High-precision ultra-short-term photovoltaic power prediction method based on MSCT-RCM model

Abstract

The invention relates to a high-precision ultra-short-term photovoltaic power prediction method based on an MSCT-RCM model. The method comprises the steps of data preprocessing and feature enhancement; an MSCT-RCM model is constructed; determining model structure parameters and training the model; and outputting a prediction result. According to the method, an MSCT-RCM model fusing a multi-scale convolution branch, a Transform backbone network and a residual error correction module is provided, local dynamic features are efficiently extracted through multi-scale convolution in the model, a Transform encoder modeling long time sequence depends on generation of baseline prediction, the residual error correction module adaptively optimizes a prediction result through historical error information, and the prediction accuracy is improved. And the problem of error accumulation in multi-step prediction is effectively relieved. And then, a two-stage training strategy is provided for the MSCT-RCM model, so that the multi-step prediction precision and stability of the MSCT-RCM model can be further improved, and finally, the high-precision prediction of the photovoltaic power can be realized. Compared with the prior art, the MSCT-RCM model provided by the invention is suitable for real-time deployment, and can provide reliable technical support for intelligent power grid dispatching, energy management and photovoltaic grid-connected operation.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power prediction technology, and specifically to a high-precision ultra-short-term photovoltaic power prediction method based on the MSCT-RCM model. Background Technology

[0002] With the rapid global deployment of photovoltaic (PV) systems, PV power forecasting has become a key technology for ensuring stable grid operation, optimizing energy dispatch, and supporting electricity market transactions. Due to rapid changes in meteorological conditions and geographical differences, PV output power exhibits significant randomness and intermittency, which introduces considerable uncertainty into power system operation.

[0003] Regarding ultra-short-term photovoltaic (PV) power prediction technology, it not only provides power plants and the grid with high-precision forward-looking information, rapidly predicting PV power fluctuations and facilitating the early deployment of adjustment measures to alleviate decision-making pressure and reduce risks such as grid frequency and voltage fluctuations, but also helps power plants and the grid more accurately match power generation and consumption demands, optimize the configuration of backup capacity such as energy storage, reduce curtailment losses, and improve economic efficiency. Furthermore, this technology can assist power plants in optimizing maintenance and cleaning plans, reducing investment and operating costs. Therefore, with the continuous expansion of PV power generation, ultra-short-term PV power prediction technology has become one of the key technologies for my country's energy structure transformation. Advances in this technology will contribute to achieving carbon peaking and carbon neutrality goals, as well as the construction of new power systems.

[0004] In existing technologies, Chinese patent document CN 120450104 A discloses a photovoltaic power generation prediction method based on the Transformer model, and Chinese patent document CN 118467940 A discloses a photovoltaic power generation prediction method based on convolutional neural networks and Transformer. Both of these methods utilize deep learning, demonstrating the significant advantages of deep learning methods in photovoltaic power prediction in recent years. Models based on attention mechanisms and the Transformer architecture exhibit strong capabilities in capturing long-term temporal dependencies, uncovering complex feature interactions, and significantly improving prediction accuracy. Furthermore, a hybrid model combining CNN and Transformer can further enhance photovoltaic power prediction performance; however, this hybrid model still has shortcomings in model generalization ability, real-time computational efficiency, and robustness under extreme weather conditions.

[0005] Meanwhile, considering the multi-scale characteristics of photovoltaic power data, signal decomposition methods such as EMD or VMD can help photovoltaic power prediction models capture time-series features more clearly by separating different frequency components. However, these decomposition methods often lead to problems such as high parameter sensitivity, increased computational complexity, and information leakage if validation is insufficient. Residual connections, attention enhancement modules, and multi-scale convolution operations can enhance the prediction model's response to high-frequency fluctuations and local anomalies. Using historical prediction residuals or error information as auxiliary inputs helps alleviate the error accumulation problem in multi-step predictions. Furthermore, to balance model prediction accuracy with practical deployment needs, strategies such as lightweight network design, parameter sharing mechanisms, or model compression have gradually gained research attention.

[0006] Therefore, developing a high-precision ultra-short-term photovoltaic power prediction scheme has become a key technical problem that urgently needs to be solved in the field of photovoltaic power generation. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the purpose of this invention is to provide a high-precision ultra-short-term photovoltaic power prediction method based on the MSCT-RCM model, which can achieve high-precision, ultra-short-term photovoltaic power prediction.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] A high-precision ultra-short-term photovoltaic power prediction method based on the MSCT-RCM model, the method includes the following steps:

[0010] S1. Obtain raw photovoltaic power and meteorological data, perform preprocessing and feature enhancement on the data, and generate an input feature matrix;

[0011] S2. Construct the MSCT-RCM model; The MSCT-RCM model is used to extract multi-scale convolutional features from the input feature matrix, fuse the multi-scale convolutional features with the original input features, generate baseline prediction results by modeling global dependencies of the sequence, and perform residual correction on the baseline prediction results based on historical residual information.

[0012] S3. Determine the network structure parameters of the MSCT-RCM model and train the model using a two-stage training strategy;

[0013] S4. Input the input feature matrix to be predicted into the trained MSCT-RCM model, and generate high-precision ultra-short-term photovoltaic power prediction results by fusing the baseline prediction results with the residual correction output.

[0014] As a further improvement to the above technical solution, step S1 specifically includes:

[0015] S11. Obtain raw photovoltaic power and meteorological data to ensure that the data covers the time series range and feature dimensions required for forecasting;

[0016] S12. Clean the data and handle outliers;

[0017] S13. Enhance the multi-scale features and select key features;

[0018] S14. Normalization and Dataset Partitioning: The Min–Max normalization method is used to map all input features in the initial input feature matrix to the [0,1] interval to generate the final input feature matrix.

[0019] The normalization formula is:

[0020] (1);

[0021] Where x is the original value of the feature, x min For the eigenvalue minimum, x max For the eigenvalue, x norm For the normalized eigenvalues ​​after applying Min–Max normalization;

[0022] S15. The sliding window method is used to divide the final input feature matrix into a fixed-length input sequence and a corresponding target output sequence, and outputs a set of paired samples of the input sequence and the target output sequence.

[0023] As a further improvement to the above technical solution, step S12 specifically includes:

[0024] S121. For records with a missing proportion lower than the set threshold, use linear interpolation to complete them; S122. Identify, remove or correct outliers with unreasonable physical meaning.

[0025] As a further improvement to the above technical solution, step S13 specifically includes:

[0026] S131. Empirical mode decomposition is used to perform multi-scale decomposition on the time series of each numerical feature, decomposing the time series x(t) into several intrinsic mode functions and residual terms;

[0027] The formula for the multi-scale decomposition is:

[0028] (2);

[0029] in, represents the intrinsic mode function; i is the index of the IMF, ranging from 1 to n, where n is the total number of IMFs obtained from the decomposition; For residual terms; This represents the original value of the time series at time point t;

[0030] S132. Based on the multi-scale decomposition results, extract the first two IMF components as enhancement features, and concatenate them with the original input features to form the input feature matrix.

[0031] S133. Use gradient boosting regression to rank all enhanced features by importance, and select the core features that rank within a set range as input variables.

[0032] As a further improvement to the above technical solution, the MSCT-RCM model includes a multi-scale convolutional branch, a Transformer backbone network, and a residual correction module.

[0033] The multi-scale convolution branch is used to extract multi-scale convolution features from the input feature matrix;

[0034] The Transformer backbone network is used to fuse multi-scale convolutional features with original input features and generate baseline prediction results by modeling global dependencies of sequences.

[0035] The residual correction module is used to correct the baseline prediction results based on historical residual information.

[0036] As a further improvement to the above technical solution, step S2 specifically includes the following steps:

[0037] S21. Multi-scale convolutional branch feature extraction: Based on the convolutional neural network (CNN), multiple parallel one-dimensional convolutional branches are designed to extract multi-scale convolutional features from the input sequence.

[0038] S22. Construction of the backbone network: Multi-scale convolutional features are fused with the original input features. The fused features are then adapted by linear projection and positional encoding, and then processed by a multi-head self-attention mechanism to model the global dependencies of the sequence, finally generating the baseline prediction results.

[0039] S23. Residual Correction Module Construction: Combining historical residual information, the baseline prediction results are dynamically corrected, and the interaction between the baseline prediction results and historical residual information is captured by a lightweight Transformer encoder to iteratively optimize the prediction results.

[0040] As a further improvement to the above technical solution, step S21 specifically includes:

[0041] S211. Define the input sequence using the following formula:

[0042] (3);

[0043] Where X represents the original input sequence matrix, Let L represent the feature vector at each time step in the sequence, and F represent the feature dimension.

[0044] S212. Calculate the features of each convolutional branch using the following formula to obtain multi-branch features:

[0045] (4);

[0046] Among them, H k This is the feature matrix extracted by the k-th convolutional kernel, where K is the kernel size and ReLU is the modified linear unit activation function. This is a one-dimensional convolution operation;

[0047] S213. Use the following formula to concatenate the multi-branch features to form multi-scale convolutional features:

[0048] (5);

[0049] in, For multi-scale convolutional features, These correspond to the outputs of convolutional branches with kernel sizes of 3, 5, and 7, respectively. Concat represents the concatenation operation, L represents the length of the input sequence, and C represents the number of convolutional output channels for each branch.

[0050] As a further improvement to the above technical solution, step S22 specifically includes:

[0051] S221. Use the following formula to convert the multi-scale convolutional features H... cnn The features are then fused with the original input features to generate fused features:

[0052] (6);

[0053] in, Represents the fused feature matrix. Represents multi-scale convolutional features. Let L represent the original input sequence matrix, F represent the feature dimension, and C represent the number of convolution output channels for each branch.

[0054] S222. Using the following formula, the fused features are mapped to the hidden space through linear projection to generate the projected features:

[0055] (7);

[0056] in, This represents the projected feature matrix. Represents the fused feature matrix. and Here, are the projection weight matrix and the bias vector, respectively, and L represents the length of the input sequence. The hidden layer dimension of the Transformer;

[0057] S223. Use the following formula to superimpose learnable positional codes onto the projected features to generate features with positional codes:

[0058] (8);

[0059] in, This is the encoded feature matrix. PE represents the projected feature matrix, where PE is the learnable location encoding matrix.

[0060] S224. Input the position-encoded features into an N-layer Transformer encoder using the following formula, and obtain the global temporal feature representation through a multi-head self-attention mechanism:

[0061] (9);

[0062] Where Z represents the output matrix of the Transformer encoder, and TransformerEncoder is the Transformer encoder module. The encoded feature matrix is ​​shown, where L represents the length of the input sequence. The hidden layer dimension of the Transformer;

[0063] S225. Take the hidden state vector Z[−1] of the last time step output of the Transformer encoder, and use the following formula to generate the baseline prediction result of photovoltaic power through mapping of the fully connected layer:

[0064] (10);

[0065] in, For the baseline prediction result vector, This is the weight matrix of the fully connected layer. This is the bias vector for the fully connected layer. Output the hidden state vector for the last step of the Transformer, where H represents the prediction step size.

[0066] As a further improvement to the above technical solution, step S23 specifically includes:

[0067] S231. Define the historical residual vector of the current sample as:

[0068] (11);

[0069] in, Represents the historical residual vector. H represents the residual corresponding to the prediction step size;

[0070] S232, Residual Correction Process

[0071] S2321, Residual Coding: Historical residual information is mapped to token representation using the following formula to adapt to Transformer coding:

[0072] (12);

[0073] in, The token representation after historical residual mapping. Represents the projection function. H represents the historical residual vector, and H represents the prediction step size. The hidden layer dimension of the Transformer;

[0074] S2322, Sequence Concatenation: Concatenate the sequences from the baseline prediction tokens... The first token is concatenated with the residual token to form the input sequence:

[0075] (13);

[0076] in, This represents the concatenated input sequence. For the baseline prediction result vector, The token representation is the result of mapping historical residuals, where H represents the prediction step size. The hidden layer dimension of the Transformer;

[0077] S2333, Lightweight Transformer Encoding: Based on a single-layer Transformer encoder, it uses the following formula to capture the interaction between baseline prediction results and historical residual information:

[0078] (14);

[0079] in, This represents the output of the lightweight Transformer encoder. For the Transformer encoder module, This represents the concatenated input sequence;

[0080] S2334, Residual Generation: The residual prediction result is obtained by mapping the first token through a multilayer perceptron using the following formula:

[0081] (15);

[0082] in, This is the vector of residual prediction results. This is the weight matrix of the residual generation layer. The bias vector of the residual generation layer. This is the first token vector output by the lightweight Transformer;

[0083] S2335. Prediction Output: The residual prediction is added to the baseline prediction result using the following formula to obtain the final prediction result:

[0084] (16);

[0085] in, This is the final prediction result vector. For the baseline prediction result vector, This is the vector of residual prediction results.

[0086] As a further improvement to the above technical solution, step S3 specifically includes the following steps:

[0087] S31. Stage I involves training the baseline model, specifically the Transformer backbone network, to provide reliable baseline power predictions for residual correction. In Stage I, the optimization objective is designed to minimize the baseline predictions. The mean square error between the true value y and the actual value y is given by the following formula:

[0088] (17);

[0089] in, This represents the mean squared error loss for stage I. For the true value of sample i, represents the baseline prediction result vector, where N is the number of samples;

[0090] S32, Phase II is end-to-end joint training, and the optimization objective is as follows:

[0091] (18);

[0092] in, The mean squared error loss for stage II. This is the final prediction result vector. Let be the true value of sample i, and N be the number of samples.

[0093] Phase II enables the proposed MSCT-RCM model to adaptively utilize historical residual information through end-to-end joint training, achieving iterative optimization of baseline output and residual correction, and significantly improving the robustness of multi-step prediction.

[0094] Compared with the prior art, the advantages of the present invention are:

[0095] (1) This invention proposes a hybrid architecture model that combines multi-scale convolutional neural networks (CNN) and Transformer models, which can simultaneously capture local temporal dynamics and global dependencies, and realize high-precision ultra-short-term photovoltaic power prediction at multiple time scales.

[0096] (2) In order to solve the problem of error accumulation in multi-step prediction of hybrid architecture model, this invention designs a lightweight residual correction module. This module adjusts the baseline prediction results by combining historical error data, which improves the long-term prediction accuracy and stability of the proposed model, especially showing stronger robustness under complex meteorological conditions.

[0097] (3) The MSCT-RCM model proposed in this invention and its two-stage training strategy ensure that it achieves the research goal of high-precision ultra-short-term photovoltaic power prediction, and is expected to play an important role in smart grid dispatch and energy management of photovoltaic power plants. Future research can further optimize the computational efficiency of the model, improve its generalization ability under more complex meteorological conditions, and explore more feature extraction techniques to enhance the robustness of the model. The two-stage training strategy first independently trains the baseline model that integrates CNN and Transformer to achieve robust multi-scale feature extraction; then, based on loading the pre-trained weights, it is jointly optimized with the designed residual correction module to alleviate the gradient explosion problem, accelerate the convergence speed, and significantly improve the multi-step prediction accuracy and stability of the model. Attached Figure Description

[0098] Figure 1 This is a flowchart illustrating the method sequence of the high-precision ultra-short-term photovoltaic power prediction method based on the MSCT-RCM model in this invention.

[0099] Figure 2 This is a diagram of the architecture of the MSCT-RCM model in this invention;

[0100] Figure 3 The graph shows the prediction performance results for different numbers of features in this invention.

[0101] Figure 4 The graph shows the prediction performance results under different input sequences in this invention.

[0102] Figure 5 This is a graph showing the performance comparison results under different prediction durations in this invention;

[0103] Figure 6 This is a comparison chart of the predicted and actual values ​​of photovoltaic power generation at different prediction times in this invention. Detailed Implementation

[0104] To provide a better understanding of the structural features and effects achieved by the present invention, a detailed description is provided below, accompanied by preferred embodiments and accompanying drawings:

[0105] To achieve high-precision ultra-short-term photovoltaic power prediction, this invention proposes an ultra-short-term photovoltaic power prediction scheme that integrates multi-scale convolutional neural networks, Transformer, and residual correction. The design idea is to first use multi-scale CNN to extract local temporal features, and then use Transformer backbone network to model long-term temporal dependencies to generate high-quality baseline prediction results. On this basis, a lightweight residual correction module is introduced to iteratively optimize the baseline prediction results by combining historical residual information.

[0106] like Figure 1 and Figure 2 The method shown is a high-precision ultra-short-term photovoltaic power prediction method based on the MSCT-RCM model. The method includes the following steps:

[0107] S1. Obtain raw photovoltaic power and meteorological data, perform preprocessing and feature enhancement on the data, and generate the input feature matrix.

[0108] S2. Construct the MSCT-RCM model; The MSCT-RCM model is used to extract multi-scale convolutional features from the input feature matrix, fuse the multi-scale convolutional features with the original input features, generate baseline prediction results by modeling global dependencies of the sequence, and perform residual correction on the baseline prediction results based on historical residual information.

[0109] S3. Determine the network structure parameters of the MSCT-RCM model and train the model using a two-stage training strategy.

[0110] S4. Input the input feature matrix to be predicted into the trained MSCT-RCM model, and generate high-precision ultra-short-term photovoltaic power prediction results by fusing the baseline prediction results with the residual correction output.

[0111] As a further improvement to the above technical solution, step S1 specifically includes:

[0112] S11. Obtain raw photovoltaic power and meteorological data to ensure that the data covers the time series range and feature dimensions required for forecasting.

[0113] The raw photovoltaic power is the original output power record collected in real time by the photovoltaic power station; the meteorological data includes the original records of meteorological parameters related to photovoltaic power generation, such as air temperature, cloud opacity, dew point temperature, direct radiation, diffuse radiation, and global horizontal radiation. After cleaning, outlier processing, and normalization of the raw photovoltaic power and meteorological data, a continuous, complete, and high-quality input feature matrix is ​​formed, providing a reliable data foundation for subsequent modeling.

[0114] S12. Clean the data and handle outliers.

[0115] S121. For records with a missing rate of less than 0.5%, linear interpolation is used to complete them, in order to preserve the continuity of the time series to the greatest extent and avoid interference from data breaks in capturing time series features. S122. Outliers with unreasonable physical meaning are identified, removed, or corrected to avoid the impact of physically unreasonable values ​​on power prediction results. These outliers include, but are not limited to, power exceeding the rated capacity of the photovoltaic power station and power records less than zero. By removing or correcting outliers, their impact on the accuracy of power prediction results is avoided.

[0116] S13, Multi-scale feature enhancement and key feature selection.

[0117] S131. To enhance the multi-scale expressive power of features, empirical mode decomposition (EMD) is used to decompose the time series of each numerical feature into multiple scales. Based on EMD, the time series x(t) is decomposed into several intrinsic mode functions and residual terms.

[0118] The formula for the multi-scale decomposition is:

[0119] (1);

[0120] in, represents the intrinsic mode function; i is the index of the IMF, ranging from 1 to n, where n is the total number of IMFs obtained from the decomposition; For the residual term, This represents the original value of the time series at time point t.

[0121] S132. Based on the multi-scale decomposition results, extract the first two IMF components as enhanced features, and concatenate them with the original input features to form the input feature matrix.

[0122] The first two IMF components capture short-term fluctuations and medium-term trends. This invention selects them as enhancement features and concatenates them with the original input to form a preliminary input feature matrix.

[0123] S133. Gradient boosting regression is used to rank the importance of all enhanced features, and the top 10 core features are selected as input variables. This aims to reduce redundant information and improve the model training efficiency and generalization ability.

[0124] S14. Normalization and Dataset Partitioning

[0125] To eliminate the influence of different units on model training, the Min-Max normalization method is used to map all input features in the initial input feature matrix to the [0,1] interval, generating the final input feature matrix; the normalization formula is:

[0126] (2);

[0127] Where x is the original value of the feature, x min For the eigenvalue minimum, x max For the eigenvalue, x norm This refers to the normalized eigenvalues ​​obtained by applying Min–Max normalization.

[0128] S15. The sliding window method is used to divide the final input feature matrix into a fixed-length input sequence and a corresponding target output sequence, and output a set of paired samples of the input sequence and the target output sequence to adapt to the multi-step prediction requirements of the model, while ensuring the temporal consistency of the training and prediction process.

[0129] As a further improvement to the above technical solution, the MSCT-RCM model includes a multi-scale convolutional branch, a Transformer backbone network, and a residual correction module. The multi-scale convolutional branch is used to extract multi-scale convolutional features from the input feature matrix; the Transformer backbone network is used to fuse the multi-scale convolutional features with the original input features, generating baseline prediction results by modeling global dependencies of the sequence; and the residual correction module is used to perform residual correction on the baseline prediction results based on historical residual information.

[0130] As a further improvement to the above technical solution, step S2 specifically includes the following steps:

[0131] S21. Multi-scale convolutional branch feature extraction: Based on the convolutional neural network (CNN), multiple parallel one-dimensional convolutional branches are designed to extract multi-scale convolutional features from the input sequence.

[0132] Ultra-short-term photovoltaic (PV) power sequences contain multi-timescale dynamic features, including short-term fluctuations (such as rapid cloud movement) and medium- to long-term trends (such as changes in solar radiation). To capture these multi-timescale features simultaneously during feature extraction, three parallel one-dimensional convolutional branches were designed based on a convolutional neural network (CNN). The kernel sizes k were set to 3, 5, and 7, respectively, to adapt to the multi-timescale characteristics of the PV power sequences. Specifically, the small kernel (k=3) captures short-term fluctuation features, such as power abrupt changes caused by rapid cloud movement; the medium kernel (k=5) captures medium-term trend features, such as intraday intensity variations; and the large kernel (k=7) captures long-term pattern features, such as seasonally related solar radiation patterns. After extracting local dynamic features at corresponding scales through convolutional operations in each branch, the output features of all branches are concatenated and fused, providing rich contextual information to support the Transformer backbone network.

[0133] Step S21 specifically includes:

[0134] S211. Define the input sequence using the following formula:

[0135] (3);

[0136] Where X represents the original input sequence matrix, Let L represent the feature vector at each time step in the sequence, and F represent the feature dimension.

[0137] S212. Calculate the features of each convolutional branch using the following formula to obtain multi-branch features:

[0138] (4);

[0139] Among them, H k This is the feature matrix extracted by the k-th convolutional kernel, where K is the kernel size and ReLU is the modified linear unit activation function. This is a one-dimensional convolution operation;

[0140] S213. Use the following formula to concatenate the multi-branch features to form multi-scale convolutional features:

[0141] (5);

[0142] in, For multi-scale convolutional features, These correspond to the outputs of convolutional branches with kernel sizes of 3, 5, and 7, respectively. Concat represents the concatenation operation, L represents the length of the input sequence, and C represents the number of convolutional output channels for each branch.

[0143] The multi-scale convolutional branch design of the MSCT-RCM model aims to preserve local dynamic features while providing richer contextual feature representations for the Transformer backbone network in the proposed model.

[0144] S22. Construction of the backbone network: Multi-scale convolutional features are fused with the original input features. The fused features are then adapted by linear projection and positional encoding, and then processed by a multi-head self-attention mechanism to model the global dependencies of the sequence, finally generating the baseline prediction results.

[0145] To capture the global dependencies of the input sequence and improve the multi-step prediction capability of the MSCT-RCM model, a Transformer backbone network is introduced into the MSCT-RCM model proposed in this invention.

[0146] Step S22 specifically includes:

[0147] S221. Use the following formula to convert the multi-scale convolutional features H... cnn The features are then fused with the original input features to generate fused features:

[0148] (6);

[0149] in, Represents the fused feature matrix. Represents multi-scale convolutional features. Let L represent the original input sequence matrix, F represent the feature dimension, and C represent the number of convolution output channels for each branch.

[0150] S222. Using the following formula, the fused features are mapped to the hidden space through linear projection to generate the projected features:

[0151] (7);

[0152] in, This represents the projected feature matrix. Represents the fused feature matrix. and Here, are the projection weight matrix and the bias vector, respectively, and L represents the length of the input sequence. The hidden layer dimension of the Transformer;

[0153] S223. Use the following formula to superimpose learnable positional codes onto the projected features to generate features with positional codes:

[0154] (8);

[0155] in, This is the encoded feature matrix. PE represents the projected feature matrix, where PE is the learnable location encoding matrix.

[0156] S224. Input the position-encoded features into an N-layer Transformer encoder using the following formula, and obtain the global temporal feature representation through a multi-head self-attention mechanism:

[0157] (9);

[0158] Where Z represents the output matrix of the Transformer encoder. For the Transformer encoder module, The encoded feature matrix is ​​shown, where L represents the length of the input sequence. The hidden layer dimension of the Transformer;

[0159] S225. Take the hidden state vector Z[-1] of the last time step output of the Transformer encoder, and use the following formula to generate the baseline prediction result of photovoltaic power through mapping of the fully connected layer:

[0160] (10);

[0161] in, For the baseline prediction result vector, This is the weight matrix of the fully connected layer. This is the bias vector for the fully connected layer. Output the hidden state vector for the last step of the Transformer, where H represents the prediction step size.

[0162] In step S22, the fused features output from the multi-scale convolutional branches are first concatenated with the original input features to form a comprehensive feature that simultaneously possesses local details and original information. Then, the comprehensive feature is mapped to the hidden space of the adapted Transformer through linear projection, and the positional encoding of the science system is superimposed to supplement the temporal positional information. Finally, the encoded features are input into a multi-layer Transformer encoder, and the global interaction relationship between features is mined through a multi-head self-attention mechanism to model the global dependency relationship of the sequence, while generating the photovoltaic power baseline prediction result.

[0163] S23. Residual Correction Module Construction: Combining historical residual information, the baseline prediction results are dynamically corrected, and the interaction between the baseline prediction results and historical residual information is captured by a lightweight Transformer encoder to iteratively optimize the prediction results.

[0164] Considering that errors accumulate gradually during multi-step prediction, leading to a decrease in prediction accuracy after multiple prediction steps, this invention designs a lightweight correction mechanism by incorporating historical prediction residual information. The historical residual vector is mapped to an adapted encoded token representation using a projection function, and this token representation is concatenated with the baseline prediction result to form an input sequence. A single-layer lightweight Transformer encoder captures the dynamic interaction between the baseline prediction and historical residuals, and the encoded features are processed to generate residual prediction values. The residual prediction values ​​are then superimposed with the baseline prediction results, thereby achieving iterative optimization of the prediction results and improving the accuracy and stability of multi-step prediction. The MSCT-RCM model proposed in this invention innovatively integrates a lightweight residual correction module. By deeply mining historical residual information from the training set and similar samples, a dynamic optimization mechanism is constructed, achieving accurate compensation and iterative optimization of prediction errors.

[0165] Step S23 specifically includes:

[0166] S231. Define the historical residual vector of the current sample as:

[0167] (11);

[0168] in, Represents the historical residual vector. H represents the residual corresponding to the prediction step size;

[0169] S232, Residual Correction Process

[0170] S2321, Residual Coding: Historical residual information is mapped to token representation using the following formula to adapt to Transformer coding:

[0171] (12);

[0172] in, The token representation after historical residual mapping. Represents the projection function. H represents the historical residual vector, and H represents the prediction step size. The hidden layer dimension of the Transformer;

[0173] S2322, Sequence Concatenation: Concatenate the sequences from the baseline prediction tokens... The first token is concatenated with the residual token to form the input sequence:

[0174] (13);

[0175] S2333, Lightweight Transformer Encoding: Based on a single-layer Transformer encoder, it uses the following formula to capture the interaction between baseline prediction results and historical residual information:

[0176] (14);

[0177] in, This represents the output of the lightweight Transformer encoder. For the Transformer encoder module, This represents the concatenated input sequence;

[0178] S2334, Residual Generation: The residual prediction result is obtained by mapping the first token through a multilayer perceptron using the following formula:

[0179] (15);

[0180] in, This is the vector of residual prediction results. This is the weight matrix of the residual generation layer. The bias vector of the residual generation layer. This is the first token vector output by the lightweight Transformer;

[0181] S2335. Prediction Output: The residual prediction is added to the baseline prediction result using the following formula to obtain the final prediction result:

[0182] (16);

[0183] in, This is the final prediction result vector. For the baseline prediction result vector, This is the vector of residual prediction results.

[0184] Step S23 enables adaptive correction of systematic errors under complex weather conditions, significantly improving the accuracy and stability of multi-step prediction. Compared with models without residual correction, the prediction accuracy is greatly improved.

[0185] As a further improvement to the above technical solution, step S3 specifically includes the following steps:

[0186] S31. Stage I involves training the baseline model, specifically the Transformer backbone network, to provide reliable baseline power predictions for residual correction. In Stage I, the optimization objective is designed to minimize the baseline predictions. The mean square error between the true value y and the actual value y is given by the following formula:

[0187] (17);

[0188] in, This represents the mean squared error loss for stage I. For the true value of sample i, represents the baseline prediction result vector, where N is the number of samples;

[0189] S32, Phase II is end-to-end joint training, and the optimization objective is as follows:

[0190] (18);

[0191] in, The mean squared error loss for stage II. This is the final prediction result vector. Let be the true value of sample i, and N be the number of samples. Stage II, through end-to-end joint training, enables the proposed MSCT-RCM model to adaptively utilize historical residual information to achieve iterative optimization of baseline output and residual correction, thereby improving the robustness of multi-step prediction.

[0192] Based on two years of measured data from a large photovoltaic power plant in northern China from 2018 to 2019, the ultra-short-term photovoltaic power prediction scheme based on the MSCT-RCM model proposed in this invention was comprehensively validated. Through comparative experiments with different prediction time domains, different input step sizes, and mainstream benchmark models, as well as ablation analysis, the prediction accuracy, robustness, and effectiveness of key modules in the MSCT-RCM model of the proposed scheme were systematically evaluated.

[0193] 1. Dataset and Sample Construction

[0194] The experimental data was sampled at 15-minute intervals, covering photovoltaic power output and multi-dimensional meteorological characteristics, including air temperature, cloud opacity, dew point temperature, direct radiation (DNI), diffuse radiation (DHI), global horizontal radiation (GHI), global tilted surface radiation (GTI), relative humidity, surface air pressure, wind speed, and wind direction. To ensure prediction effectiveness, only the effective time period from 08:00 to 18:00 daily was selected, avoiding the period of zero power at night. Considering the data sampling frequency, every four consecutive 15-minute sampling points were used as one sample for input. Based on a two-year time span, the total number of samples was approximately 8,030. To ensure the rationality of model training and testing, the training set and test set were divided in an 8:2 ratio, with the training set containing 6,424 samples and the test set containing 1,606 samples. The sample division maintained a continuous temporal order, and there was no overlap between the training set and the test set, ensuring the fairness of the experiment and the reliability of the results.

[0195] 2. Experimental Environment

[0196] All experiments were conducted on a computer with an Intel Xeon 3.0 GHz CPU, 64 GB of DDR4 memory, and an NVIDIA RTX 2080 Ti GPU (11 GB of VRAM). The operating system was Windows 11.0, the deep learning framework was PyTorch 2.1, and the CUDA version was 12.1.

[0197] 3. Evaluation Indicators

[0198] To quantify the power prediction performance of the proposed model from multiple perspectives, the following three types of evaluation metrics are used.

[0199] 1) Mean Absolute Error (MAE)

[0200] (19);

[0201] In this context, MAE measures the average deviation of predictions in kW. This evaluation index provides a direct assessment of the power prediction accuracy of the proposed model, where N represents the number of data points. For the true value, These are predicted values.

[0202] 2) Root Mean Square Error (RMSE)

[0203] (20);

[0204] Where N is the number of data points. For the true value, These are predicted values.

[0205] RMSE imposes a more severe penalty on larger errors, making it suitable for evaluating the power prediction performance of proposed models under extreme weather conditions.

[0206] 3) Coefficient of determination (R) 2 )

[0207] (twenty one);

[0208] Among them, R 2 This represents the proportion of variance explained by the model; a value closer to 1 indicates a better fit. N is the number of data points. For the true value, These are predicted values.

[0209] 4. Baseline Model and Hyperparameters

[0210] To comprehensively evaluate the ultra-short-term photovoltaic power prediction performance of the proposed scheme, four mainstream benchmark deep learning models—LSTM, BiLSTM, GRU, and standard Transformer—were selected as comparison models. The parameters of each benchmark model were optimized to ensure the fairness of the comparison. Their key architectures and hyperparameters are shown in Table 1.

[0211] Table 1. Baseline Model and Hyperparameters

[0212]

[0213] 5. Experimental Results and Analysis

[0214] 1) Influence of the number of features

[0215] Figure 3 The prediction performance of the proposed MSCT-RCM model with different numbers of features is shown. When the number of features increases from 5 to 10, the MAE and RMSE of the model in the three prediction time domains of 15 minutes, 30 minutes, and 1 hour all show a significant decreasing trend. 2 The continuously increasing values ​​reveal that introducing more highly relevant features can enhance the model's feature representation ability. However, when the number of features exceeds 10, the model performance slightly decreases, mainly due to the introduction of redundant or noisy information affecting the model's generalization ability. Therefore, given the optimal balance between information richness and model complexity achieved by the top 10 features, the proposed scheme sets the top 10 features as input variables.

[0216] 2) Sensitivity analysis for different input sequence lengths

[0217] Table 2 Performance Comparison for Different Input Sequence Lengths

[0218]

[0219] The impact of different input sequence lengths (seq_len=4, 8, 12, 16, 24) on the performance of the proposed MSCT-RCM model in three prediction ranges (15 minutes, 30 minutes, 1 hour) is summarized in Table 2 and [Table 2 is missing from the original text]. Figure 4 The results show that increasing the input length can improve accuracy to some extent, but the gains beyond this level are negligible. For 15-minute predictions, the MAE decreased from 2.75 kW (seq_len=4) to 1.91 kW (seq_len=12), and the R² increased from 0.9766 to 0.9944. Further extending the sequence to 16 or 24 steps only slightly improved performance, and similar trends were observed at the 30-minute and 1-hour horizons, revealing that an input length of 12–16 steps can achieve a balance between history-dependent capture and computational efficiency.

[0220] 3) Comparative analysis with the baseline model

[0221] Table 3 Performance Comparison Results for Different Prediction Durations

[0222]

[0223] Table 3 and Figure 5 The performance of the proposed MSCT-RCM model is compared with that of four mainstream deep learning baseline models (LSTM, BiLSTM, GRU, and the standard Transformer) at different prediction durations of 15 minutes, 30 minutes, 1 hour, 24 hours, and 48 hours. Overall, the MSCT-RCM model outperforms all baseline models on all evaluation metrics, especially demonstrating extremely high accuracy in ultra-short-term power prediction.

[0224] Specifically, with a 15-minute prediction duration, the MSCT-RCM model achieved a MAE of 1.91 kW and an RMSE of 2.73 kW, representing reductions of 59.4% and 56.4% respectively compared to the best-performing standard Transformer models (4.71 kW and 6.26 kW). The corresponding R² reached 0.9944, indicating a goodness-of-fit close to the ideal value. As the prediction duration increased, the errors of all models increased; however, the MSCT-RCM model consistently maintained a significant performance advantage. For a 1-hour time span, the MAE and RMSE were 3.22 kW and 5.20 kW respectively, more than 55% lower than the BiLSTM model (7.35 kW and 9.80 kW), while the R² value reached 0.9799, far exceeding other baseline models. For 24-hour and 48-hour time spans, the proposed model also achieved the national standard requirement of 85% photovoltaic power prediction accuracy. This significant performance improvement is attributed to the multi-stage feature modeling and error correction mechanism of the proposed MSCT-RCM model. The multi-scale convolutional branches effectively capture local short-term fluctuations and global trends, enhancing feature representation. The Transformer backbone network models global dependencies, ensuring stable predictions under complex time dynamics. The residual correction module can dynamically adjust predictions, reduce error accumulation in multi-step predictions, and maintain high accuracy and stability over a longer period of time.

[0225] 4) Prediction results for different time periods

[0226] Figure 6The typical prediction results of the MSCT-RCM model in the time periods of 15 minutes, 30 minutes, and 1 hour are shown. The predicted sequences are closely related to the actual power generation. For short-term prediction (15 minutes), the prediction results basically coincide with the actual situation, indicating that the local feature fitting accuracy is high. As the time period is extended to 30 minutes and 1 hour, some peaks and troughs show slight deviations, but the overall trend remains consistent. The error is significantly lower than that of the baseline model, which confirms that the residual correction module in the proposed model can effectively use historical errors to adjust the prediction, enhance the response capability, and reduce the lag or bias under rapidly changing weather conditions. It can achieve high-precision and robust ultra-short-term photovoltaic power prediction under different meteorological conditions.

[0227] 5) Ablation Research

[0228] Table 4 Results of Ablation Experiment

[0229]

[0230] Table 4 summarizes the ablation study results and evaluates the contributions of the multi-scale CNN branch, the Transformer backbone network, and the residual correction module to the ultra-short-term photovoltaic power prediction performance.

[0231] (1) Residual Correction Module (RCM): After removing RCM (using only CNN+Transformer), MAE increased from 2.57 kW to 3.24 kW, RMSE increased from 3.96 kW to 4.44 kW, and R² decreased to 0.9847, revealing that RCM in the proposed model effectively alleviates error accumulation in multi-step prediction by dynamically compensating for historical errors.

[0232] (2) Multi-scale CNN branch: After removing the CNN branch, the MAE increased to 4.12 kW, the RMSE increased to 5.30 kW, and the R² decreased to 0.9709, indicating that the CNN branch in the proposed model plays a key role in extracting local short-term and medium-term time features, thereby enhancing the feature representation capability of the Transformer.

[0233] (3) Transformer backbone: After removing the Transformer, the MAE is 3.90 kW, the RMSE is 5.13 kW, and the R² is 0.9736, which confirms its key role in capturing long-range dependencies and global feature interactions.

[0234] (4) Single module configuration: The accuracy of pure Transformer (w / o CNN + RCM) and pure CNN (w / o Transformer + RCM) is significantly reduced (R² is 0.9583 and 0.9610 respectively), which indicates that it is impossible to accurately perceive the multi-scale dynamics of ultra-short-term photovoltaic power sequences by using either module alone.

[0235] (5) Module synergy effect: The complete MSCT-RCM achieved the best performance on all indicators (MAE = 2.57kW, RMSE = 3.96kW, R² = 0.9876), indicating that the fusion of multi-scale convolution, global dependency modeling and residual correction in the proposed model has generated a strong synergy effect, ensuring high accuracy and robust ultra-short-term photovoltaic power prediction.

[0236] In summary, this invention proposes an ultra-short-term photovoltaic power prediction scheme based on the MSCT-RCM model. The MSCT-RCM model integrates multi-scale convolutional branches, a Transformer backbone network, and a residual correction module, achieving multi-scale feature extraction, globally dependent models, and dynamic error correction. Furthermore, the proposed scheme incorporates a two-stage MSCT-RCM model training strategy: first, a baseline model fusing CNN and Transformer is trained independently to achieve robust multi-scale feature extraction; then, based on pre-trained weights, joint optimization is performed in conjunction with the residual correction module. Experimental results based on measured data show that the proposed MSCT-RCM model outperforms LSTM, GRU, BiLSTM, and the standard Transformer model in prediction tasks of 15 minutes, 30 minutes, and 1 hour; particularly in the 15-minute prediction task, the MAE and RMSE are reduced by 59.4% and 56.4%, respectively. The MSCT-RCM model outperforms benchmark models (including LSTM, GRU, BiLSTM, and the standard Transformer) across multiple prediction durations. Sensitivity and ablation analyses confirm the effectiveness of the three core functional modules of the MSCT-RCM model and its strong generalization ability. The synergy between the MSCT-RCM model and its two-stage training strategy proposed in this invention ensures the achievement of the research goal of high-precision ultra-short-term photovoltaic power prediction, and is expected to play an important role in smart grid dispatching and energy management of photovoltaic power plants.

[0237] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A high-precision ultra-short-term photovoltaic power prediction method based on the MSCT-RCM model, characterized in that, The method includes the following steps: S1. Obtain raw photovoltaic power and meteorological data, perform preprocessing and feature enhancement on the data, and generate an input feature matrix; S2. Construct the MSCT-RCM model; The MSCT-RCM model is used to extract multi-scale convolutional features from the input feature matrix, fuse the multi-scale convolutional features with the original input features, generate baseline prediction results by modeling global dependencies of the sequence, and perform residual correction on the baseline prediction results based on historical residual information. S3. Determine the network structure parameters of the MSCT-RCM model and train the model using a two-stage training strategy; S4. Input the input feature matrix to be predicted into the trained MSCT-RCM model, and generate high-precision ultra-short-term photovoltaic power prediction results by fusing the baseline prediction results with the residual correction output.

2. The high-precision ultra-short-term photovoltaic power prediction method based on the MSCT-RCM model according to claim 1, characterized in that, Step S1 specifically includes: S11. Obtain raw photovoltaic power and meteorological data to ensure that the data covers the time series range and feature dimensions required for forecasting; S12. Clean the data and handle outliers; S13. Enhance the multi-scale features and select key features; S14. Normalization and Dataset Partitioning: The Min–Max normalization method is used to map all input features in the initial input feature matrix to the [0,1] interval to generate the final input feature matrix. The normalization formula is: （1）； Where x is the original value of the feature, x min For the eigenvalue minimum, x max x is the eigenvalue. norm For the normalized eigenvalues ​​after applying Min–Max normalization; S15. The sliding window method is used to divide the final input feature matrix into a fixed-length input sequence and a corresponding target output sequence, and outputs a set of paired samples of the input sequence and the target output sequence.

3. The high-precision ultra-short-term photovoltaic power prediction method based on the MSCT-RCM model according to claim 2, characterized in that, Step S12 specifically includes: S121. For records with a missing proportion lower than the set threshold, use linear interpolation to complete them; S122. Identify, remove or correct outliers with unreasonable physical meaning.

4. The high-precision ultra-short-term photovoltaic power prediction method based on the MSCT-RCM model according to claim 2, characterized in that, S131. Empirical mode decomposition is used to perform multi-scale decomposition on the time series of each numerical feature, decomposing the time series x(t) into several intrinsic mode functions and residual terms; The formula for the multi-scale decomposition is: （2）； in, represents the intrinsic mode function; i is the index of the IMF, ranging from 1 to n, where n is the total number of IMFs obtained from the decomposition; For residual terms; This represents the original value of the time series at time point t; S132. Based on the multi-scale decomposition results, extract the first two IMF components as enhancement features, and concatenate them with the original input features to form the input feature matrix. S133. Use gradient boosting regression to rank all enhanced features by importance, and select the core features that rank within a set range as input variables.

5. The high-precision ultra-short-term photovoltaic power prediction method based on the MSCT-RCM model according to claim 1, characterized in that, The MSCT-RCM model includes a multi-scale convolutional branch, a Transformer backbone network, and a residual correction module. The multi-scale convolution branch is used to extract multi-temporal convolutional features from the input feature matrix; The Transformer backbone network is used to fuse multi-scale convolutional features with original input features and generate baseline prediction results by modeling global dependencies of sequences. The residual correction module is used to correct the baseline prediction results based on historical residual information.

6. The high-precision ultra-short-term photovoltaic power prediction method based on the MSCT-RCM model according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21. Multi-scale convolutional branch feature extraction: Based on the convolutional neural network (CNN), multiple parallel one-dimensional convolutional branches are designed to extract multi-time convolutional features from the input sequence. S22. Construction of the backbone network: Multi-temporal convolutional features are fused with the original input features. The fused features are then adapted by linear projection and positional encoding, and then processed by a multi-head self-attention mechanism to model the global dependencies of the sequence, finally generating the baseline prediction results. S23. Residual Correction Module Construction: Combining historical residual information, the baseline prediction results are dynamically corrected, and the interaction between the baseline prediction results and historical residual information is captured by a lightweight Transformer encoder to iteratively optimize the prediction results.

7. The high-precision ultra-short-term photovoltaic power prediction method based on the MSCT-RCM model according to claim 6, characterized in that, Step S21 specifically includes: S211. Define the input sequence using the following formula: （3）； Where X represents the original input sequence matrix, Let L represent the feature vector at each time step in the sequence, and F represent the feature dimension. S212. Calculate the features of each convolutional branch using the following formula to obtain multi-branch features: （4）； Among them, H k This is the feature matrix extracted by the k-th convolutional kernel, where K is the kernel size and ReLU is the modified linear unit activation function. This is a one-dimensional convolution operation; S213. Use the following formula to concatenate the multi-branch features to form multi-scale convolutional features: （5）； in, For multi-scale convolutional features, These correspond to the outputs of convolutional branches with kernel sizes of 3, 5, and 7, respectively. Concat represents the concatenation operation, L represents the length of the input sequence, and C represents the number of convolutional output channels for each branch.

8. The high-precision ultra-short-term photovoltaic power prediction method based on the MSCT-RCM model according to claim 7, characterized in that, Step S22 specifically includes: S221. Use the following formula to convert the multi-scale convolutional features H... cnn The features are then fused with the original input features to generate fused features: （6）； in, Represents the fused feature matrix. Represents multi-scale convolutional features. Let L represent the original input sequence matrix, F represent the feature dimension, and C represent the number of convolution output channels for each branch. S222. Using the following formula, the fused features are mapped to the hidden space through linear projection to generate the projected features: （7）； in, This represents the projected feature matrix. Represents the fused feature matrix. and Here, are the projection weight matrix and the bias vector, respectively, and L represents the length of the input sequence. The hidden layer dimension of the Transformer; S223. Use the following formula to superimpose learnable positional codes onto the projected features to generate features with positional codes: （8）； in, This is the encoded feature matrix. PE represents the projected feature matrix, where PE is the learnable location encoding matrix. S224. Input the position-encoded features into an N-layer Transformer encoder using the following formula, and obtain the global temporal feature representation through a multi-head self-attention mechanism: （9）； Where Z represents the output matrix of the Transformer encoder. For the Transformer encoder module, The encoded feature matrix is ​​shown, where L represents the length of the input sequence. The hidden layer dimension of the Transformer; S225. Take the hidden state vector Z[−1] of the last time step output of the Transformer encoder, and use the following formula to generate the baseline prediction result of photovoltaic power through mapping of the fully connected layer: （10）； in, This is the baseline prediction result vector. This is the weight matrix of the fully connected layer. For the bias vector of the fully connected layer, Output the hidden state vector for the last step of the Transformer, where H represents the prediction step size.

9. The high-precision ultra-short-term photovoltaic power prediction method based on the MSCT-RCM model according to claim 8, characterized in that, Step S23 specifically includes: S231. Define the historical residual vector of the current sample as: （11）； in, Represents the historical residual vector. H represents the residual corresponding to the prediction step size; S232, Residual Correction Process S2321, Residual Coding: Historical residual information is mapped to token representation using the following formula to adapt to Transformer coding: （12）； in, The token representation after mapping historical residuals. Represents the projection function. H represents the historical residual vector, and H represents the prediction step size. The hidden layer dimension of the Transformer; S2322, Sequence Concatenation: Concatenate the sequences from the baseline prediction token... The first token is concatenated with the residual token to form the input sequence: （13）； in, This represents the concatenated input sequence. This is the baseline prediction result vector. The token representation is the result of mapping historical residuals, where H represents the prediction step size. The hidden layer dimension of the Transformer; S2333, Lightweight Transformer Encoding: Based on a single-layer Transformer encoder, it uses the following formula to capture the interaction between baseline prediction results and historical residual information: （14）； in, This represents the output of the lightweight Transformer encoder. For the Transformer encoder module, This represents the concatenated input sequence; S2334, Residual Generation: The residual prediction result is obtained by mapping the first token through a multilayer perceptron using the following formula: （15）； in, This is the vector of residual prediction results. This is the weight matrix of the residual generation layer. The bias vector of the residual generation layer. This is the first token vector output by the lightweight Transformer; S2335. Prediction Output: The residual prediction is added to the baseline prediction result using the following formula to obtain the final prediction result: （16）； in, This is the final prediction result vector. This is the baseline prediction result vector. This is the vector of residual prediction results.

10. The high-precision ultra-short-term photovoltaic power prediction method based on the MSCT-RCM model according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31. Phase I is for training the baseline model, specifically training the Transformer backbone network. In Phase I, the optimization objective is designed to minimize the baseline predictions. The mean square error between the true value y and the actual value y is given by the following formula: （17）； in, This represents the mean squared error loss for stage I. For the true value of sample i, represents the baseline prediction result vector, where N is the number of samples; S32. Stage II involves end-to-end joint training, performing iterative optimization of baseline prediction results and residual corrections. The optimization objective for this stage is shown in the following formula: （18）； in, The mean squared error loss for stage II. This is the final prediction result vector. Let be the true value of sample i, and N be the number of samples.

Images