A photovoltaic power generation short-term power prediction method

Abstract

The application discloses a kind of photovoltaic power generation short-term power prediction methods, belong to photovoltaic power generation field, comprising: obtaining the historical power generation data and meteorological forecast data of photovoltaic power station;The historical power generation data and meteorological forecast data are input into the confrontation reversible instance normalization module, and normalized characteristic data is obtained;Normalized characteristic data is input into multi-scale dynamic time series convolution network module, and multi-scale time series feature is obtained;Multi-scale time series feature is input into Reformer network module, and deep time series feature representation is obtained;Deep time series feature representation is input into the inverse normalization processing of confrontation reversible instance normalization module, and obtains photovoltaic power generation short-term power prediction result.The application effectively solves the problem of data distribution drift, significantly improves the prediction robustness under weather mutation scene.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power generation, and in particular relates to a method for short-term power prediction of photovoltaic power generation. Background Technology

[0002] With the advancement of global energy structure transformation, photovoltaic (PV) power generation, as a clean and renewable energy source, continues to increase its share in the power system. However, PV power generation is significantly affected by multiple meteorological factors such as irradiance, temperature, and humidity, exhibiting significant randomness, intermittency, and non-stationarity, posing a severe challenge to the safe and stable operation and optimized dispatch of the power grid. Short-term power forecasting, as a key technology for smoothing fluctuations and improving absorption capacity, directly impacts the peak-shaving capacity and economic operation level of the power grid. Traditional forecasting methods are mainly based on statistical models such as ARIMA or physical modeling methods, which are difficult to effectively capture complex nonlinear temporal dependencies. In recent years, deep learning technology has been widely applied in this field. Models represented by Transformer and its variants (such as Informer, Autoformer, iTransformer, and Reformer) have achieved breakthroughs in forecasting accuracy by capturing long-range dependencies through self-attention mechanisms; Temporal Convolutional Networks (TCNs) effectively extract local temporal features through causal convolution and dilated convolution; and Reversible Instance Normalization (RevIN) technology has been introduced to alleviate the problem of data distribution drift. The above methods provide a technical basis for photovoltaic power prediction, but they still face many bottlenecks in practical engineering applications.

[0003] The existing technical solutions have the following technical problems: First, traditional RevIN only relies on the current sample statistics for normalization processing and does not explicitly consider the data distribution alignment under different weather patterns, resulting in a significant decrease in the model's prediction robustness in sudden weather scenarios (such as a sudden change from sunny to rainy). Second, the TCN structure with a single expansion coefficient cannot simultaneously capture the multi-scale coupling features of short-term mutations, medium-term fluctuations, and long-term trends in photovoltaic power generation data, and its ability to perceive sudden power fluctuations is insufficient. Third, Transformer-type models have high computational complexity, while lightweight variants such as Reformer, while enhancing the efficiency of long sequence processing, have limited local feature extraction capabilities, making it difficult to balance the extraction of global dependencies and local details. Fourth, existing models have insufficient adaptive processing capabilities for non-stationary time series data, poor training stability under conditions of few samples or poor data quality, weak cross-seasonal generalization ability, and poor prediction accuracy. Summary of the Invention

[0004] The purpose of this invention is to propose a hybrid deep learning method based on Reformer network combined with Multi-Scale Dynamic Temporal Convolutional Network (MS-DTCN) and Adversarial Training Instance Normalization (AdvReVin) to address the problems of data distribution drift and multi-scale coupling of temporal features in photovoltaic power generation prediction. This method utilizes Reformer to capture long-term dependencies in time-series data, while the multi-branch parallel structure of MS-DTCN (with different expansion coefficients) captures local abrupt changes, medium-term variations, and long-term trend features of the time series, enhancing the ability to perceive sudden power fluctuations and weather changes. AdvReVin explicitly considers the alignment of data distribution under different weather patterns, making the model more robust to predictions during sudden weather changes (such as a sudden shift from sunny to cloudy).

[0005] To achieve the above objectives, the present invention provides a method for short-term photovoltaic power generation prediction, comprising: Acquire historical power generation data and weather forecast data from photovoltaic power plants; The historical power generation data and weather forecast data are input into the adversarial reversible instance normalization module to obtain normalized feature data; The normalized feature data is input into the multi-scale dynamic temporal convolutional network module to obtain multi-scale temporal features; The multi-scale temporal features are input into the Reformer network module to obtain a deep temporal feature representation. The deep time series feature representation is input into the adversarial reversible instance normalization module for inverse normalization processing to obtain the short-term power prediction result of photovoltaic power generation.

[0006] Preferably, the process of acquiring historical power generation data and weather forecast data of photovoltaic power plants includes: Acquire historical power generation data of photovoltaic power plants with a time resolution of 15 minutes; Obtain meteorological forecast data with a time resolution of 15 minutes corresponding to the historical power generation data. The meteorological forecast data includes solar irradiance, air temperature, and relative humidity.

[0007] Preferably, the process of obtaining normalized feature data includes: Input the historical power generation data and weather forecast data into the condition generator to obtain normalized parameters; The historical power generation data and weather forecast data are normalized according to the normalization parameters to obtain initial normalized features; The initial normalized features and weather condition labels are input into the discriminator to obtain the discrimination result; The normalization parameters are adjusted based on the discrimination result to obtain the normalized feature data.

[0008] Preferably, the discriminator is a lightweight conditional discriminator, whose input includes normalized features and weather condition labels, and whose output is the discrimination probability of matching features with weather conditions; By alternating the training of the condition generator and discriminator using the adversarial loss function, the normalized feature distribution under different weather conditions tends to be consistent.

[0009] Preferably, the multi-scale dynamic temporal convolutional network module includes: Multiple parallel convolutional branches, including short-time convolutional branches, medium-time convolutional branches, and long-time convolutional branches; The short-term, medium-term, and long-term convolutional branches all contain two one-dimensional convolutional layers and an efficient channel attention module, and adopt an improved residual connection structure. An enhanced residual connection structure is employed in each convolutional branch, wherein the enhanced residual connection structure is the sum of the convolutional layer output and the input after a 1×1 convolution; An efficient channel attention mechanism is integrated into each convolutional branch, which includes global average pooling, one-dimensional convolutional local channel interaction, and channel attention weight generation.

[0010] Preferably, the process of obtaining multi-scale temporal features includes: The normalized feature data is input into the short-time convolution branch, the medium-time convolution branch, and the long-time convolution branch respectively to obtain the short-time features, medium-time features, and long-time features accordingly; The short-term, medium-term, and long-term features are concatenated to obtain the concatenated features; The spliced ​​features are input into a multilayer perceptron to obtain the weight coefficients of each branch; The short-term, medium-term, and long-term features are weighted and fused according to the weighting coefficients to obtain the multi-scale temporal features.

[0011] Preferably, the short-term convolutional branch, the medium-term convolutional branch, and the long-term convolutional branch use different dilation coefficients, which correspond to local mutation feature extraction, medium-term change feature extraction, and long-term trend feature extraction, respectively. Among them, the short-term branch expansion coefficient is 1, the medium-term branch expansion coefficient is 4, and the long-term branch expansion coefficient is 16; The multilayer perceptron outputs weight coefficients normalized by the Softmax function, enabling data-driven adaptive feature fusion.

[0012] Preferably, the process of obtaining deep temporal feature representations includes: The multi-scale temporal features are divided into blocks to obtain multiple feature blocks; Calculate local sensitive hash attention for each feature block to obtain the attention output for each block; The attention outputs of each block are concatenated to obtain the deep temporal feature representation.

[0013] Preferably, the process of calculating local sensitive hash attention for each feature block and obtaining the attention output for each block includes: A locality-sensitive hash attention mechanism is used to perform attention calculation on multi-scale time-series feature data. The query vector and key vector are mapped and bucketed through a hash function, and attention is calculated only within the same bucket. A reversible network layer structure is used for forward and backward propagation, eliminating the need to store intermediate activation values; The input sequence is divided into multiple blocks, self-attention is calculated independently in each block, and the attention outputs of all blocks are concatenated to obtain the deep temporal feature representation.

[0014] Preferably, the short-term power prediction result of photovoltaic power generation is a power generation prediction sequence at 15-minute intervals within the next day; The input sequence length of the historical power generation data and weather forecast data to the adversarial reversible instance normalization module is 96 time steps, with each time step representing 15 minutes. The short-term power forecast results for photovoltaic power generation are the power forecast values ​​for the next 96 time steps.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: The photovoltaic power generation short-term power prediction method provided by this invention establishes an explicit mapping relationship between normalization parameters and weather conditions by introducing a reversible instance normalization module with adversarial training. This forces the feature distribution under different weather conditions to approach consistency, effectively solving the data distribution drift problem and significantly improving the prediction robustness under sudden weather changes. A multi-scale dynamic temporal convolutional network is used to extract local, medium-term, and long-term temporal features in parallel, and a dynamic weight fusion mechanism is combined to adaptively adjust the contribution of each branch, enhancing the model's ability to capture multi-scale temporal coupled features. Combined with the efficient long-sequence modeling capability of the Reformer network, the method achieves effective fusion of global dependencies and local features while maintaining computational efficiency. Experiments show that on real photovoltaic power plant datasets, this method improves prediction accuracy by 1%-3% and reduces the mean absolute error by approximately 15% compared to existing mainstream models, demonstrating excellent cross-seasonal generalization ability and engineering applicability. Attached Figure Description

[0016] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the TCR-Reformer network structure according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the hash bucketing, sorting, and block division steps in an embodiment of the present invention; Figure 3 This is a schematic diagram of the TCN network structure according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the ReVIN module structure according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the AdvReVin module structure according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the MS-DTCN module structure according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the time series spectra of photovoltaic-related variables in an embodiment of the present invention; Figure 8 This is a fitting diagram of the predicted power and actual power generation of each model for photovoltaic power station A in this embodiment of the invention; Figure 9 This is a fitting diagram of the predicted power and actual power generation of each model for photovoltaic power station B in this embodiment of the invention. Figure 10 This is a scatter plot of the predicted power and actual power generation of photovoltaic power plant A in a typical month according to an embodiment of the present invention. Figure 11 This is a scatter plot of the predicted power and actual power generation of photovoltaic power plant B in a typical month according to an embodiment of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0019] This embodiment provides a short-term photovoltaic (PV) power prediction method. It addresses the issues of data distribution drift and multi-scale coupling of temporal features in PV power prediction through an improved TCR-Reformer deep learning algorithm. This algorithm is based on Reformer and incorporates an AdvReVin normalization mechanism and a multi-scale dynamic temporal convolutional neural network (MS-DTCN) to enhance the model's ability to handle non-stationary time series and local receptive fields, while maintaining the Reformer's advantage in efficiently handling long sequences. The method specifically includes: Acquire historical power generation data and weather forecast data from photovoltaic power plants; Historical power generation data and weather forecast data are input into the adversarial reversible instance normalization module to obtain normalized feature data; Normalized feature data is input into a multi-scale dynamic temporal convolutional network module to obtain multi-scale temporal features; Multi-scale temporal features are input into the Reformer network module to obtain deep temporal feature representations; The deep temporal feature representation is input into the adversarial reversible instance normalization module for inverse normalization processing to obtain the short-term power prediction results of photovoltaic power generation.

[0020] Furthermore, the process of obtaining historical power generation data and weather forecast data for photovoltaic power plants includes: Acquire historical power generation data of photovoltaic power plants with a time resolution of 15 minutes; Acquire meteorological forecast data with a time resolution of 15 minutes corresponding to historical power generation data. The meteorological forecast data includes solar irradiance, air temperature, and relative humidity.

[0021] Furthermore, the process of obtaining normalized feature data includes: Input historical power generation data and weather forecast data into the condition generator to obtain normalized parameters; The historical power generation data and weather forecast data are normalized according to the normalization parameters to obtain the initial normalized characteristics; Input the initial normalized features and weather condition labels into the discriminator to obtain the discrimination results; Adjust the normalization parameters based on the discrimination results to obtain normalized feature data.

[0022] Furthermore, the discriminator is a lightweight conditional discriminator, whose input includes normalized features and weather condition labels, and whose output is the discriminant probability of matching features with weather conditions; By alternating the training of the condition generator and discriminator using the adversarial loss function, the normalized feature distribution under different weather conditions tends to be consistent.

[0023] Furthermore, the multi-scale dynamic temporal convolutional network module includes: Multiple parallel convolution branches, including short-time convolution branches, medium-time convolution branches, and long-time convolution branches; The short-term, medium-term, and long-term convolution branches all contain two one-dimensional convolutional layers and an efficient channel attention module, and adopt an improved residual connection structure. An enhanced residual connection structure is used in each convolutional branch, which is the sum of the convolutional layer output and the input after a 1×1 convolution. An efficient channel attention mechanism is integrated into each convolutional branch. This efficient channel attention mechanism includes global average pooling, one-dimensional convolutional local channel interaction, and channel attention weight generation.

[0024] Furthermore, the process of obtaining multi-scale temporal features includes: Normalized feature data are input into the short-time convolution branch, the medium-time convolution branch, and the long-time convolution branch respectively to obtain short-time features, medium-time features, and long-time features respectively; The short-term, medium-term, and long-term features are concatenated to obtain the concatenated features; The concatenated features are input into a multilayer perceptron to obtain the weight coefficients of each branch; The short-term, medium-term, and long-term features are weighted and fused based on the weighting coefficients to obtain multi-scale time-series features.

[0025] Furthermore, the short-term convolutional branch, the medium-term convolutional branch, and the long-term convolutional branch use different dilation coefficients, which correspond to local mutation feature extraction, medium-term change feature extraction, and long-term trend feature extraction, respectively. Among them, the short-term branch expansion coefficient is 1, the medium-term branch expansion coefficient is 4, and the long-term branch expansion coefficient is 16; The multilayer perceptron outputs weight coefficients normalized by the Softmax function, enabling data-driven adaptive feature fusion.

[0026] Furthermore, the process of obtaining deep temporal feature representations includes: Multi-scale temporal features are divided into blocks to obtain multiple feature blocks; Calculate local sensitive hash attention for each feature block to obtain the attention output for each block; The attention outputs of each block are concatenated to obtain a deep temporal feature representation.

[0027] Furthermore, the process of calculating local sensitive hash attention for each feature block and obtaining the attention output for each block includes: A locality-sensitive hash attention mechanism is used to perform attention calculation on multi-scale time-series feature data. The query vector and key vector are mapped and bucketed through a hash function, and attention is calculated only within the same bucket. A reversible network layer structure is used for forward and backward propagation, eliminating the need to store intermediate activation values; The input sequence is divided into multiple blocks, and self-attention is calculated independently within each block. The attention outputs of all blocks are then concatenated to obtain a deep temporal feature representation.

[0028] Furthermore, the short-term power forecast results for photovoltaic power generation are a sequence of power generation forecasts at 15-minute intervals within the next day; The input sequence length of the adversarial reversible instance normalization module is 96 time steps, with each time step representing 15 minutes; The short-term power forecast for photovoltaic power generation is the power forecast value for the next 96 time steps.

[0029] Furthermore, in the above embodiments, this embodiment proposes a short-term photovoltaic power prediction model based on TCR-Reformer. The TCR-Reformer network structure diagram is shown below. Figure 1 As shown.

[0030] Specifically, this embodiment uses the Reformer network as the baseline model and introduces two innovative modules, AdvRevin and MS-DTCN, for enhancement. The Reformer, as an efficient variant of the Transformer, has the advantage of handling long sequences, and its structure includes: (1) Reversible Layers; Reversible layers are a type of neural network architecture that allows the input to be reconstructed from the output without storing intermediate activation values, significantly reducing memory usage and making them particularly suitable for training long sequences and deep networks. Let the input be... The output is The functional relationship of the reversible layer is: (1) (2) Through this relationship, it is possible to propagate backwards using only... Calculate x without storing intermediate variables.

[0031] (2) Locality Sensitive Hash (LSH) attention mechanism; LSH uses a set of hash functions to map high-dimensional data points to a low-dimensional space, ensuring that similar data points are hashed into the same or similar buckets. The standard self-attention mechanism has a computational complexity of O(n²), where n is the sequence length. Reformer uses Locality Sensitive Hashing (LSH) to approximate the attention distribution, reducing the computational complexity to O(nlogn). Assuming a query vector Q and a key vector K, they are bucketed using a hash function h: (3) (4) In LSH attention, a hash function is used to bucket queries and keys, and attention is only calculated within the same bucket. In this way, LSH can effectively group similar data points together, thus significantly reducing computation in large-scale data processing. A simplified version of the LSH attention mechanism is as follows: Figure 2 As shown.

[0032] (3) Segmented query and weight sharing; When processing long sequences, the Reformer model uses a segmented query approach, which divides the long sequence into smaller blocks. Then, self-attention is calculated independently within each block. Assuming the input sequence has a length of n, it is divided into m blocks, each with a length of l = n / m. The calculation formula is as follows: (5) Where d is the feature dimension and l is the size of each block. For each block Calculate self-attention separately.

[0033] (6) , , They are blocks X i The query, key, and value matrix. The attention outputs of all blocks are then reassembled into a complete sequence. The formula is as follows: (7) In Reformer, some parameters are shared across multiple layers. Specifically, segmented queries reduce the computational complexity of each layer, while weight sharing reduces the number of parameters in the overall model. These two aspects complement each other, making Reformer both efficient and resource-saving when processing long sequences.

[0034] Furthermore, the TCN (Temporal Convolutional Network) module, as an efficient temporal convolutional network structure, employs causal convolution and dilated convolution to effectively capture long-term temporal dependencies. The TCN module stacks multiple residual blocks, each containing a one-dimensional convolutional layer, layer normalization, and a non-linear activation function, resulting in powerful temporal feature extraction capabilities. By employing residual connections and weight normalization techniques, TCN effectively mitigates the vanishing gradient problem. TCN uses causal convolution to ensure that the model's predictions depend only on the current time step and previous inputs, preventing data leakage issues. The formula is as follows: (8) Where k is the kernel size. It is the input at time t. This is the output. Furthermore, to capture long-range dependencies without increasing the number of parameters, TCN introduces dilated convolutions.

[0035] (9) Where d is the inflation factor, which increases exponentially with the number of network layers (usually d=2). l (where l is the number of layers). This mechanism enables TCN to effectively model the relationship between solar irradiance and power generation at different time scales. TCN employs a residual block design, with each residual block containing two layers of structures with dilated convolutions, as well as skip connections.

[0036] (10) in This represents the convolution operation within the residual block. This design helps alleviate the difficulties in training deep networks and improves the model's tracking performance of power generation capacity under variable weather conditions. The TCN module, for example... Figure 3 As shown.

[0037] Furthermore, RevIN (Reversible Instance Normalization), as a reversible instance normalization method, is used to address the problem of data distribution variation in time series forecasting. Its core idea is to consider the time dimension of the sequence during the normalization process to handle data distribution drift and enhance the model's ability to handle non-stationary time series. RevIN comprises three processes: normalization, affine transformation, and denormalization, as shown below: (1) Normalization stage: For a given input data X∈RB×T×N (X is a B×T×N dimensional real tensor), where B is the batch size, T is the number of time steps, and N is the feature dimension. ReVIN calculates the mean and standard deviation of each instance along the time dimension.

[0038] (11) (12) in, and These are the mean and standard deviation of the i-th instance, respectively. It is a small constant used to ensure numerical stability. The normalization process standardizes each time series instance to a distribution with a mean of 0 and a standard deviation of 1.

[0039] (13) (2) Affine transformation: After instance normalization, ReVIN introduces learnable affine parameters, allowing the model to adjust the distribution of the normalized data.

[0040] (14) in, and These are learnable scaling and offset parameters, corresponding to each feature dimension. The affine transformation parameters are optimized during training via backpropagation, enabling the model to learn the optimal scaling and offset coefficients for each feature.

[0041] (3) Denormalization stage: The key innovation of ReVIN lies in its reversibility. In the model prediction stage, the normalized data can be restored to the original distribution using the previously saved mean and standard deviation, as shown in Equations 15 and 16.

[0042] (15) (16) This normalization operation ensures that the predicted results and the original data lie in the same distribution space, thus solving the distribution shift problem. Through learnable affine parameters, ReVIN can adapt to the distribution requirements of different features, improving the model's representational power. The ReVIN module is as follows: Figure 4 As shown.

[0043] To further optimize the technical solution, this embodiment enhances it through two specially designed modules, forming a more complete prediction framework.

[0044] To address the issues of distribution shift and non-stationarity in time-series data, this embodiment proposes an Adversarial Learning-based Invertible Instance Normalization (AdvReVin) module. This module improves upon traditional ReVIN by introducing an adversarial training mechanism and non-stationary information processing components, enabling the model to better adapt to changes in the distribution of time-series data. AdversarialReVin normalizes the input sequence and captures the mapping relationship between the source and target distributions through an adversarial learning framework. This adversarial training strategy allows the model to learn more robust feature representations, maintaining prediction consistency across different time periods and weather conditions, effectively mitigating the impact of distribution shifts caused by weather variations. This reduces prediction errors caused by seasonal changes and weather fluctuations.

[0045] To capture the multi-scale temporal dependencies in photovoltaic power generation data, this embodiment designs a Multi-Scale Dynamic Temporal Convolutional Network (MS-DTCN) module. Based on the traditional Temporal Convolutional Network (TCN) architecture, this module introduces multiple convolutional branches with different receptive field sizes (d=1, d=4, and d=16, respectively) to simultaneously model short-term, medium-term, and long-term temporal patterns. Each branch contains two convolutional layers and an ECA (Efficient Channel Attention) module, integrating multi-scale features through a dynamic weight fusion mechanism, thus enhancing the model's ability to perceive changes at different time scales.

[0046] The overall framework of this embodiment adopts an end-to-end design concept. The data flow includes the following steps: the input time series is preprocessed by the AdvReVin instance normalization module, and then the normalized data is input into the MS-DTCN module for multi-scale feature extraction; then the extracted features are passed to the TCR-Reformer module for deep time series relationship modeling; finally, the model output is processed by AdvReVin inverse normalization to obtain the final day-ahead prediction result of short-term power of photovoltaic power plants.

[0047] Furthermore, traditional ReVin eliminates data distribution differences through instance normalization, but its normalization parameters (mean μ, variance σ) only depend on the current sample and do not explicitly consider distribution alignment under different weather patterns. This leads to a decrease in the model's predictive robustness when weather changes abruptly (e.g., from sunny to cloudy). To address this issue, this embodiment proposes a normalization method called AdvReVin based on the ReVin module. It considers introducing dynamic thresholds and adversarial normalization mechanisms, predicts the normalization scaling and offset parameters through a lightweight LSTM network, and forces the normalized feature distributions to align under different weather conditions through a discriminator network, thereby improving the model's adaptability to weather changes. Specifically: Design a lightweight conditional discriminator D, whose input is the normalized feature z.t The weather condition label c (e.g., sunny / rainy / cloudy) is used as the output, which is the discrimination probability D(z). t |c), representing feature z t Does it match weather condition c? The adversarial loss function is as follows: (17) in, The normalized characteristic distribution under real weather conditions. The normalized feature distribution generated for the model. This is achieved by minimizing... This forces z under different weather conditions t The distributions tend to be uniform. Next, in ReVin's normalization layer, the generation process of the mean μ and variance σ is combined with adversarial training. The discriminator gradient is backpropagated to the normalization parameters μ and σ.

[0048] (18) (19) in, Main task prediction loss (e.g., MAE), balance coefficient It is a key hyperparameter used to balance prediction accuracy and the model's ability to generalize under different weather conditions.

[0049] This embodiment determines its optimal value through grid search. In the experiment, this embodiment tested... ∈{0.01,0.1,1,10}, etc., and finally found that when When the value is 0.1, the model prediction achieves the lowest root mean square error (RMSE). To enhance explicit control over weather conditions, conditional parameters are introduced in ReVin. Normalized parameters are dynamically generated based on the weather label c.

[0050] (20) The normalization formula is adjusted to: (twenty one) By introducing the aforementioned adversarial training process, further constraints can be imposed. Consistency across weather distributions. The AdvReVin module, for example... Figure 5 As shown.

[0051] Furthermore, the MS-DTCN (Multi-Scale Dynamic Temporal Convolution Network) module involved in this embodiment is an improved multi-scale dynamic temporal convolutional network based on the TCN module, employing a multi-branch parallel structure to replace the fixed dilation coefficient in TCN. Specifically, small, medium, and large dilation coefficients (assigned values ​​of 1, 4, and 16 respectively in this embodiment's experiments) replace the fixed dilation coefficient in TCN. These three dilation coefficients correspond to the short-term, medium-term, and long-term branches, respectively. This is used to capture local abrupt changes, intermediate-term variations, and long-term trends in time-series data. The convolution kernel size used in different branches of MS-DTCN is 3×1, and the network has two layers; the difference lies in the dilation coefficient.

[0052] Simultaneously, an attention mechanism is introduced to adaptively adjust the weight contributions of each branch, achieving a dynamic weight fusion mechanism. This includes designing a feature importance evaluation module to dynamically allocate the weights of each branch based on the characteristics of the input data, and implementing data-driven adaptive feature fusion to enhance the model's adaptability to different weather patterns. The formula is as follows: (twenty two) In the formula, , , The output features of the three branches are respectively. It is a multilayer perceptron used to learn feature importance, and finally outputs the feature F_final= [0]× + [1]× + [2]× .

[0053] Next, cross-layer residual connections are introduced to improve the residual structure of traditional TCN and alleviate the gradient vanishing problem at deeper levels; a feature enhancement module is added to improve the feature representation capability in the residual path. The specific formula is as follows.

[0054] Original residual connection: (twenty three) Enhanced residual connectivity: (twenty four) Finally, the ECA (Efficient Channel Attention) concept is integrated into the TCN structure, adding channel attention modules to each timescale branch and adaptively adjusting the importance of different feature channels to enhance the expression of key features. Specifically, global average pooling is performed on temporal features to obtain channel descriptors, 1D convolution is used for local channel interactions to capture cross-channel correlations, and then channel attention weights are generated to highlight important feature channels. Finally, the 1D convolution kernel size is adaptively selected based on the number of input channels, where the kernel size k = f(C), and C is the number of channels, and f is the mapping function.

[0055] Photovoltaic power generation is influenced by factors at multiple time scales. MS-DTCN achieves multi-scale feature extraction through parallel branching, enhancing the model's ability to perceive changes at different time granularities. It also introduces a dynamic weight fusion mechanism, which is more flexible than the fixed structure of traditional TCN. Furthermore, the channel attention mechanism and optimized residual connections used in MS-DTCN significantly improve the model's feature representation capabilities. MS-DTCN modules include... Figure 6 As shown.

[0056] As an additional implementation method, to verify the practicality of the proposed method in photovoltaic power generation application scenarios, this embodiment uses data from two real photovoltaic power stations in the Tibet Autonomous Region, divided by season, to conduct experimental verification, demonstrating the universal applicability and stability of the proposed method.

[0057] It should be noted that all data used in this embodiment were provided by the Tibet Autonomous Region Meteorological Bureau. The real photovoltaic power generation data used in the experiment were collected from two independent sensor systems, and the two photovoltaic power plants involved have different fluctuation patterns in parameters such as installed capacity, radiation, temperature, and power generation. Their data distributions are not the same under similar seasonal conditions. The time resolution of PV power and NWP data in the data source is 15 minutes, which meets the relevant standards of the power industry. NWP data improves the prediction accuracy of the model by providing key meteorological parameters based on physical mechanisms such as solar irradiance, temperature, and humidity, and can predict cloud dynamics and distribution, thereby more accurately estimating the solar energy reaching the photovoltaic panels and supporting the prediction of radiation changes at different time scales (hourly to daily). In addition, this embodiment also performs dynamic error correction processing, which can calibrate the NWP output through historical data to reduce systematic bias and adapt to different weather types and seasons.

[0058] In short-term power forecasting research for photovoltaic (PV) power generation, a refined forecasting and evaluation mechanism is crucial for measuring the practical application value of forecasting models. A standardized evaluation system is adopted for the short-term power forecasting results of PV power plants, with daily statistics and monthly assessments. Accuracy calculations differentiate between different power conditions: when the actual power is greater than or equal to 20% of the installed capacity, the actual power is used as the standardized benchmark; while when the actual power is less than 20% of the installed capacity, 20% of the installed capacity is used as the evaluation benchmark. This effectively solves the problem of excessive evaluation deviation under low-power conditions. According to grid industry standards, the accuracy of short-term power forecasting for PV power plants should not be lower than 65%. This refined evaluation system directly reflects the importance of short-term forecasting in actual power dispatching and is also the core evaluation basis for the day-ahead short-term power forecasting based on the TCR-Reformer network in this embodiment.

[0059] This embodiment's verification experiment is conducted in two stages. First, considering the requirements of practical engineering applications, the seasons with higher radiation levels (July, August, and September 2023) are selected to examine and optimize the model's fitting ability, prediction accuracy, stability, and generalization ability. For this part, data from October 1, 2022 to June 30, 2023 (a total of nine months) are used as the training set for both power plants. In addition, this embodiment also conducts cross-seasonal verification experiments to assess the model's adaptability to different environmental and weather conditions. March, June, September, and December 2023 are selected as representative months for spring, summer, autumn, and winter, respectively, to verify the model's practicality. In this part, the data from the remaining months, excluding the test set, is shuffled and used as the training set for this embodiment. All measurements are recorded at 15-minute intervals to enable high-resolution modeling of short-term fluctuations and long-term trends.

[0060] Figure 7 The data presents time series plots of selected variables for two different photovoltaic power plants, A and B, from October 1, 2022 to June 30, 2023. It can be seen that the radiation from photovoltaic power plant A is relatively more stable, while the radiation from photovoltaic power plant B fluctuates more dramatically. Furthermore, photovoltaic power plant A experiences a significant low-temperature period in the middle of the observation period, while photovoltaic power plant B exhibits more uniform temperature fluctuations without a particularly pronounced low-temperature period. The humidity parameters are largely similar for both. Finally, the actual power generation of photovoltaic power plant A is relatively more stable and fluctuates more evenly, while photovoltaic power plant B experiences significant periods of low power output. These data further confirm that even under similar seasonal conditions, their data distributions are not identical. This allows this embodiment to verify the applicability and robustness of the prediction model under different environmental and scale conditions using these two power plants.

[0061] like Figure 7The time series plot of photovoltaic-related variables shown in the figure displays time series data for four key variables, from top to bottom: total solar irradiance (W / m²), air temperature (°C), relative humidity (%), and photovoltaic output power (MW). The horizontal axis represents the time step index, where each step represents a 15-minute time interval.

[0062] Furthermore, during the training and testing phases, the learning rate was set to 0.0001, and the model was trained for a maximum of 20 epochs. To prevent overfitting and improve training efficiency, this embodiment employed an early stopping mechanism with a patience period of 4 epochs. An early stopping criterion was set to monitor the validation loss, requiring the loss to decrease by at least 0.001 within the patience window before resetting the counter; otherwise, training would be terminated. The experimental environment and specific parameter settings in this experiment are shown in Table 1.

[0063] Table 1

[0064] Furthermore, to evaluate the accuracy of the prediction method proposed in this embodiment, the accuracy of the model prediction is verified by using the calculation method for the day-ahead prediction accuracy of short-term power in the power grid. The calculation formula is as follows: (25) Among them: power rationing time Let i be the available power at time i, and let i be the power at time i without power constraints. Let i be the actual power at time i. The predicted power for time i is given by [previous value]. This represents the total installed capacity of the photovoltaic power plant for the current month. The day-ahead forecast accuracy in the short-term power forecast results for photovoltaic power plants should be no less than 65%.

[0065] Furthermore, in the subsequent experimental section of this embodiment, the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) will be used to measure the goodness of fit between the predicted values ​​and the actual observed values. The specific formulas are as follows: (26) in, For the first One actual value, For the first There are 10 predicted values, where n is the total number of samples. The absolute values ​​of all error terms are summed to avoid the cancellation of positive and negative errors.

[0066] (27) in, For the first One actual value, For the first There are n predicted values, where n is the total number of samples. RMSE first calculates the square of the difference between the predicted and actual values, then averages the results and takes the square root.

[0067] (28) in, For the i-th actual value, For the i-th predicted value, This is the average of all actual values. The denominator is the total sum of squares (SST), representing the total variance of the original data; the numerator is the residual sum of squares (SSR), representing the variance that the model failed to explain.

[0068] Furthermore, to verify the reliability of the proposed TCR-Reformer method, this embodiment also conducted comparative experiments with other classic time series prediction models (including Transformer, Informer, Autoformer, iTransformer, and Reformer). On test sets from July to September 2023 at two different photovoltaic power plants, the prediction performance of the models is shown in Tables 2 and 3. The comparison between the predicted power and the actual generated power of each model is as follows: Figure 8 , Figure 9 As shown, TCR-Reformer achieves the best performance in terms of accuracy and mean absolute percentage error on the test set, demonstrating a certain advantage over all benchmark models.

[0069] Figure 8 A graph showing the fitting of predicted power and actual power generation from various photovoltaic power generation models. To more clearly observe the degree of fitting between the model predicted power and actual power generation, three sub-graphs were drawn, corresponding to certain dates in July, August, and September, respectively. Note: In the comparison of the power prediction results of the six models, the black, light blue, blue, green, orange, and purple lines correspond to Transformer, Informer, Autoformer, iTransformer, Reformer, and TCR-Reformer, respectively, while the red line represents the true value.

[0070] This embodiment evaluates the performance of different variants of the Transformer model on photovoltaic power prediction tasks through a series of experiments. Figure 8 and Figure 9The experimental results reveal several key findings: the TCR-Reformer model's predictions are closer to the actual power generation on most test dates, especially during periods of significant data fluctuation, where its performance advantage is more pronounced; at power peaks and in rapidly changing regions, the TCR-Reformer model's predictions are closest to the actual values. Further analysis shows that when illumination conditions are stable, all models exhibit relatively high prediction accuracy. However, on days with variable weather conditions, the performance differences between models widen significantly. This indicates that the non-stationarity caused by weather changes is a key factor affecting prediction accuracy. Notably, the TCR-Reformer model proposed in this embodiment effectively mitigates this non-stationarity by introducing the AdvRevin module, demonstrating higher stability on days with significant fluctuations in actual power generation. The performance comparison results of similar models (PV power plant A, July-September) are shown in Table 2, and the performance comparison results of similar models (PV power plant B, July-September) are shown in Table 3.

[0071] Table 2

[0072] Table 3

[0073] Figure 9 Figure B shows the fitting plots between the model-predicted power and the actual power generation for photovoltaic power station B. To more clearly observe the degree of fitting between the model-predicted power and the actual power generation, three subplots were drawn, corresponding to certain dates in July, August, and September, respectively.

[0074] Furthermore, to further explore the predictive performance, generalization ability, and robustness of the model proposed in this embodiment under varying meteorological conditions such as light and temperature in different seasons, this embodiment designs the test set as independent samples with seasonal representativeness to verify the model's generalization ability in different seasons. Specifically, this embodiment selects data from March, June, September, and December 2023 as the test set, representing the four seasons of spring, summer, autumn, and winter, respectively. These test data are independent of the training process and are used only for final evaluation. The training set consists of all remaining data that does not overlap with the test set, i.e., data from October 2022 to January 2024, excluding the data from the aforementioned four months. This partitioning method ensures that the training set has a sufficient time span and data volume, providing ample support for the model to learn the long-term variation pattern of photovoltaic power generation, while avoiding data leakage and ensuring the objectivity of the test results. Data sets from photovoltaic power plants A and B were selected as data sources, and the prediction performance of the TCR-Reformer model and five other baseline models (Autoformer, Informer, iTransformer, Reformer, and Transformer) was evaluated on the test set for each of the four seasons. Experimental results are shown in Tables 4 and 5. Furthermore, to more intuitively demonstrate the prediction performance of each model, this embodiment plots scatter plots of the predicted power and actual power generation of each model for photovoltaic power plants A and B in typical months, as shown in Tables 4 and 5. Figure 10 as well as Figure 11 The performance comparison results of the cross-seasonal model for photovoltaic power station A are shown in Table 4, and the performance comparison results of the cross-seasonal model for photovoltaic power station B are shown in Table 5.

[0075] Table 4

[0076] Table 5

[0077] By comparing scatter plots from four different seasons, it can be observed that the TCR-Reformer model predicts more concentrated and less dispersed power generation values ​​compared to other models. This intuitively demonstrates that the TCR-Reformer model has higher accuracy and smaller errors in predicting photovoltaic power generation. The prediction accuracy of the TCR-Reformer model exhibits strong consistency. On most datasets (e.g., March, June, and September), it consistently demonstrates better prediction performance than other models. Even in December, the prediction accuracy is very close to that of the optimal model, proving that the TCR-Reformer model has good robustness and seasonal adaptability, maintaining high prediction accuracy under different meteorological conditions and seasonal variations.

[0078] In summary, the experimental results demonstrate that the TCR-Reformer model proposed in this embodiment achieves satisfactory performance in photovoltaic power prediction tasks, outperforming existing classic Transformer variant models in terms of accuracy and stability. This is mainly attributed to the effective handling of data distribution shifts by the AdvRevin module and the accurate capture of multi-scale time dependencies by the MS-DTCN module. Furthermore, this embodiment also verifies that when the test set consists of months from different seasons, even with significant differences in data distribution, the prediction accuracy of the TCR-Reformer model is more stable than other transformer-type models. This confirms the good applicability and robustness of the TCR-Reformer model proposed in this embodiment. These results provide a new technical approach for photovoltaic power generation prediction and offer more reliable predictive support for renewable energy grid connection management.

[0079] In summary, to address the issues of data distribution drift and multi-scale coupling of time-series features in photovoltaic power forecasting, this embodiment proposes a TCR-Reformer model based on the Reformer network framework, fusing adversarial training-based reversible instance normalization (AdvReVin) with a multi-scale dynamic temporal convolutional network (MS-DTCN). Validation of the proposed method using historical photovoltaic data and NWP numerical weather prediction data from October 1, 2022 to January 1, 2024 from a power company in the Tibet Autonomous Region shows that, compared to other classic Former models, this method improves prediction accuracy by 1%-3% using grid evaluation standards and reduces the mean absolute error (MAE) by approximately 15%. Furthermore, the modular design of AdvReVin and MS-DTCN proposed in this embodiment can provide a cross-domain methodological reference for energy time-series data analysis tasks such as wind power forecasting and regional load forecasting, contributing to the construction of a highly resilient new power system. The research results of this embodiment not only supplement existing research findings but also provide strong support for power grid companies to optimize power dispatch and reduce operating costs.

[0080] The above experiments demonstrate that the model proposed in this embodiment has wide applicability and strong robustness, but it also has certain limitations. Future research can improve the model's portability to adapt to regional characteristics in different areas, and consider introducing more dimensional data and more complex meteorological conditions to conduct more comprehensive experiments to further verify the model's robustness and accuracy.

[0081] This embodiment evaluates the performance of the proposed method using real photovoltaic power generation data from multiple photovoltaic power plants in Guizhou Province from October 2022 to January 2024, provided by the Tibet Autonomous Region Power Company. The results show that the proposed method achieves prediction accuracies of 75.67% and 73.60% on this dataset using the Chinese power grid industry evaluation standards. Compared with classic models such as Transformer, Informer, Autoformer, iTransformer, and Reformer, this method improves prediction accuracy by 1%-3% and reduces the mean absolute error (MAE) by approximately 15%. Experimental results demonstrate that this method can provide good prediction accuracy in photovoltaic power systems.

[0082] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for short-term photovoltaic power generation prediction, characterized in that, include: Acquire historical power generation data and weather forecast data from photovoltaic power plants; The historical power generation data and weather forecast data are input into the adversarial reversible instance normalization module to obtain normalized feature data; The normalized feature data is input into the multi-scale dynamic temporal convolutional network module to obtain multi-scale temporal features; The multi-scale temporal features are input into the Reformer network module to obtain a deep temporal feature representation. The deep time series feature representation is input into the adversarial reversible instance normalization module for inverse normalization processing to obtain the short-term power prediction result of photovoltaic power generation.

2. The method according to claim 1, characterized in that, The process of obtaining historical power generation data and weather forecast data for photovoltaic power plants includes: Acquire historical power generation data of photovoltaic power plants with a time resolution of 15 minutes; Obtain meteorological forecast data with a time resolution of 15 minutes corresponding to the historical power generation data. The meteorological forecast data includes solar irradiance, air temperature, and relative humidity.

3. The method according to claim 1, characterized in that, The process of obtaining normalized feature data includes: Input the historical power generation data and weather forecast data into the condition generator to obtain normalized parameters; The historical power generation data and weather forecast data are normalized according to the normalization parameters to obtain initial normalized features; The initial normalized features and weather condition labels are input into the discriminator to obtain the discrimination result; The normalization parameters are adjusted based on the discrimination result to obtain the normalized feature data.

4. The method according to claim 3, characterized in that, The discriminator is a lightweight conditional discriminator. Its input includes normalized features and weather condition labels, and its output is the discriminant probability of matching features with weather conditions. By alternating the training of the condition generator and discriminator using the adversarial loss function, the normalized feature distribution under different weather conditions tends to be consistent.

5. The method according to claim 1, characterized in that, The multi-scale dynamic temporal convolutional network module includes: Multiple parallel convolutional branches, including short-time convolutional branches, medium-time convolutional branches, and long-time convolutional branches; The short-term, medium-term, and long-term convolutional branches all contain two one-dimensional convolutional layers and an efficient channel attention module, and adopt an improved residual connection structure. An enhanced residual connection structure is employed in each convolutional branch, wherein the enhanced residual connection structure is the sum of the convolutional layer output and the input after a 1×1 convolution; An efficient channel attention mechanism is integrated into each convolutional branch, which includes global average pooling, one-dimensional convolutional local channel interaction, and channel attention weight generation.

6. The method according to claim 1, characterized in that, The process of obtaining multi-scale temporal features includes: The normalized feature data is input into the short-time convolution branch, the medium-time convolution branch, and the long-time convolution branch respectively to obtain the short-time features, medium-time features, and long-time features accordingly; The short-term, medium-term, and long-term features are concatenated to obtain the concatenated features; The spliced ​​features are input into a multilayer perceptron to obtain the weight coefficients of each branch; The short-term, medium-term, and long-term features are weighted and fused according to the weighting coefficients to obtain the multi-scale temporal features.

7. The method according to claim 6, characterized in that, The short-term, medium-term, and long-term convolutional branches use different dilation coefficients, which correspond to local mutation feature extraction, medium-term change feature extraction, and long-term trend feature extraction, respectively. Among them, the short-term branch expansion coefficient is 1, the medium-term branch expansion coefficient is 4, and the long-term branch expansion coefficient is 16; The multilayer perceptron outputs weight coefficients normalized by the Softmax function, enabling data-driven adaptive feature fusion.

8. The method according to claim 1, characterized in that, The process of obtaining deep temporal feature representations includes: The multi-scale temporal features are divided into blocks to obtain multiple feature blocks; Calculate local sensitive hash attention for each feature block to obtain the attention output for each block; The attention outputs of each block are concatenated to obtain the deep temporal feature representation.

9. The method according to claim 8, characterized in that, The process of calculating local sensitive hash attention for each feature block and obtaining the attention output for each block includes: A locality-sensitive hash attention mechanism is used to perform attention calculation on multi-scale time-series feature data. The query vector and key vector are mapped and bucketed through a hash function, and attention is calculated only within the same bucket. A reversible network layer structure is used for forward and backward propagation, eliminating the need to store intermediate activation values; The input sequence is divided into multiple blocks, self-attention is calculated independently in each block, and the attention outputs of all blocks are concatenated to obtain the deep temporal feature representation.

10. The method according to claim 1, characterized in that, The short-term power prediction results of photovoltaic power generation are a power generation prediction sequence at 15-minute intervals within the next day; The input sequence length of the historical power generation data and weather forecast data to the adversarial reversible instance normalization module is 96 time steps, with each time step representing 15 minutes. The short-term power forecast results for photovoltaic power generation are the power forecast values ​​for the next 96 time steps.

Images