High-resolution long-term photovoltaic power prediction method and system based on double-branch architecture

Abstract

This invention belongs to the fields of new energy technology and artificial intelligence technology, and provides a high-resolution long-term photovoltaic power prediction method and system based on a dual-branch architecture. By constructing a parallel dual-branch architecture, the baseline branch extracts multi-scale features, and the ramp branch identifies power ramping events. The features from both branches are embedded through multiple channels and then fed into a BiGRU-Enhanced Transformer model with shared weights. By strategically deploying BiGRU layers at the input and between the encoder and decoder, local instantaneous dependencies and global long-term trends are captured simultaneously. A multi-task joint optimization loss function is designed for the ramp branch, using weighted cross-entropy and... L The probability, temporal location, and fluctuation amplitude of slope occurrence are jointly optimized using the L1 norm. Finally, nonlinear residual correction is applied to the outputs of the two branches to achieve the final prediction. This invention effectively overcomes the oversmoothing effect of deep learning models, demonstrating extremely high capture accuracy and robustness in long-term prediction of one month's worth of data at 1-minute resolution.

Description

Technical Field

[0001] This invention belongs to the fields of new energy technology and artificial intelligence technology, and particularly relates to a high-resolution long-term photovoltaic power prediction method and system based on a dual-branch architecture. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] The proportion of photovoltaic (PV) power generation in the energy system is growing rapidly. Its output is significantly affected by meteorological conditions, exhibiting strong intermittent, fluctuating, and nonlinear characteristics. Accurate PV power forecasting is crucial for grid dispatch, frequency stability, and voltage control. However, in high-resolution and long-term forecasting scenarios, existing forecasting methods generally suffer from oversmoothing, making it difficult to accurately capture sharp fluctuations and high-frequency ramp-up events in the power curve. This leads to a rapid accumulation of forecast errors, posing a serious challenge to the safe and stable operation of the power system.

[0004] Currently, photovoltaic power prediction technologies are mainly divided into three categories: physical models, statistical models, and machine learning models.

[0005] Physical models rely on detailed physical parameters, resulting in high computational costs and stringent requirements for input data accuracy. Statistical models are suitable for stationary sequences but struggle to handle the strong nonlinearity and nonstationarity of photovoltaic data. While deep learning models have improved their nonlinear fitting capabilities, they still exhibit smoothing effects in large-scale, long-sequence, high-resolution predictions, making it difficult to effectively capture ramp-up events, and their long-term prediction accuracy decays significantly.

[0006] While existing technologies include improved schemes based on signal decomposition, dual-branch structures, and hybrid Transformers, they have not yet achieved high-precision separation of trend and climbing events, effective balance between local short-term and global long-term dependencies, or intelligent residual correction based on nonlinear fusion. This limits the accuracy and robustness of high-resolution long-term predictions. Summary of the Invention

[0007] To address at least one of the technical problems mentioned above, this invention provides a high-resolution long-term photovoltaic power prediction method and system based on a dual-branch architecture, overcoming the shortcomings of insufficient accuracy, excessive smoothing, and weak ability to capture ramp-up events in the prior art for high-resolution long-term photovoltaic power prediction.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture, comprising the following steps: Acquire historical photovoltaic power generation data and corresponding meteorological data, and preprocess the historical photovoltaic power generation data to obtain a photovoltaic power generation sequence; The power sequence is fed into the baseline branch and the ramp branch in parallel for processing; in the baseline branch, the power sequence is decomposed into multiple intrinsic mode functions; in the ramp branch, the ramp event sequence and ramp flags are extracted. The results of the dual-branch processing are input in parallel into the trained power prediction model to obtain the baseline predicted power and the ramp predicted power; the construction process of the power prediction model includes: The features extracted from each branch are concatenated with meteorological data and time-coded features to obtain the multi-channel input embedding of each branch; the input embeddings of the two branches are fed into a shared BiGRU enhanced Transformer network for prediction, and the baseline power prediction sequence and the probability, start and end time and amplitude prediction of the slope event are obtained respectively. The baseline predicted power and the ramp-up predicted power are nonlinearly fused to generate the final high-resolution long-term photovoltaic power prediction result.

[0009] Furthermore, the preprocessing of historical power generation data includes outlier correction, which includes outlier handling for three different scenarios.

[0010] Furthermore, in the baseline branch, the power sequence is decomposed into multiple intrinsic mode functions, including MF1-IMF7. MF1 characterizes the low-frequency baseline trend of daily operation determined by the solar altitude angle, capturing long-term trends. IMF2-IMF4 capture minute-level mid-frequency fluctuations caused by cloud movement. IMF5-IMF7 lock in subtle high-frequency noise and random disturbances, capturing high-frequency disturbances.

[0011] Furthermore, the process of embedding the inputs of the two branches into a shared BiGRU-enhanced Transformer network for prediction includes the following steps for each branch: The input is embedded into the front BiGRU layer to extract high-dimensional hidden features. The hidden features are linearly added to the fixed-position encoding based on the sine-cosine function and then sent to the decoder of the Transformer model. The Transformer encoder performs multi-head self-attention global dependency modeling to obtain global semantic features. The global semantic features are further refined by the inter-layer BiGRU layer and the output is used as high-order historical context features. The future weather forecast features for the period to be predicted are fused with location codes and used as the initial input to the decoder. A masked multi-head self-attention mechanism is then used to extract future expected features and generate a query vector. Q Subsequently, the query vector and higher-order historical context features are used as the key matrix. K Sum matrix VThe corresponding branch prediction sequence is obtained by performing association matching and information fusion in the encoder-decoder cross-attention layer.

[0012] Furthermore, the location code calculation formula is as follows: , , In the formula, and These represent the positional encoding values ​​of the input sequence in the odd / even feature dimension, respectively. Indicates the time step position. i For dimensional indexing, d Represents the encoding dimension.

[0013] Furthermore, the ramp branch utilizes a dedicated multi-task joint optimization loss function. Loss Training is defined as: , in, L pred The weighted cross-entropy loss is designed for slope occurrence indicators to address class imbalance caused by fewer climbing events. L time The loss is the prediction error at the start and end time points. L amp This is the prediction error loss for the slope power change amplitude. , , These are the weighting coefficients.

[0014] Furthermore, the baseline predicted power and the slope predicted power are nonlinearly fused. The baseline power prediction sequence, the slope event prediction results and meteorological characteristics are used as inputs. The LightGBM model is used to predict the residual between the baseline prediction and the actual power. The fused prediction value is the sum of the baseline prediction value and the LightGBM prediction residual.

[0015] A second aspect of the present invention provides a high-resolution long-term photovoltaic power prediction system based on a dual-branch architecture, comprising: The data preprocessing module is used to acquire historical photovoltaic power generation data and corresponding meteorological data, and to preprocess the historical photovoltaic power generation data to obtain a photovoltaic power generation sequence. The feature extraction module is used to feed the power sequence into the baseline branch and the ramp branch in parallel for processing; in the baseline branch, the power sequence is decomposed into multiple intrinsic mode functions; in the ramp branch, the ramp event sequence and ramp flags are extracted. The branch power prediction module is used to input the results of the dual-branch processing in parallel into the trained power prediction model to obtain the baseline predicted power and the ramp-up predicted power; the construction process of the power prediction model includes: The features extracted from each branch are concatenated with meteorological data and time-coded features to obtain the multi-channel input embedding of each branch; the input embeddings of the two branches are fed into a shared BiGRU enhanced Transformer network for prediction, and the baseline power prediction sequence and the probability, start and end time and amplitude prediction of the slope event are obtained respectively. The power prediction output module is used to nonlinearly fuse the baseline predicted power and the ramp-up predicted power to generate the final high-resolution long-term photovoltaic power prediction result.

[0016] A third aspect of the present invention provides a computer-readable storage medium.

[0017] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture as described above.

[0018] A fourth aspect of the present invention provides a computer device.

[0019] A computer device 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 steps in the high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture as described above.

[0020] Compared with the prior art, the beneficial effects of the present invention are: This invention utilizes a dual-branch parallel architecture and dedicated ramp-up event extraction. Combined with a dual-branch photovoltaic power prediction model, it captures high-resolution data through the ramp-up prediction branch. By incorporating the multi-head self-attention mechanism of BiGRU-Enhanced Transformer, it significantly enhances the ability to capture short-term local dependencies and model long-term global dependencies, reducing the loss of long-sequence information. It effectively solves the over-smoothing problem of deep learning models, enabling the prediction of high-resolution data. It also addresses the shortcomings of existing technologies in the field of high-resolution long-term photovoltaic power prediction.

[0021] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0023] Figure 1 This is a flowchart of the high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture provided in an embodiment of the present invention; Figure 2 This is a diagram of the BiGRU-Enhanced Transformer model provided in this embodiment of the invention; Figure 3 This is a comparison chart of the prediction accuracy of the five models provided in the embodiments of the present invention. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0026] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0027] Example 1 like Figure 1 As shown, this embodiment provides a high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture, including the following steps: Step 1: Obtain historical power generation data and corresponding meteorological data, and preprocess the historical power generation data to obtain the power generation sequence; In this embodiment, taking the 1-minute resolution measured power data of a 100 MWp ground-mounted photovoltaic power station (geographic coordinates 30.25°N, 120.16°E) as an example, the preprocessing process includes: First, outlier correction is performed, including three cases: Zero or near-zero power readings occurring during periods of high daytime irradiance are identified as potential equipment failures or significant sensor errors and are replaced using time-series interpolation methods. Small non-zero power values ​​recorded during nighttime periods when irradiance is zero are considered sensor drift and corrected to zero. Power values ​​exceeding the power plant's installed capacity are limited and replaced with the maximum capacity.

[0028] Meanwhile, the original 15-minute resolution meteorological data and irradiance data were upsampled to 1-minute resolution using linear interpolation to ensure data synchronization.

[0029] Step 2: The power sequence is fed into the baseline branch and the ramp branch in parallel for processing; in the baseline branch, the power sequence is decomposed into multiple intrinsic mode functions; in the ramp branch, the ramp event sequence and ramp flags are extracted. In this embodiment, in the baseline branch, the power sequence is decomposed into multiple intrinsic mode functions and VMD decomposition is performed. The preprocessed power sequence is decomposed into 7 intrinsic mode functions, MF1-IMF7. Among them, MF1 represents the low-frequency baseline trend of daily operation determined by the solar altitude angle, accurately capturing the long-term trend; IMF2-IMF4 capture minute-level mid-frequency fluctuations caused by cloud movement; and IMF5-IMF7 lock in subtle high-frequency noise and random disturbances, capturing high-frequency disturbances. As a further implementation, when performing VMD decomposition, the core constraint variational problem of VMD is: , , In the formula, The decomposition yields the first... k One modal component, Represents the Dirac function, This represents the preprocessed original photovoltaic power generation sequence. Indicates the first k The center frequencies of the intrinsic mode components This represents the total number of modes in the decomposition; Specifically, by introducing a quadratic penalty factor and Lagrange multipliers, the problem is transformed into an unconstrained problem, and then iteratively solved in the Fourier domain using the alternating direction multiplier method until convergence. In this embodiment, at the slope branch, the slope event sequence and slope marker are extracted, the power sequence is scanned using a peak-valley detection algorithm, and the slope judgment criteria are defined: , In the formula, This represents the photovoltaic power value at the end of the sliding window. Indicates the start time of the sliding window t Photovoltaic power value at the location, Indicates the length of the time window for determining the hill climb; In this embodiment, it is set Within a sliding window of minutes, if the power change exceeds 15% (i.e. 15 MW) of the power plant's installed capacity, it is determined to be a ramp event. Once a ramp event is determined to have occurred, the start time, end time, and power change amplitude of the event are recorded simultaneously, and a binary ramp flag is generated, which is 1 if it has occurred and 0 if it has not occurred.

[0030] Step 3: Input the results of the dual-branch processing in parallel into the trained power prediction model to obtain the baseline predicted power and the ramp predicted power; Specifically, the steps include the following: Step 301: Concatenate the features extracted from each branch with meteorological data and time-coded features to obtain the multi-channel input embedding of each branch; Before performing feature stitching, the Pearson correlation coefficients of various meteorological factors and power series are first calculated. Among them, irradiance, temperature, precipitation, wind speed at 50 meters height and humidity have strong correlations, while humidity is excluded due to low prediction accuracy and cloud coverage is also excluded due to large prediction uncertainty. The IMF components, slope event sequences, and slope markers generated in step 2, along with the filtered meteorological features and related time series features, are concatenated to form a 17-dimensional comprehensive feature vector containing 7 VMD components, 3 meteorological features, 1 irradiance feature, 5 time-related periodic features, and 1 slope marker. Each feature group is mapped to a unified hidden space through an independent linear projection layer, and a multi-channel input embedding is generated through a connection operation, represented as: , In the formula, E feat The final multi-channel embedding represents a high-dimensional feature vector formed by dimensional alignment and fusion of all independent feature groups. This indicates a splicing operation. and This represents the independent linear projection layers that act on VMD decomposition and meteorological features. and This represents the original input vector for VMD decomposition and meteorological feature group; Step 302: The inputs of the two branches are embedded and fed into the shared BiGRU enhanced Transformer network in parallel for prediction, to obtain the baseline power prediction sequence and the probability of occurrence, start and end time and amplitude prediction of the ramp event, respectively. In this embodiment, the BiGRU-enhanced Transformer network includes a pre-input BiGRU layer, a Transformer encoder, a BiGRU layer between the encoder and the decoder, and a decoder. like Figure 2As shown, the parallel processing flow for the two branches is the same. Taking one branch as an example, it specifically includes: The baseline power input sequence is first fed into a pre-stage BiGRU layer to extract short-time local bidirectional dependencies; In this embodiment, the BiGRU layer is based on bidirectional GRU units. The update gates, reset gates, and candidate hidden states of the forward GRU and backward GRU are calculated separately and then concatenated to obtain the final hidden representation. The gating mechanism formula for a single GRU unit is as follows: , , , , In the formula, x t Indicates the current time step t The input feature vector, This represents the hidden state vector from the previous time step. z t and r t These represent the activation outputs of the update door and the reset door, respectively. Indicates the candidate hidden state. h t Indicates the final hidden state. , , , , and Represents the weight matrix. , , For bias vectors, It is the Sigmoid activation function. This indicates element-wise multiplication; The forward and backward hidden states are concatenated to output a high-dimensional hidden feature. This hidden feature is then linearly added to a fixed-position code based on a sine-cosine function and sent to the decoder of the Transformer model. The formula for calculating the position code is as follows: , , In the formula, pos Indicates the time step position. i For dimensional indexing, d Represents the encoding dimension. PE (pos,2i) and PE (pos,2i+1)These represent the positional encoding values ​​of the input sequence in the odd-even feature dimension, respectively. This design ensures that the positional encoding exhibits different frequency characteristics in different dimensions, enabling the model to perceive the temporal order of the input sequence. The Transformer encoder performs multi-head self-attention global dependency modeling, and the attention calculation formula is: , in, Q , K and V These represent the query matrix, key matrix, and value matrix, respectively. dk The feature dimension of the key matrix is ​​represented by Attention. Q , K , V () represents the final attention output. The multi-head attention mechanism maps the input to multiple subspaces to achieve parallel computation. , , in, This represents the linear projection matrix of different heads, while head i This represents the output mapping matrix. W O This represents the final mapping matrix used after concatenating the outputs of all independent attention heads and then performing a linear fusion. The global semantic features output by the encoder are further refined by the inter-layer BiGRU layer to alleviate the information loss and semantic gap caused by long sequence self-attention. The refined output is used as a high-order historical context feature for later use. Finally, in the decoding stage, the future weather forecast features for the period to be predicted are fused with the location code and used as the initial input to the decoder. The future expected features are then extracted through a masked multi-head self-attention mechanism to generate a query vector. Q Subsequently, the query vector and the historical context features output by the inter-layer BiGRU layer are used as the key matrix. K Sum matrix V The encoder-decoder cross-attention layer performs correlation matching and information fusion, and the decoder finally generates preliminary prediction sequences for the baseline branch and the ramp branch through autoregression.

[0031] This invention uses a pre-processed BiGRU to extract local transient dependencies in high-resolution sequences by splicing forward and backward hidden states, thereby enhancing the perception of sudden fluctuations. Combined with sinusoidal position coding, the data enters the Transformer encoder, which uses a multi-head self-attention mechanism to capture long-range correlations lasting several days. An inter-layer BiGRU is embedded between the encoder and decoder to refine the global semantics of the encoder output, thereby mitigating semantic loss in the generation of long sequences. The baseline branch outputs the initial power sequence; the ramp branch then optimizes the loss function through a dedicated multi-task joint optimization. Loss The training aims to simultaneously optimize event classification, time localization, and amplitude regression. Its formula is defined as: , in, L pred The weighted cross-entropy loss is designed for slope occurrence indicators to address class imbalance caused by fewer climbing events. L time The loss is the prediction error at the start and end time points. L amp This represents the prediction error loss for the slope power variation amplitude. In this embodiment, the optimal combination of weighting coefficients is determined by performing a grid search on the validation set. , , This ensures that the model maximizes slope capture rate while maintaining baseline accuracy.

[0032] This loss function introduces a weighted cross-entropy loss. L pred To solve the problem of imbalanced samples, utilize L 1-norm constraint on slope start and end time deviation L time and slope amplitude deviation L amp Furthermore, by using adjustable weight coefficients to achieve collaborative optimization among subtasks, the model's ability to capture extreme instantaneous fluctuations is significantly improved.

[0033] Step 4: Nonlinearly fuse the baseline predicted power and the ramp-up predicted power to generate the final high-resolution long-term photovoltaic power prediction result.

[0034] Using the baseline power prediction sequence, slope event prediction results, and meteorological characteristics as input, the LightGBM model is used to predict the residual between the baseline prediction and the actual power. In this embodiment, the LightGBM model is used to nonlinearly fuse the prediction results of the baseline branch and the prediction results of the climbing branch. The LightGBM model takes the baseline prediction results, climbing event characteristics, key meteorological indicators, lag power characteristics and time index as input feature vectors and feeds them into the LightGBM model to learn the nonlinear residual mapping. LightGBM is a gradient boosting decision tree, and its regression loss function is:

[0035] in N For the sample size, l It is a differentiable convex loss function used to measure the true value. y i Compared with the current set of predicted values The difference between them, and This is a regularization term used to control model complexity; LightGBM is explicitly trained with the goal of predicting the true residual error, which is defined as the actual observed power. P actual Compared with the baseline forecast P baseline Deviation between: , Therefore, the final fusion prediction value P final This is the sum of the baseline prediction and the LightGBM prediction residuals: , , By leveraging LightGBM's nonlinear gradient boosting decision tree fusion, context-aware intelligent residual correction is achieved, further enhancing the accuracy of hill-climbing event capture and overall robustness.

[0036] Slope prediction branch P ramp The output is not used directly as an addition term, but rather incorporated as a core feature into the meta-feature vector of LightGBM. X meta ; P ramp and P baseline The lag characteristics and time index together constitute a comprehensive X meta ; LightGBM learns intelligent correction rules by leveraging its robust non-linear decision tree structure.

[0037] The system adopts a 7-day rolling forecast strategy, which uses historical data to predict the 1-minute resolution power for the next 7 days. By continuously executing this strategy, a high-resolution long-term forecast for the next month is finally completed.

[0038] Analysis of photovoltaic power prediction results: To verify the superiority of the proposed dual-branch architecture in long-term high-resolution photovoltaic power prediction accuracy, four models—SVR, Transformer, LSTM-Transformer, and Informer—were constructed for comparison. Taking the prediction results as of October 22, 2024, as an example, the prediction accuracy of the five models was compared. Figure 3 As shown in Table 1, the error indicators for the entire October forecast are as follows.

[0039] Table 1 Comparison of prediction errors of different models

[0040] In the actual measurement of this photovoltaic power plant, the proposed dual-branch prediction architecture achieved a MAE of 2.32MW and an RMSE of 2.31MW, significantly outperforming other mainstream prediction models. Experiments show that this invention effectively overcomes the oversmoothing effect of deep learning models, demonstrating extremely high capture accuracy and robustness in long-term prediction of one month's worth of data at 1-minute resolution.

[0041] Example 2 This embodiment provides a high-resolution long-term photovoltaic power prediction system based on a dual-branch architecture, including: The data preprocessing module is used to acquire historical photovoltaic power generation data and corresponding meteorological data, and to preprocess the historical photovoltaic power generation data to obtain a photovoltaic power generation sequence. The feature extraction module is used to feed the power sequence into the baseline branch and the ramp branch in parallel for processing; in the baseline branch, the power sequence is decomposed into multiple intrinsic mode functions; in the ramp branch, the ramp event sequence and ramp flags are extracted. The branch power prediction module is used to input the results of the dual-branch processing in parallel into the trained power prediction model to obtain the baseline predicted power and the ramp-up predicted power; the construction process of the power prediction model includes: The features extracted from each branch are concatenated with meteorological data and time-coded features to obtain the multi-channel input embedding of each branch; the input embeddings of the two branches are fed into a shared BiGRU enhanced Transformer network for prediction, and the baseline power prediction sequence and the probability, start and end time and amplitude prediction of the slope event are obtained respectively. The power prediction output module is used to nonlinearly fuse the baseline predicted power and the ramp-up predicted power to generate the final high-resolution long-term photovoltaic power prediction result.

[0042] It should be noted that the specific implementation of the high-resolution long-term photovoltaic power prediction system based on the dual-branch architecture in this embodiment of the invention is similar to the specific implementation of the high-resolution long-term photovoltaic power prediction method based on the dual-branch architecture in this embodiment of the invention. For details, please refer to the description in the method section. To reduce redundancy, it will not be repeated here.

[0043] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture as described above.

[0044] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture as described above.

[0045] 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 hardware embodiments, software embodiments, or embodiments 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 and optical storage) containing computer-usable program code.

[0046] 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, and 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0047] 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.

[0048] These computer program instructions may 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.

[0049] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0050] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture, characterized in that, Includes the following steps: Acquire historical photovoltaic power generation data and corresponding meteorological data, and preprocess the historical photovoltaic power generation data to obtain a photovoltaic power generation sequence; The power sequence is fed into the baseline branch and the ramp branch in parallel for processing; in the baseline branch, the power sequence is decomposed into multiple intrinsic mode functions; in the ramp branch, the ramp event sequence and ramp flags are extracted. The results of the dual-branch processing are input in parallel into the trained power prediction model to obtain the baseline predicted power and the ramp predicted power; the construction process of the power prediction model includes: The features extracted from each branch are concatenated with meteorological data and time-coded features to obtain the multi-channel input embedding of each branch; the input embeddings of the two branches are fed into a shared BiGRU enhanced Transformer network for prediction, and the baseline power prediction sequence and the probability, start and end time and amplitude prediction of the slope event are obtained respectively. The baseline predicted power and the ramp-up predicted power are nonlinearly fused to generate the final high-resolution long-term photovoltaic power prediction result.

2. The high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture as described in claim 1, characterized in that, The preprocessing of historical power generation data includes outlier correction, which includes outlier handling for three different scenarios.

3. The high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture as described in claim 1, characterized in that, In the baseline branch, the power sequence is decomposed into multiple intrinsic mode functions, including MF1-IMF7. MF1 characterizes the low-frequency baseline trend of daily operation determined by the solar altitude angle, capturing long-term trends. IMF2-IMF4 capture minute-level mid-frequency fluctuations caused by cloud movement. IMF5-IMF7 lock in subtle high-frequency noise and random disturbances, capturing high-frequency disturbances.

4. The high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture as described in claim 1, characterized in that, The process of embedding the inputs of the two branches into a shared BiGRU-enhanced Transformer network for prediction includes the following steps for each branch: The input is embedded into the front BiGRU layer to extract high-dimensional hidden features. The hidden features are linearly added to the fixed-position encoding based on the sine-cosine function and then sent to the decoder of the Transformer model. The Transformer encoder performs multi-head self-attention global dependency modeling to obtain global semantic features. The global semantic features are further refined by the inter-layer BiGRU layer and the output is used as high-order historical context features. The future weather forecast features for the period to be predicted are fused with location codes and used as the initial input to the decoder. A masked multi-head self-attention mechanism is then used to extract future expected features and generate a query vector. Q Subsequently, the query vector and higher-order historical context features are used as the key matrix. K Sum matrix V The corresponding branch prediction sequence is obtained by performing association matching and information fusion in the encoder-decoder cross-attention layer.

5. The high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture as described in claim 4, characterized in that, The formula for calculating the location code is: ， ， In the formula, and These represent the positional encoding values ​​of the input sequence in the odd / even feature dimension, respectively. Indicates the time step position. i For dimensional indexing, d Represents the encoding dimension.

6. The high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture as described in claim 1, characterized in that, The ramp branch then uses a dedicated multi-task joint optimization loss function. Loss Training is defined as: ， in, L pred The weighted cross-entropy loss is designed for slope occurrence indicators to address class imbalance caused by fewer climbing events. L time The loss is the prediction error at the start and end time points. L amp This is the prediction error loss for the slope power change amplitude. , , These are the weighting coefficients.

7. The high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture as described in claim 1, characterized in that, The method involves nonlinearly fusing the baseline predicted power and the slope predicted power, using the baseline power prediction sequence, slope event prediction results, and meteorological characteristics as inputs, and employing the LightGBM model to predict the residual between the baseline prediction and the actual power. The fused prediction value is the sum of the baseline prediction value and the LightGBM prediction residual.

8. A high-resolution long-term photovoltaic power prediction system based on a dual-branch architecture, characterized in that, include: The data preprocessing module is used to acquire historical photovoltaic power generation data and corresponding meteorological data, and to preprocess the historical photovoltaic power generation data to obtain a photovoltaic power generation sequence. The feature extraction module is used to feed the power sequence into the baseline branch and the ramp branch in parallel for processing; in the baseline branch, the power sequence is decomposed into multiple intrinsic mode functions; in the ramp branch, the ramp event sequence and ramp flags are extracted. The branch power prediction module is used to input the results of the dual-branch processing in parallel into the trained power prediction model to obtain the baseline predicted power and the ramp-up predicted power; the construction process of the power prediction model includes: The features extracted from each branch are concatenated with meteorological data and time-coded features to obtain the multi-channel input embedding of each branch; the input embeddings of the two branches are fed into a shared BiGRU enhanced Transformer network for prediction, and the baseline power prediction sequence and the probability, start and end time and amplitude prediction of the slope event are obtained respectively. The power prediction output module is used to nonlinearly fuse the baseline predicted power and the ramp-up predicted power to generate the final high-resolution long-term photovoltaic power prediction result.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the high-resolution long-term photovoltaic power prediction method based on a dual-branch architecture as described in any one of claims 1-7.

Images