverage its powerful spatiotemporal dependency capture capabilities.
[0058] In this invention, the prediction model learns the mapping from historical data to future power output through an encoder-decoder structure. The encoder layer receives multi-dimensional input (including historical power values of each decomposed mode and meteorological variables at the corresponding time) and extracts high-level spatiotemporal feature representations. The decoder layer predicts future photovoltaic power values step by step based on the encoded representation and its own recursive output. Through a self-attention mechanism, the model can focus on the correlation information between different modes and different time steps, such as the pattern of a certain IMF component within a certain period, or the correspondence between a sudden drop in irradiance and a decrease in power, thereby capturing complex cross-variable and cross-time dependencies. It is worth emphasizing that, since the input data after CEEMDAN decomposition has more physical meaning in each mode, the Transformer model can learn corresponding patterns for different modes, such as learning short-period fluctuations of high-frequency IMF (cloud-induced disturbances) and smooth changes of low-frequency IMF (solar radiation trend), thereby improving the precision and accuracy of prediction. In addition, the hyperparameter configuration provided by TOC optimization ensures that the Transformer model will not overfit a single scenario during training, but will learn under globally optimal parameters, giving it a good fitting ability for diverse weather data.
[0059] The modeling process begins with the integration and representation of multi-source input data. Before entering the prediction model, the historical photovoltaic power sequence has been decomposed into multiple subsequences with clear physical meaning by the CEEMDAN algorithm (such as high-frequency IMF representing cloud disturbances and low-frequency IMF representing diurnal cycles), and aligned and concatenated with meteorological observation data (such as irradiance and temperature) at the corresponding time points to form a multivariate time series input. The prediction model first projects all features at each time step into a high-dimensional vector space through an embedding layer and injects positional encoding information, enabling the model to understand the order and relative position of the data on the time axis, laying the foundation for subsequent deep relational learning.
[0060] The encoder, as the first stage of the module, has the core task of mining and fusing complex spatiotemporal dependencies in the historical sequence. It operates based on the multi-head self-attention mechanism in the Transformer architecture. This mechanism allows the model to compute the correlation strength between all features at any two historical time steps in the sequence in parallel, thereby automatically identifying key patterns affecting photovoltaic power and their time spans. For example, it can discover a strong correlation between "the sudden drop in power yesterday afternoon due to cloud movement" and "the current trend of decreasing irradiance." Through this global, non-sequential attention computation, the encoder effectively compresses and refines the entire historical context into a set of high-level "memory" vectors. These vectors not only capture the evolution of power itself at different time scales but also deeply integrate the coupling relationship between meteorological factors and power fluctuations, forming a comprehensive and abstract feature representation of past states.
[0061] The decoder is responsible for generating the future prediction sequence autoregressively. It works in a step-by-step manner: when predicting the power value at a specific future moment, the decoder first reviews earlier future values it has generated (or knows), and ensures that the prediction does not depend on unknown future information by masking self-attention. Then, it performs a crucial step—encoder-decoder cross-attention. In this step, the decoder dynamically and selectively "questions" the historical memory matrix generated by the encoder, retrieving the historical context information most relevant to the current prediction moment. For example, when predicting the power tomorrow afternoon, it might focus on referencing illumination patterns during similar periods of clear weather in history and recent similar weather evolution fragments. Finally, the decoder combines the retrieved historical context information with the generated prediction context, calculates and outputs the preliminary prediction value for that moment using a feedforward neural network. This "retrieval-generation" process is recursively performed time-step until a preliminary prediction sequence for the complete future time period is generated.
[0062] S4. Error compensation steps (residual error compensation).
[0063] This step models and corrects the residual errors of the initial predictions to obtain the final prediction output. Its input is the main model's prediction error sequence (the residual sequence obtained by subtracting the initial prediction value from the actual value), and its output is the predicted value for that error, i.e., the compensation amount. The core idea is that the initial prediction errors contain physical information that the main model has not learned. Therefore, instead of directly allowing these errors to affect the final result, it is better to build a new model for the error sequence itself to capture and correct the patterns within it.
[0064] Preferably, this invention uses the EEMD-GRU combined model as an example of the error compensation step. First, the error sequence of the main model is decomposed into several IMF error components using the EEMD (Ensemble Empirical Mode Decomposition) algorithm. EEMD decomposes complex residuals into more regular sub-error signals (e.g., certain periodic underestimation or overestimation patterns). Then, a GRU (Gated Recurrent Unit) neural network is introduced as a lightweight learning model to perform secondary prediction on the error IMF sequence. GRU is a recurrent neural network unit, chosen for its fewer parameters and ability to capture short-term dependencies in time series, making it efficient for modeling the dynamics of error sequences. The GRU model takes the past time period of the error sequence as input and outputs the error prediction for the next time period. Each IMF component can be predicted separately using the GRU model, and then summed to obtain the total error prediction, or it can be directly input into a single GRU model in the form of a multivariate time series for prediction. The overall idea is to perform another prediction correction in the error space. Since the error compensation step does not repeat the prediction of the power itself, but focuses on capturing and correcting the residual bias of the main model, it can specifically compensate for the deficiencies of the main model. From a physics perspective, this step provides a feedback mechanism, reintroducing complex effects that the main model hasn't learned well into the error sequence for the secondary model to learn. For example, in multi-factor coupled scenarios like rainy days, the main model might struggle to fully learn the precipitous drop in output caused by sudden rainfall, but these uncaptured phenomena will be reflected in the residuals. The error module learns these residual patterns and corrects them, improving the overall prediction accuracy and robustness. Introducing this error compensation mechanism maximizes the framework's adaptability to various anomalous perturbations, essentially providing the model with a self-correcting function, pushing prediction accuracy close to the theoretical limit.
[0065] In summary, this invention first extracts multi-scale features from historical power data using CEEMDAN decomposition; secondly, it optimizes the model hyperparameter configuration for global data using the TOC algorithm; next, it uses the optimized Transformer model to learn temporal relationships and obtain preliminary predictions; finally, it uses EEMD decomposition + GRU model to perform secondary learning and correction on the preliminary prediction residuals, outputting the final prediction result. Each step works closely together through a data interface: the IMF subsequence output from the decomposition step serves as the training input for the modeling module; the hyperparameters provided by the optimization step guide the training process of the prediction model (i.e., the Transformer model); and the error compensation step uses the prediction residuals of the main model for correction. This multi-stage serial architecture fully leverages the advantages of each stage, forming a systematic solution for photovoltaic power prediction under complex climatic conditions.
[0066] By employing a fully ensemble empirical mode decomposition algorithm with added noise to decompose historical photovoltaic power sequences, the original strongly non-stationary, multi-modal mixed power signal is adaptively decoupled into multiple eigenmode functions and residual subsequences with clear physical meaning. This solves the problem of mixed original signal modes and difficulty in unified modeling, reducing sequence complexity from the data source. Secondly, a global optimization algorithm targeting the overall prediction error covering multiple weather types is introduced to optimize the hyperparameters of the prediction model. This feature overcomes the inherent defects of traditional models with fixed parameters and difficulty in adapting to climate mechanism changes. It forces the model to learn and balance the mapping relationship under different weather scenarios at the parameter level, ensuring stable prediction performance during rapid transitions in extreme weather such as sunny, rainy, and dust storms caused by the ENSO phase transition. Furthermore, through secondary decomposition of the initial prediction residuals and error compensation, the residual physical effects that the main prediction model failed to learn due to its limitations are systematically captured and corrected, especially the complex nonlinear responses under multi-factor coupled scenarios. This adds a closed-loop feedback and self-correction link to the prediction system, enabling targeted compensation for systematic biases left over from the initial prediction. Ultimately, this endows the prediction model with consistent high accuracy and robustness across all climate scenarios, solving the problem of insufficient reliability in existing technologies for climate transition predictions.
[0067] Experimental verification: This example demonstrates how to use the four-stage framework described in this invention to predict the power generation of a photovoltaic power station in a study area under three weather conditions during the ENSO climate transition period. The actual power generation under the three weather conditions is as follows: Figure 5 As shown.
[0068] Evaluation metrics R² and RMSE: R² (coefficient of determination) and RMSE (root mean square error) are commonly used evaluation metrics in photovoltaic power prediction. R² reflects the goodness of fit of the model to the actual data, ranging from 0 to 1. The closer to 1, the better the predicted value matches the actual value (R²=1 indicates a completely accurate prediction). RMSE measures the absolute magnitude of the prediction error; the smaller the value, the more accurate the prediction. The combination of these two metrics comprehensively reflects the model performance. In various scenarios, the R² of this invention is close to 1, and the RMSE is significantly lower than that of the comparison model. These quantitative metrics have been detailed in the previous section on technical effects. Readers can more intuitively understand the performance advantages of this invention through the comparison of these metrics.
[0069] Example 1: Photovoltaic power prediction under sunny weather conditions.
[0070] Choose a clear, cloudless period from July 7th to 9th, 2024, such as... Figure 5The green curve serves as an example for analysis. During this period, solar irradiance is stable and cloud-free, resulting in a smooth solar power output curve, primarily governed by the deterministic physical laws governing Earth's rotation and the relative positions of the Earth and the Sun.
[0071] like Figure 6 As shown, in a sunny day example, the final prediction of the method of this invention almost perfectly reproduces the actual power curve. Particularly noteworthy is that at the most dramatic turning points in output change—early sunrise and late sunset—the error compensation module corrects the slight underestimation / overestimation of the Transformer model, resulting in a high degree of agreement between the predicted and measured curves. The RMSE throughout the daytime is only about 1 kW, making the predicted curve virtually indistinguishable from the actual curve to the naked eye. In contrast, while traditional single-stage models are easy to fit the overall trend on sunny days, they often struggle to capture the detailed changes at dawn and dusk, exhibiting phase lag or amplitude errors of a few minutes. This method, by decomposing high-frequency components and performing secondary correction, significantly improves the ability to characterize details, achieving an R² index of over 0.99 under sunny conditions. This means that under ideal irradiance conditions, this invention achieves almost zero-error accurate prediction.
[0072] Example 2: Photovoltaic power prediction in rainy weather scenarios.
[0073] The period from August 2nd to 4th, 2024, characterized by showers and dense clouds, was selected as a representative rainy weather scenario for testing. During this period, clouds rapidly converged and dispersed, rainbands moved frequently, and photovoltaic power output exhibited rapid and random fluctuations, often experiencing precipitous drops or rebounds within a short time. This type of weather poses a significant challenge to forecasting models because the discontinuous changes in output are difficult to approximate with smoothing functions.
[0074] like Figure 7As shown, in the most challenging scenario of rainy weather, the method of this invention demonstrates performance far superior to traditional models. The final prediction curve closely follows the dramatic fluctuations in actual power, successfully predicting the trend and approximate magnitude of multiple large power jumps. Although there may still be slight errors at individual moments, the overall trend is highly consistent with reality. After adopting this method, the overall prediction R² for rainy weather is close to 0.978, significantly higher than the level of less than 0.9 typically achieved by conventional methods, with a root mean square error of only about 9.7 kW. In contrast, the baseline model without any special processing often fails to capture cloud-induced dramatic changes during this period, resulting in a prediction curve that is almost a smooth mean line, unable to keep up with the large fluctuations in actual power, leading to large errors and very low fit. Through the multi-stage processing of this invention, the model not only identifies the risk of impending power drop in advance (under the effect of TOC optimization), but also further reduces the originally difficult-to-eliminate random errors through error compensation. This means that in actual grid dispatching, this method can timely predict sudden drops in photovoltaic output, giving dispatching time to react and greatly improving the practicality and reliability of photovoltaic output prediction under severe weather conditions such as rainy days.
[0075] Example 3: Photovoltaic power prediction under sandstorm weather scenario.
[0076] The period from December 9th to 11th, 2024, during which a dust storm occurred, was selected as the third typical example. Dust storms often involve a large amount of dust in the sky, leading to a significant reduction in solar radiation, and may be accompanied by strong winds and dry conditions. Its impact on photovoltaic power is manifested as a significant decrease in output and irregular fluctuations. Unlike rainy days, while the power changes caused by dust storms are dramatic, they are relatively dominated by radiation changes. Random factors (such as dust concentration and wind speed) also contribute to complex effects, making it a scenario with multiple factors at play, but with radiation reduction as the core mechanism.
[0077] like Figure 8As shown, in a dust storm example, the method of this invention successfully tracked the trend of power change with dust storm intensity. The final prediction curve reflects every significant drop and rebound in power during the dust storm, highly consistent with the actual curve trend. Quantitative results show that the R² of this method in a dust storm scenario is approximately 0.979, and the RMSE is only on the order of several kilowatts, a significant improvement compared to models without special measures (R² is usually <0.95, RMSE can reach tens of kilowatts). Notably, ablation experiments show that even without the error compensation module, the CEEMDAN-Transformer, after TOC optimization, already achieves good performance (R² close to 0.973) on dust storm days; this indicates that the first few modules of this invention (decomposition and optimization) have enabled the model to adapt well to the special environment dominated by radiation, such as dust storms. The introduction of error compensation further enhances the model, eliminating the slightly lagging systematic error and making the prediction trajectory more realistic. In actual photovoltaic power plant operation and maintenance, dust storms often pose an extremely challenging situation for output prediction. This method significantly reduces the prediction uncertainty in such cases, demonstrating strong practical value.
[0078] As can be seen from the three examples above, the application process of the method of this invention is largely consistent under different weather conditions, all demonstrating a complete closed loop of preprocessing decomposition + optimization parameter tuning + deep model prediction + error correction. Each step plays a role based on the data characteristics of the specific scenario, ultimately achieving near-optimal prediction results in each scenario. This ability to "use one method for multiple weather conditions" is a prominent advantage of this invention, indicating that the method possesses good versatility and robustness, sufficient to address the challenges of photovoltaic power prediction posed by various extreme weather events during the ENSO climate phase transition.
[0079] This invention has achieved significant progress in the field of photovoltaic power prediction through an innovative combination of technologies. Its technical effects are mainly reflected in the following aspects: 1. A significant leap forward in prediction accuracy: This invention achieves near-limit prediction accuracy across all weather scenarios through a multi-stage collaborative framework. Experimental data shows that the R² index reaches 0.98999, 0.97782, and 0.97967 under three typical weather conditions: sunny, rainy, and dust storm. Particularly noteworthy is the reduction in root mean square error (RMSE) during the most complex dust storm weather, which significantly decreased from 112.1 kW in the baseline model to 4.8672 kW, fully demonstrating the powerful ability of the "decomposition-compensation" architecture to capture complex physical effects.
[0080] 2. Significantly enhanced climate adaptability and robustness: This invention employs a global optimization strategy to endow the model with inherent resilience in the face of dramatic climate change. The model maintains top-tier performance under significantly different physical scenarios, including sunny, rainy, and dust storm conditions, without requiring model switching or parameter reconfiguration for different weather conditions. In cross-year generalization tests, the performance degradation of this invention's framework is far less than that of the baseline model, indicating that it learns robust physical laws rather than overfitting to climate noise specific to a particular year.
[0081] 3. The interpretability of the model has been effectively improved: The modular design of this invention not only improves performance but also enhances the interpretability of the model. The output of the physical fluctuation decoupling module naturally corresponds to physical processes at different time scales, enabling researchers to trace the specific physical sources affecting power fluctuations. Simultaneously, the residual sequence learned by the error compensation module clearly reveals physical effects not yet fully understood by the main model, providing a clear direction for subsequent model improvements.
[0082] 4. Excellent architectural versatility and outstanding engineering application value: The framework design of this invention demonstrates broad applicability. Experiments show that even when the core spatiotemporal dependency learning module is replaced with other models with significantly different structures, the complete framework still maintains excellent performance. This indicates that the core performance improvement stems from the architecture and compositional ideas of this invention, rather than the advantages of a specific algorithm. This characteristic lays a solid foundation for subsequent model lightweighting and deployment on edge computing devices, and has significant engineering application value.
[0083] Based on the above inventive concept, the present invention also provides a photovoltaic power generation prediction system, comprising: The acquisition module is used to acquire historical photovoltaic power time series data of photovoltaic power plants and their corresponding historical meteorological data; The decomposition module is used to perform adaptive mode decomposition on historical photovoltaic power time series data to obtain multiple eigenmode function subsequences with different time scale characteristics and a residual subsequence. The module is used to construct an enhanced feature set based on the intrinsic mode function subsequences, residual subsequences, and corresponding historical meteorological data. The optimization module is used to construct a training dataset that includes all weather types. With the goal of minimizing the prediction error of the training dataset, the hyperparameters of the prediction model based on the neural network architecture are globally optimized using the tornado optimization algorithm with Coriolis force to obtain the globally optimal model hyperparameters. The prediction module is used to input the enhanced feature set into the prediction model configured with the global optimal hyperparameters to obtain the preliminary prediction sequence of photovoltaic power; calculate the residual sequence between the preliminary prediction value sequence and the actual value; perform secondary decomposition on the residual sequence to obtain multiple error mode components; and use a gated recurrent unit neural network to learn and predict the temporal pattern of the error mode components to obtain the error compensation value. The results output module is used to add the preliminary prediction sequence to the error compensation value to obtain the final photovoltaic power prediction sequence.
[0084] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The steps of the provided photovoltaic power prediction method.
[0085] This invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for various operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above-mentioned functions. Figure 1 The steps of the provided photovoltaic power prediction method.
[0086] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0087] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0088] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0089] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0090] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the patent of the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A method for predicting photovoltaic power generation, characterized in that, The method includes: Acquire historical photovoltaic power time series data of photovoltaic power plants and their corresponding historical meteorological data; Adaptive mode decomposition is performed on the historical photovoltaic power time series data to obtain multiple eigenmode function subsequences with different time scale characteristics and a residual subsequence; Based on the intrinsic mode function subsequences, residual subsequences, and corresponding historical meteorological data, an enhanced feature set is constructed; A training dataset including all weather types is constructed. With the goal of minimizing the prediction error of the training dataset, a tornado optimization algorithm with Coriolis force is used to globally optimize the hyperparameters of the prediction model based on the neural network architecture to obtain the globally optimal model hyperparameters. The enhanced feature set is input into a prediction model configured with globally optimal hyperparameters to obtain a preliminary prediction sequence of photovoltaic power; the residual sequence between the preliminary prediction value sequence and the actual value is calculated; the residual sequence is decomposed in a second step to obtain multiple error mode components; a gated recurrent unit neural network is used to learn and predict the temporal patterns of the error mode components to obtain error compensation values. The preliminary prediction sequence is added to the error compensation value to obtain the final photovoltaic power prediction sequence.
2. The method according to claim 1, characterized in that, Adaptive mode decomposition is performed on the historical photovoltaic power time series data. The adaptive mode decomposition adopts the complete set empirical mode decomposition algorithm with added noise.
3. The method according to claim 1, characterized in that, The hyperparameters of the prediction model based on the neural network architecture are globally optimized, wherein the global optimization is performed using the tornado optimization algorithm with Coriolis force.
4. The method according to claim 1, characterized in that, The prediction model of the neural network architecture is a neural network model based on the Transformer architecture. The prediction model learns historical data through encoder and decoder structures to obtain the mapping of power output at future time. The prediction model based on the globally optimal hyperparameter configuration processes the input enhanced feature set and outputs a preliminary prediction sequence of photovoltaic power. Specifically, the enhanced feature set is input into the encoder of the prediction model to extract spatiotemporal dependent features. The spatiotemporal dependent features are input into the decoder of the prediction model, and the power values for multiple future time steps are mapped through recursive or forward computation to form the preliminary prediction sequence.
5. The method according to claim 1, characterized in that, The residual sequence is then subjected to a secondary decomposition, specifically by using the ensemble empirical mode decomposition algorithm to perform mode decomposition on the error sequence.
6. The method according to claim 1, characterized in that, The step of using a gated recurrent unit neural network to perform time-series prediction on the multiple error mode components to obtain an error compensation value involves predicting each error mode component separately through a preset neural network, and then summing the prediction results of each preset neural network to obtain the total error compensation value.
7. The method according to claim 1, characterized in that, The multiple error mode components are time-series predicted using a gated recurrent unit neural network. Specifically, the multiple error mode components are treated as multivariate time series and input into a gated recurrent unit neural network for prediction to obtain the total error compensation value sequence.
8. A photovoltaic power generation prediction system, characterized in that, include: The acquisition module is used to acquire historical photovoltaic power time series data of photovoltaic power plants and their corresponding historical meteorological data; The decomposition module is used to perform adaptive mode decomposition on the historical photovoltaic power time series data to obtain multiple eigenmode function subsequences with different time scale characteristics and a residual subsequence. The construction module is used to construct an enhanced feature set based on the intrinsic mode function subsequences, residual subsequences, and corresponding historical meteorological data; The optimization module is used to construct a training dataset that includes all weather types. With the goal of minimizing the prediction error of the training dataset, the hyperparameters of the prediction model based on the neural network architecture are globally optimized using the tornado optimization algorithm with Coriolis force to obtain the globally optimal model hyperparameters. The prediction module is used to input the enhanced feature set into a prediction model configured with globally optimal hyperparameters to obtain a preliminary prediction sequence of photovoltaic power; calculate the residual sequence between the preliminary prediction value sequence and the actual value; and perform a secondary decomposition on the residual sequence to obtain multiple error mode components. A gated recurrent unit neural network is used to learn and predict the temporal patterns of the error mode components to obtain error compensation values; The results output module is used to add the preliminary prediction sequence to the error compensation value to obtain the final photovoltaic power prediction sequence.
9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 7.
10. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 7.