Photovoltaic power prediction method and device based on enhanced deep learning

Abstract

The invention discloses a photovoltaic power prediction method and device based on enhanced deep learning, and the method comprises the steps: obtaining photovoltaic time sequence data in a historical period of time, including NWP data, LMD data and power data; the photovoltaic time sequence data passes through an input embedding layer to obtain first time sequence data representation; decomposing the first time sequence data representation into season information and trend information through a logarithm moving average function; the seasonal information passes through an improved attention network, a dynamic and static feedforward neural network and an activation function to obtain a final seasonal information learning result; processing the trend information, the second time sequence data representation and the third time sequence data representation through a gating cross fusion module to obtain a final trend information learning result; and processing the final season information learning result and the final trend information learning result through a gating learning mechanism to obtain fourth time sequence data representation, and obtaining a photovoltaic power prediction result in a period of time in the future through an output embedding layer.

Description

Technical Field

[0001] This application relates to a photovoltaic power prediction method and device based on reinforcement deep learning, belonging to the field of photovoltaic power prediction technology. Background Technology

[0002] With increasing global emphasis on environmental protection and sustainable development, photovoltaic (PV) power generation has gained popularity due to its ability to drive energy structure transformation and reduce greenhouse gas emissions. Consequently, photovoltaic power forecasting (PVPF) has become a research hotspot in recent years. PV power is a time series influenced by multidimensional characteristics such as solar radiation intensity, temperature, wind speed, humidity, and its own properties, exhibiting randomness, volatility, intermittency, and nonlinearity. Traditional PV power forecasting methods, including physical and time series analysis, have limitations in handling complex and variable PV power generation data, making it difficult to achieve high prediction accuracy.

[0003] In recent years, with the extensive research on photovoltaic power prediction methods, these methods can be broadly divided into physical methods (based on physical principles) and data-driven methods (based on traditional statistics, traditional machine learning, and deep learning). These methods use various input variables such as weather data, historical power generation data, and solar irradiance measurements to predict the output of solar photovoltaic systems.

[0004] Specifically, physical methods describe the physical behavior of solar photovoltaic (PV) systems. These methods use physical weather parameters (such as cloud cover, ambient temperature, and solar irradiance) as input vectors to the physical equations predicting power output, simulating the conversion of solar radiation into electrical energy. These equations are generally very complex, requiring a detailed understanding of the underlying physics and system characteristics of the PV system being modeled (such as installation details, electrical configuration, geographical location, and technical specifications). Furthermore, because they utilize weather conditions as input for PV power prediction, the model performance is heavily dependent on the accuracy of weather forecast data.

[0005] Data-driven methods refer to approaches that rely on the use and analysis of historical data to obtain patterns and relationships for predicting the output of solar photovoltaic (PV) systems. These methods are based on statistical data and algorithms, learning patterns and relationships from the data to make accurate predictions. However, these methods require large amounts of high-quality data to be effective, and the accuracy of predictions can be affected by data bias or incomplete data. Data-driven methods can be further subdivided into traditional statistics, traditional machine learning, and deep learning methods. Traditional statistical methods are based on statistical analysis of historical data to identify trends and patterns that can be used to predict future solar PV power output. They use statistical techniques and mathematical methods to fit historical data and derive relationships between different variables, including methods such as Autoregressive (AR), Moving Average (MA), and Autoregressive Integrated Moving Average (ARIMA). These methods generally only handle stable, reliable, and linear data, and often perform poorly with nonlinear data such as PV power output. Traditional machine learning methods can learn patterns and relationships in data, handle more complex nonlinear relationships between input and output variables, and often generate more accurate predictions than traditional statistical methods. Traditional machine learning methods employ a segregated learning strategy, primarily including Support Vector Machines (SVM) and Random Forests (RF). Their shallow structures require manual feature selection, hindering further improvements in photovoltaic power prediction performance. Deep learning methods, on the other hand, can automatically learn and extract features from input data, reducing manual intervention and workload. They can handle large-scale, high-dimensional data and are suitable for various complex photovoltaic systems. Furthermore, deep learning possesses strong nonlinear modeling capabilities, better capturing the complex relationships in the photovoltaic power generation process. Deep learning-based photovoltaic power prediction predicts the output power of a photovoltaic system over a future period by analyzing historical and real-time meteorological data. This includes Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN) and their variants, Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), Transformers and their variants, as well as hybrid models combining two or more deep learning methods. CNNs use local features extracted from sequential data for prediction, but they may encounter difficulties when dealing with long sequence data.RNNs and their variants are used for sequence prediction problems because they can handle the temporal dynamics of input data and capture sequential dependencies in time series, but they struggle to capture long-term dependencies. Transformers, through their self-attention mechanism, can capture long-distance dependencies, thus improving their predictive performance, but this mechanism is more complex. Self-attention is a model that mimics human visual attention, allowing the model to focus on important parts of information while ignoring less important ones. Its core idea is to determine which parts should receive more weight by calculating the correlation or similarity between different parts of the model. In sequence models, this typically involves calculating the relationships between queries, keys, and values, and generating weighted outputs based on these relationships. Existing Transformers and their variants often embed multiple variables at the same timestamp into indistinguishable channels, focusing attention on these timestamps to capture their temporal dependencies. However, multiple variables at the same timestamp represent completely different physical meanings recorded by inconsistent measurements, and these variables are embedded in a feature designed to eliminate multivariate correlations. Due to problems such as excessively localized receptive fields and temporal misalignment at the same time point, features formed by a single time step may fail to reveal their implicit information. Furthermore, the permutation-invariant attention mechanism employed by these methods ignores the influence of sequence order. Therefore, the performance of the Transformer in capturing sequence representations and depicting multivariate correlations is weakened, limiting its generalization ability to different photovoltaic data and failing to fully exploit the inherent coupling relationships between data, resulting in insufficient prediction accuracy. In addition, the Transformer also has quadratic complexity, limiting its deployment in resource-constrained scenarios and real-time applications. Hybrid models, combining the feature extraction capabilities and time-series data processing capabilities of two or more deep learning methods, can improve prediction accuracy, but may increase model complexity, prolong training time, increase the risk of overfitting, increase the difficulty of parameter tuning, and limit generalization ability.

[0006] Given the numerous problems with the aforementioned physical and data-driven methods, realizing a photovoltaic power prediction method based on reinforcement deep learning has become a challenging task. Summary of the Invention

[0007] Objective: In view of at least one of the above technical problems, this application provides a photovoltaic power prediction method and apparatus based on reinforcement deep learning.

[0008] The technical solution adopted in this application is:

[0009] Firstly, this application provides a photovoltaic power prediction method based on reinforcement deep learning, including:

[0010] Acquire photovoltaic time-series data over a historical period, including NWP data, LMD data, and power data. NWP data includes global radiation, direct radiation, temperature, humidity, wind speed, wind direction, and air pressure, while LMD data includes global radiation, diffuse radiation, temperature, air pressure, wind direction, and wind speed.

[0011] The photovoltaic time series data is input into the embedding layer to obtain the first time series data representation;

[0012] The first time series data is decomposed into seasonal and trend information using an exponential-log moving average function.

[0013] The seasonal information is processed through an improved attention network to obtain a second temporal data representation;

[0014] The second time-series data representation is processed by a dynamic-static feedforward neural network to obtain the third time-series data representation.

[0015] The third temporal representation is processed through an activation function to obtain the final seasonal information learning result;

[0016] The trend information, the second time series data representation, and the third time series data representation are processed by a gated cross-fusion module to obtain the final trend information learning result.

[0017] The final seasonal information learning results and the final trend information learning results are processed through a gating learning mechanism to obtain the fourth time series data representation;

[0018] The fourth time-series data characterization, after passing through the output embedding layer, yields photovoltaic power prediction results for a future period.

[0019] Secondly, this application provides a photovoltaic power prediction device based on reinforcement deep learning, including a processor and a storage medium;

[0020] The storage medium is used to store instructions;

[0021] The processor is configured to operate according to the instructions to execute the method according to the first aspect.

[0022] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0023] Fourthly, this application provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0024] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0025] Beneficial Effects: The photovoltaic power prediction method and device based on reinforcement deep learning provided in this application have the following advantages: It can deeply mine long-term dependency information, dynamic information, and static information in complex nonlinear time-series data, while reducing the corresponding time complexity, and has good time-series prediction capabilities. It improves the accuracy and precision of photovoltaic power prediction. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a photovoltaic power prediction method based on reinforcement deep learning according to an embodiment of this application. Detailed Implementation

[0027] The present application will be further described below with reference to the accompanying drawings and embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and should not be used to limit the scope of protection of the present application.

[0028] In the description of this application, "several" means one or more, "multiple" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0029] In the description of this application, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0030] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0031] Example 1: This example provides a photovoltaic power prediction method based on reinforcement deep learning, such as... Figure 1 As shown, it includes:

[0032] Step S1: Obtain photovoltaic time-series data over a historical period, including NWP data, LMD data, and power data;

[0033] It should be noted that NWP data includes global radiation, direct radiation, temperature, humidity, wind speed, wind direction, and air pressure, while LMD data includes global radiation, diffuse radiation, temperature, air pressure, wind direction, and wind speed.

[0034] Step S2: The photovoltaic time series data is passed through the input embedding layer to obtain the first time series data representation.

[0035] Furthermore, step S2 specifically includes:

[0036] ;

[0037] in, The input is photovoltaic time-series data. It is the first time series data representation. For gated Swish activation functions, It is the Swish activation function. and It is a learnable matrix. Represents the product of matrices.

[0038] It should be noted that the input embedding layer uses a gated Swish activation function, which can effectively learn the temporal dependencies of time series data, and at the same time map the time series data to space for easy processing.

[0039] Step S3: The first time series data representation is decomposed into seasonal information and trend information through the exponential-logarithmic moving average function.

[0040] Furthermore, step S3 specifically includes:

[0041] ;

[0042] in, It is the first time series data representation. It is trend information;

[0043] Yes The processed logarithmic moving average function specifically includes:

[0044] set up It contains photovoltaic time-series data at t time points. They are exist The value of the time, then go through After function processing exist Values ​​at any given time They are respectively:

[0045] ;

[0046] ;

[0047] ;

[0048] ;

[0049] ;

[0050] in, It is a logarithmic function. and These are super parameters that can be manually adjusted;

[0051] The trend information is obtained using the above formula. Then, the corresponding seasonal information is obtained using the interpolation method:

[0052] ;

[0053] in, It's seasonal information. It is the first time series data representation. It's trend information.

[0054] It should be noted that the exponential-log moving average function can effectively decompose time series data into seasonal and trend information, thereby learning different time models in the time series data, simplifying complex problems, and achieving the goal of divide and conquer.

[0055] Step S4: The seasonal information is processed through an improved attention network to obtain a second time-series data representation.

[0056] Furthermore, step S4 specifically includes:

[0057] ;

[0058] in, It's seasonal information. It is a second time-series data representation;

[0059] Yes Improved attention networks for processing, specifically including:

[0060] ;

[0061] in, It is an activation function. yes Dimensions It is a query matrix. It is a key matrix. It is a value matrix. It is a pulse network function. It is a depthwise separable convolution function. It is a point-by-point separable convolution function. It is the Kolmogorov–Arnold network.

[0062] It should be noted that the improved attention network combines depthwise separable convolutional functions, spiking network functions, pointwise separable convolutional functions, Kolmogorov–Arnold networks, and self-attention mechanisms, thereby increasing the model's long-term predictive ability.

[0063] Step S5: The second time series data representation is processed through a dynamic-static feedforward neural network to obtain the third time series data representation.

[0064] Furthermore, step S5 specifically includes:

[0065] ;

[0066] in, It is a second time-series data representation. It is a third-order time-series data representation;

[0067] It is aimed at The processed dynamic and static feedforward neural networks specifically include:

[0068] ;

[0069] in, Represents matrix multiplication. This represents the Hadamard product. This represents a one-dimensional convolutional neural network. This indicates a depthwise separable convolution function. It is a depthwise separable convolution function. It is a fully connected function. This is the average pooling function.

[0070] It should be noted that dynamic-static feedforward neural networks differ from traditional feedforward neural networks. They can learn time series data from both dynamic and static perspectives, thereby extracting dynamic and static information from the time series data and improving the performance of time series prediction.

[0071] Step S6: Pass the third temporal representation through an activation function to obtain the final seasonal information learning result.

[0072] Furthermore, step S6 specifically includes:

[0073] ;

[0074] in For the final seasonal information learning results, For the activation function, the actual selection Activation function.

[0075] It should be noted that activation functions can process the final seasonal information, thereby improving the model's non-linear learning ability.

[0076] Step S7: The trend information, the second time series data representation, and the third time series data representation are processed by the gated cross-fusion module to obtain the final trend information learning result.

[0077] Furthermore, step S7 specifically includes:

[0078] First, trend information After processing by the gating attention module, including:

[0079] (1) Trend Information The first trend information is obtained after trend normalization. ;

[0080] ;

[0081] in, Trend normalization function: , It is an exponential moving average function. For learnable parameters, for The variance, t is The total number of time points, i is The i-th time value The serial number, yes The mean, It is a learnable trend polynomial. and It is a learnable value;

[0082] (2) First trend information The second trend information is obtained through learning via a dual-gated deep learning network. ;

[0083] ;

[0084] in, For activation function, This represents a one-dimensional convolutional neural network. This indicates a depthwise separable convolution function. It is a depthwise separable convolution function. It is a point-by-point separable convolution function. It is a partial convolution function;

[0085] Secondly, the second time series data representation The frequency domain enhancement deep learning module obtained ;

[0086] ;

[0087] in, To employ a frequency domain-enhanced deep learning module for the representation of the second time-series data The result of learning For Fast Fourier Transform, To extend the concatenation operation of depthwise separable convolution functions, batch normalization functions, and ReLU activation functions, This represents a one-dimensional convolutional neural network. This is the inverse fast Fourier transform;

[0088] Third, the third time series data representation After multi-scale linear attention Mamba processing, the following was obtained. ;

[0089] ;

[0090] in, Multi-scale linear attention Mamba for characterizing third-time series data The result of learning For activation function, This represents a one-dimensional convolutional neural network. For gating mechanism, For linear attention functions, It is a multilayer perceptron. It is a 5-dimensional depthwise separable convolution function. It is a 3D depthwise separable convolution function;

[0091] Finally, , and A multi-expert mechanism is used for fusion to obtain the final trend information learning result. ;

[0092] ;

[0093] in, In response to , and Multi-expert functions.

[0094] It should be noted that the gated cross-fusion module combines the gated attention module, the frequency domain enhanced deep learning module, and the multi-scale linear attention Mamba module to further improve the model's ability to learn trend information.

[0095] Step S8: The final seasonal information learning results and the final trend information learning results are processed through a gating learning mechanism to obtain the fourth time series data representation.

[0096] Furthermore, step S8 specifically includes:

[0097] ;

[0098] in This is for the fourth time series data representation. For activation function, For the final seasonal information learning results, For matrix addition, This represents the final trend information learning result.

[0099] It should be noted that the gating learning mechanism can effectively improve the model's learning ability.

[0100] Step S9: The fourth time series data characterization is processed by the output embedding layer to obtain the photovoltaic power prediction results for a future period of time.

[0101] Furthermore, step S9 specifically includes:

[0102] ;

[0103] Where Y represents the photovoltaic power prediction result. For self-attention functions, The root mean square layer normalization function, It is a fully connected function.

[0104] Specific simulation application examples: Table 1 shows the quantitative indicators of the method disclosed in this invention compared with other time-series forecasting methods in photovoltaic power prediction. The comparison methods selected in the experiment are classic methods such as Transformer, Autoformer, Informer, FEDformer, Reformer, Flowformer, Flashformer, iTransformer, LSTM, and GRU. The quantitative indicators selected are Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). The experiment used measured data from two different photovoltaic power stations in Hebei Province, collected from 16:00:00 on August 15, 2018 to 15:45:00 on June 13, 2019, and from 16:00:00 on July 22, 2018 to 15:45:00 on May 21, 2019, respectively. The installed capacities were 6.6MW and 20MW, respectively. Data included NWP (Numerical Weather Prediction) data (7 features: global radiation, direct radiation, temperature, humidity, wind speed, wind direction, and air pressure), LMD (Low Mite Measurement) data (6 features: global radiation, diffuse radiation, temperature, air pressure, wind direction, and wind speed), and power data, collected every 15 minutes. Because the multi-dimensional features of radiation, temperature, humidity, wind speed, wind direction, and air pressure included in the above data can provide a fine-grained characterization of meteorological environments such as rainy, cloudy, and sunny days, the results of this simulation experiment can verify the adaptability of DPIT to these special conditions. The simulation experiments preprocessed the two datasets. First, a large number of missing values ​​were filled in using linear interpolation, and cases that could not be interpolated were deleted. Then, outliers were detected using box plots, and the outliers shown in the box plots were deleted and filled in using linear interpolation. Finally, the min-max normalization method was used to normalize the NWP, LMD, and power data. After preprocessing, the data from the first 60% of the days in both datasets were used as the training set, the data from the last 20% of the days were used as the test set, and the data from the remaining days were used as the validation set. PyTorch was chosen as the implementation framework for the experiments, and a GTX 3090 GPU was used for training and inference. To verify the prediction performance of the proposed method on ultra-short-term, short-term, and medium-term data, this paper used data from the past 96 points (24 hours) to predict data from the next 96 points (24 hours) and 720 points (180 hours), respectively. In addition, this paper selects appropriate parameters through a large number of experiments, with an epoch of 10, a batch size of 32, a learning rate of 0.0001, a dropout rate of 0.1, and 2 IIT module layers.

[0105] Table 1:

[0106]

[0107] As shown in Table 1, for power plant 1, the MAE and RMSE of the proposed method are 0.4578 and 0.7931, respectively; for power plant 2, the MAE and RMSE of the proposed DPIT are 0.4216 and 0.6956, respectively. The proposed method has the lowest MAE and RMSE values ​​at both sites, demonstrating the best prediction results and the highest prediction accuracy. The iTransformer method performs second best, while other Transformer-based methods such as Transformer, Autoformer, and Informer have higher MAE and RMSE values ​​compared to iTransformer and DPIT. Meanwhile, traditional deep learning methods, represented by LSTM and GRU, have the worst prediction performance. This is because these two types of methods have poor learning capabilities due to limitations in their mechanisms. Furthermore, the prediction accuracy of Transformer-based photovoltaic power prediction methods such as Transformer, Autoformer, Informer, FEDformer, Reformer, Flowformer, Flashformer, and iTransformer is superior to that of RNN-based photovoltaic power prediction methods such as LSTM and GRU.

[0108] Example 2: Based on Example 1, this example provides a photovoltaic power prediction device based on reinforcement deep learning, including a processor and a storage medium;

[0109] The storage medium is used to store instructions;

[0110] The processor is configured to operate according to the instructions to execute the method according to Embodiment 1.

[0111] Example 3: Based on Example 1, this example provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in Example 1.

[0112] Example 4: Based on Example 1, this example provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in Example 1.

[0113] Example 5: Based on Example 1, this example provides a computer program product, including a computer program that, when executed by a processor, implements the method described in Example 1.

[0114] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0115] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0116] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0117] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0118] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A photovoltaic power prediction method based on reinforcement deep learning, characterized in that, include: Acquire photovoltaic time-series data over a historical period, including NWP data, LMD data, and power data. NWP data includes global radiation, direct radiation, temperature, humidity, wind speed, wind direction, and air pressure, while LMD data includes global radiation, diffuse radiation, temperature, air pressure, wind direction, and wind speed. The photovoltaic time series data is input into the embedding layer to obtain the first time series data representation; The first time series data is decomposed into seasonal and trend information using an exponential-log moving average function. The seasonal information is processed through an improved attention network to obtain a second temporal data representation; The second time-series data representation is processed by a dynamic-static feedforward neural network to obtain the third time-series data representation. The third temporal representation is processed through an activation function to obtain the final seasonal information learning result; The trend information, the second time series data representation, and the third time series data representation are processed by a gated cross-fusion module to obtain the final trend information learning result. The final seasonal information learning results and the final trend information learning results are processed through a gating learning mechanism to obtain the fourth time series data representation; The fourth time-series data characterization, after passing through the output embedding layer, yields photovoltaic power prediction results for a future period.

2. The method according to claim 1, characterized in that, The photovoltaic time series data is processed through an input embedding layer to obtain the first time series data representation, including: ； in, The input is photovoltaic time-series data. It is the first time series data representation. For gated Swish activation functions, It is the Swish activation function. and It is a learnable matrix. Represents the product of matrices.

3. The method according to claim 1, characterized in that, The first time series data is decomposed into seasonal and trend information using an exponential-logarithmic moving average function, including: ； in, It is the first time series data representation. It is trend information; Yes The processed logarithmic moving average function specifically includes: set up It contains photovoltaic time-series data at t time points. They are exist The value of the time, then go through After function processing exist Values ​​at any given time They are respectively: ； ； ； ； ； in, It is a logarithmic function. and These are super parameters that can be manually adjusted; The trend information is obtained using the above formula. Then, the corresponding seasonal information is obtained using the interpolation method: ； in, It's seasonal information. It is the first time series data representation. It's trend information.

4. The method according to claim 1, characterized in that, The seasonal information is processed through an improved attention network to obtain a second temporal data representation, including: ； in, It's seasonal information. It is a second time-series data representation; Yes Improved attention networks for processing, specifically including: ； in, It is an activation function. yes Dimensions It is a query matrix. It is a key matrix. It is a value matrix. It is a pulse network function. It is a depthwise separable convolution function. It is a point-by-point separable convolution function. It is the Kolmogorov–Arnold network.

5. The method according to claim 1, characterized in that, The second time-series data representation is processed through a dynamic-static feedforward neural network to obtain the third time-series data representation, including: ； in, It is a second time-series data representation. It is a third-order time-series data representation; It is aimed at The processed dynamic and static feedforward neural networks specifically include: ； in, Represents matrix multiplication. This represents the Hadamard product. This represents a one-dimensional convolutional neural network. This indicates a depthwise separable convolution function. It is a depthwise separable convolution function. It is a fully connected function. This is the average pooling function.

6. The method according to claim 1, characterized in that, The third temporal representation is processed through an activation function to obtain the final seasonal information learning result, including: ； in For the final seasonal information learning results, For the activation function, the actual selection Activation function.

7. The method according to claim 1, characterized in that, The trend information, second time-series data representation, and third time-series data representation are processed by a gated cross-fusion module to obtain the final trend information learning result, including: First, trend information After processing by the gating attention module, including: (1) Trend Information The first trend information is obtained after trend normalization. ; ； in, Trend normalization function: , It is an exponential moving average function. For learnable parameters, for The variance, t is The total number of time points, i is The i-th time value The serial number, yes The mean, It is a learnable trend polynomial. and It is a learnable value; (2) First Trend Information The second trend information is obtained through learning via a dual-gated deep learning network. ; ； in, For activation function, This represents a one-dimensional convolutional neural network. This indicates a depthwise separable convolution function. It is a depthwise separable convolution function. It is a point-by-point separable convolution function. It is a partial convolution function; Secondly, the second time series data representation The frequency domain enhancement deep learning module obtained ; ； in, To employ a frequency domain-enhanced deep learning module for the representation of the second time-series data The result of learning For Fast Fourier Transform, To extend the concatenation operation of depthwise separable convolution functions, batch normalization functions, and ReLU activation functions, This represents a one-dimensional convolutional neural network. This is the inverse fast Fourier transform; Third, the third time series data representation After multi-scale linear attention Mamba processing, the following was obtained. ; ； in, Multi-scale linear attention Mamba for characterizing third-time series data The result of learning For activation function, This represents a one-dimensional convolutional neural network. For gating mechanism, For linear attention functions, It is a multilayer perceptron. It is a 5-dimensional depthwise separable convolution function. It is a 3D depthwise separable convolution function; Finally, , and A multi-expert mechanism is used for fusion to obtain the final trend information learning result. ; ； in, In response to , and Multi-expert functions.

8. The method according to claim 1, characterized in that, The final seasonal information learning results and the final trend information learning results are processed through a gating learning mechanism to obtain the fourth time-series data representation, including: ； in This is for the fourth time series data representation. For activation function, For the final seasonal information learning results, For matrix addition, This represents the final trend information learning result.

9. The method according to claim 1, characterized in that, The fourth time-series data representation, after passing through the output embedding layer, yields photovoltaic power prediction results, including: ； Where Y represents the photovoltaic power prediction result. For self-attention function, The root mean square layer normalization function, It is a fully connected function.

10. A photovoltaic power prediction device based on reinforcement deep learning, characterized in that, Including processor and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the method according to any one of claims 1 to 9.

Images