Multi-scale photovoltaic generating capacity prediction method fusing explicit periodic modeling and KAN-MLP mixed structure

Abstract

The invention discloses a multi-scale photovoltaic power generation capacity prediction method fusing explicit periodic modeling and a KAN-MLP mixed structure, and the method comprises the steps: carrying out the normalization processing of multivariable input data collected by a photovoltaic power station, decomposing the input time series data into a trend term, a periodic term and a disturbance term through a Fourier decomposition and hierarchical downsampling mode; designing a learnable periodic compensation module, and introducing an offset vector for each time position by training a periodic matrix to realize phase alignment and feature compensation of periodic signals; aiming at a dependency relationship among multiple variables, a channel interaction mechanism of star topology is adopted; the features after multi-scale and multi-variable fusion are input into a double-branch structure, and finally adaptive weighted fusion is carried out on two branch outputs through an attention mechanism; a trend error perception item is introduced to correct the traditional mean square error loss, so that the model pays more attention to the consistency of a trend item and a periodic item in training. According to the invention, high-precision and robustness prediction of the photovoltaic generating capacity is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence and new energy power generation prediction, and particularly relates to a multi-scale photovoltaic power generation prediction method. BACKGROUND

[0002] With the continuous advancement of industrialization, the reserves of traditional fossil fuels are decreasing, and are facing the severe challenge of exhaustion. Under this background, solar energy, as a green and renewable clean energy, has attracted widespread attention. Among them, photovoltaic power generation is the most common and widely used solar energy utilization technology. Accurate and reliable prediction of photovoltaic power generation is not only a key technical means to improve the efficiency of new energy access, but also plays an important role in ensuring the stable operation of the power grid, optimizing the dispatching plan, reducing the light rejection rate and improving the utilization efficiency of energy storage. With the continuous improvement of new energy penetration rate, its strategic position in the modern energy system is increasingly prominent. Photovoltaic power generation is affected by many meteorological factors, such as solar radiation intensity, temperature, wind speed and relative humidity. These factors have high uncertainty and are difficult to intervene artificially. In addition, photovoltaic power generation has significant diurnal periodicity, showing strong randomness, volatility and intermittency. This poses a great challenge to the accurate prediction of photovoltaic power generation. Therefore, it is of great practical significance and application value to carry out in-depth research and accurate prediction of photovoltaic power generation.

[0003] At present, many photovoltaic power generation prediction methods have been proposed by domestic and foreign scholars, mainly including physical models, statistical methods, machine learning methods and deep learning methods. The first method is based on the basic principle of photovoltaic power generation, which maps the meteorological input to the power generation output through a physical formula, such as numerical weather prediction models and satellite cloud evolution models. Although this method does not rely on a large amount of historical data, the model structure is complex, the adaptability to external disturbances is poor, and the parameters need to be adjusted regularly to maintain the prediction performance. The second method relies on historical observation data, and builds a statistical relationship between meteorological variables and power generation through time series analysis, regression analysis and other means, such as statistical models, regression analysis, grey models and autoregressive models. This method has certain interpretability, but when facing sudden weather changes or non-stationary sequences, the prediction stability and accuracy are poor, and it is sensitive to data quality. The third method is based on machine learning, which extracts features from historical data and builds a supervised or unsupervised learning model for modeling and prediction. This method has good generalization ability and stability in small sample scenarios, and has certain interpretability, but it depends heavily on feature engineering, and its performance is limited in dealing with complex spatio-temporal dynamic relationships, and it is easy to fall into a local optimal solution in the training process. In summary, the traditional three methods generally have poor adaptability, weak generalization ability or limited expression ability, which makes it difficult to meet the demand for high-precision prediction of modern photovoltaic systems.

[0004] The deep learning method (the fourth method) has shown significant advantages in photovoltaic power generation prediction in recent years due to its excellent modeling ability and robustness. By constructing a multi-layer neural network, the deep learning model can effectively capture the nonlinear dynamic relationship between meteorological factors and power generation, reduce the dependence on artificial feature construction, and show stronger adaptability and prediction accuracy. However, the existing deep learning model still has shortcomings: on the one hand, in the photovoltaic power generation prediction scenario, the power generation is affected by multiple factors such as day-night cycle, meteorological fluctuation and equipment characteristics, and has significant periodicity and trend. If only relying on implicit feature learning, the model is difficult to capture stable periodic rules and long-time dependence, thereby affecting the stability and generalization performance of the prediction. Therefore, it is necessary to introduce an explicit modeling mechanism to decompose and reconstruct the periodic components, trend items and noise features of the sequence in an interpretable manner, thereby establishing a closer relationship between physical priori and data-driven; on the other hand, the mainstream method mostly uses Transformer and Long Short-Term Memory (LSTM), which has strong expression ability, but its structure is complex, the parameter size is large and the interpretability is weak, which is difficult to balance between precision, computational efficiency and controllability. In recent years, although some research has tried to introduce Kolmogorov-Arnold Network (KAN) into the field of power system, such as achieving good accuracy and interpretability performance in power load prediction, state estimation and other tasks, but the research on photovoltaic power generation prediction is still relatively limited. Existing works such as RFKAN model have explored its feasibility in photovoltaic power prediction, but most researches still stay at the method verification level, and have not systematically combined the multivariate meteorological driving features and the periodicity of photovoltaic power generation for modeling.

[0005] With the development of deep learning, recurrent neural networks, convolutional neural networks, and Transformer based on attention mechanism have gradually become mainstream. These methods can alleviate the dependence on artificial features to a large extent and capture nonlinear time series features, thus showing better performance in photovoltaic prediction. Some people use MLP network to establish a nonlinear activation function to realize feature cross and output prediction value. Because MLP cannot capture the periodicity and trend of photovoltaic power generation, the prediction accuracy is limited. Some people use deep convolutional neural network (CNN) and meta-learning to predict photovoltaic power generation. Although CNN is good at spatial feature extraction, it is not good at modeling long-term dependence of time series and needs to rely on additional structures (such as LSTM) to make up for it. Some people use LSTM for photovoltaic power prediction, and combine it with a gray model to improve the accuracy. However, the model has poor generalization ability and low processing efficiency. In order to overcome the limitations of single model structure, researchers have widely explored the integration of learning and multi-module collaborative modeling strategies in recent years, combining multiple network structures to achieve the purpose of complementation. Some people propose a LSTM-CNN joint model, which first uses LSTM to extract the time dependence information of the sequence, and then uses CNN to further mine the local spatial features of the hidden state. Some people build a CNN-LSTM-XGBoost multi-model fusion structure to improve the overall prediction accuracy and robustness. However, in the above methods, there is a lack of interaction mechanism between sub-models, and the depth coupling of information cannot be realized. Some people use attention mechanism to pay attention to key information and improve the credibility of information. However, existing deep learning models still have obvious shortcomings in many aspects. On the one hand, they lack explicit modeling of day-night and seasonal cycles in photovoltaic power generation sequences, often relying on implicit learning to capture periodic information, which leads to phase shift in the prediction results at peak and valley, and the trend change is not accurate enough. On the other hand, photovoltaic data has multi-scale characteristics, including short-term high-frequency fluctuations and long-term trend changes. Existing methods still have shortcomings in extracting and utilizing multi-scale information. Although the method introduces sequence decomposition or multi-scale mixing strategy, there are still limitations in the fine description of periodic structure and the flexibility of scale division.

[0006] In addition, photovoltaic power generation prediction usually involves multiple meteorological variables, and there is a complex correlation between the variables. Existing channel modeling methods are roughly divided into two categories: channel independent modeling and channel dependent modeling. Although channel independent modeling can reduce network complexity and improve robustness in distribution drift scenarios, it ignores the potential dependence between variables; channel dependent modeling methods capture the interaction information between multiple variables through attention mechanisms and other means, but the high computational complexity limits its application in large-scale scenarios, and it is easy to introduce redundant features under noise disturbance, resulting in a decrease in prediction performance. At the same time, photovoltaic power generation sequences are often influenced by multiple factors such as day-night changes, weather disturbances and seasonal cycles, and often contain rich multi-scale structures. In recent years, multi-scale modeling methods have shown certain effectiveness in long-term prediction, but they still face problems such as inaccurate cycle modeling, rigid scale division, and insufficient response to high-frequency disturbances in the photovoltaic scenario.

[0007] In summary, although the existing technology has made certain progress in the accuracy and adaptability of photovoltaic power generation prediction, it still generally has problems such as insufficient cycle modeling, insufficient use of multi-scale features, low multi-variable fusion efficiency, high model complexity, and poor interpretability. These deficiencies limit its application effect in actual photovoltaic power station prediction and scheduling scenarios, and there is an urgent need for a new prediction method that can balance periodic and trend modeling, multi-scale feature extraction, multi-variable efficient fusion, and has strong generalization ability. SUMMARY

[0008] In order to overcome the problems of insufficient cycle modeling, difficulty in accurately capturing day-night and seasonal patterns, low multi-variable fusion efficiency, high computational complexity, and difficulty in balancing nonlinear fitting ability and computational efficiency in existing photovoltaic power generation prediction methods, the present application provides a multi-scale photovoltaic power generation prediction method that combines explicit cycle modeling and KAN-MLP hybrid structure. By decomposing the trend and cycle terms at different time granularities and explicitly compensating for the loss of cycle information caused by downsampling, the present application realizes fine modeling of periodic and trend features, effectively avoiding phase shift and trend distortion in the predicted curve. The present application proposes a star-shaped channel interaction structure, which realizes global dependence modeling between multiple variables through the interaction mode of the center channel and the edge channel, while significantly reducing computational overhead, thereby improving the running efficiency and adaptability of the model while ensuring prediction accuracy. The present application constructs a dual-branch prediction structure of KAN and MLP, and introduces an attention mechanism to adaptively weight and fuse the outputs of the two branches, so that the model can not only utilize the nonlinear expression ability of KAN, but also utilize the efficient feature combination ability of MLP, thereby improving the overall prediction performance.

[0009] The technical scheme adopted by the present application to solve its technical problems is:

[0010] A multi-scale photovoltaic power generation prediction method fusing explicit periodic modeling and KAN-MLP hybrid structure, comprising the following steps:

[0011] Step 1: Normalizing the multi-variable input data (including solar radiation, temperature, humidity, wind speed, historical power generation, etc.) collected by the photovoltaic power station to automatically eliminate the numerical scale difference of different variables and alleviate the volatility of the time series in a long time span; Since the purpose of this stage is to unify the distribution of input features, so that the model can more stably perform subsequent feature extraction and training, the normalization method can quickly realize data standardization and improve the convergence efficiency of the model, and optionally, the normalization method can adopt instance normalization, standardization or minimum-maximum scaling, etc.

[0012] Step 2: The input time series data is decomposed into trend items, periodic items and disturbance items through Fourier decomposition and hierarchical downsampling, so as to capture dynamic features at different time scales; Since photovoltaic power generation contains both long-term trends and short-term fluctuations, hierarchical downsampling can effectively separate signal components of different scales.

[0013] Optionally, the decomposition method can be sliding average, convolution decomposition or multi-scale decomposition based on wavelet transform. The number of downsampling layers can be selected (2, 4, 6, 8).

[0014] Step 3: To avoid the loss of periodic information caused by downsampling, a learnable periodic compensation module is designed, which introduces an offset vector for each time position through training a periodic matrix, realizing phase alignment and feature compensation of periodic signals.

[0015] Optionally, the periodic compensation method can be matrix lookup compensation, position coding compensation or periodic embedding based on a sine function.

[0016] Step 4: For the dependent relationship between multiple variables, a channel interaction mechanism with star topology is adopted, which realizes efficient fusion between multiple variables through the interaction path of the center channel and the edge channel, while keeping low computational complexity.

[0017] Optionally, the channel interaction method can be attention mechanism, graph neural network modeling or convolution fusion.

[0018] Step 5: The multi-scale and multi-variable fused features are input into a double-branch structure, one branch of which is a Kolmogorov-Arnold network (KAN) for nonlinear function fitting, and the other branch is a multi-layer perception (MLP) for efficient feature combination, and finally the two branches are adaptively weighted and fused through an attention mechanism.

[0019] Step 6: Loss function improvement In the prediction stage, a trend error awareness term is introduced to modify the traditional mean square error loss, so that the model pays more attention to the consistency of the trend and periodic terms in the training process, thereby improving the overall prediction accuracy and robustness.

[0020] The beneficial effects of the present application mainly include:

[0021] 1) Through multi-scale decomposition and a learnable period compensation mechanism, the present application can simultaneously extract the periodic and trend information of photovoltaic power generation sequences at different time granularities, and explicitly compensate for the periodic features that may be lost in the downsampling process, thereby achieving fine modeling of photovoltaic power generation changes and effectively improving prediction accuracy.

[0022] 2) By introducing a star-shaped channel interaction structure, the present application realizes efficient fusion and global dependence modeling among multiple variables while maintaining low computational complexity, avoiding the defects of independent channel modeling that ignore variable relationships, and overcoming the problem of excessive computational overhead of dependent channel modeling, thereby improving the efficiency of the prediction model.

[0023] 3) By constructing a dual-branch structure of KAN and MLP and using attention mechanisms for adaptive fusion, the present application fully utilizes the advantages of KAN in nonlinear function fitting and the advantages of MLP in efficient feature combination, significantly improving the model's expression ability and generalization performance, and ensuring the accuracy and stability of the prediction results.

[0024] 4) By introducing a trend error awareness term in the loss function, the present application enables the model to simultaneously focus on the consistency of periodic and trend components during the training process, further reducing the deviation between the predicted results and the true data in terms of trend changes, thereby improving the robustness and practicality under complex weather conditions and different application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0025] Fig. 1 is a multi-scale feature fusion method flowchart.

[0026] Fig. 2 is a multi-channel feature star topology fusion method flowchart.

[0027] Fig. 3 is an explicit period and KAN-MLP feature fusion method flowchart. DETAILED DESCRIPTION

[0028] The present application will be further described below with reference to the accompanying drawings.

[0029] Referring to Figs. 1-3 A multi-scale photovoltaic power generation prediction method that combines explicit period modeling and KAN-MLP hybrid structure includes the following steps:

[0030] Step 1: Normalize the multi-variable input data collected by the photovoltaic power station (including solar radiation, temperature, humidity, wind speed, historical power generation, etc.) to automatically eliminate the numerical scale difference of different variables and alleviate the volatility of time series over a long time span; since the purpose of this stage is to unify the distribution of input features, making the model more stable for subsequent feature extraction and training, the normalization method can quickly realize data standardization and improve the convergence efficiency of the model. Optionally, the normalization method can use instance normalization, standardization or minimum-maximum scaling, etc.

[0031] Specifically, in step 1, the purpose is to weaken the volatility and non-stationarity of time series over a long time span, avoid direct modeling leading to a decline in key feature capture ability, and thus improve the accuracy of photovoltaic power generation prediction. The process is as follows:

[0032] Step 1.1: Obtain the continuously collected photovoltaic power generation related time series data, which includes meteorological features and power generation features.

[0033] Step 1.2: Identify the volatility and non-stationarity features of the time series, such as variance fluctuations caused by the decrease in component efficiency due to rising temperature, or significant fluctuations in power generation caused by changes in solar radiation intensity.

[0034] Step 1.3: For the input time series data, calculate the mean and standard deviation along the time dimension, and perform instance normalization on the time series based on the mean and standard deviation, with the specific formula as follows:

[0035]

[0036] where X represents the input vector Mean L represents the operation of calculating the mean along the time dimension, Std L represents the operation of calculating the standard deviation along the time dimension, i represents the current channel, and L represents the channel length. After normalization, the sequence data distribution is more stable, which is beneficial for capturing periodic and trend features in the next step. At the same time, the model can better capture local dynamic features, thereby improving the anti-interference ability of the model.

[0037] Step 1.4: Use the normalized time series as input data for the subsequent modeling stage.

[0038] Step 2: Decompose the input time series data into trend items, periodic items and disturbance items through Fourier decomposition and hierarchical downsampling, in order to capture dynamic features at different time scales; since photovoltaic power generation contains both long-term trends and short-term fluctuations, hierarchical downsampling can effectively separate signal components of different scales.

[0039] Optionally, the decomposition method can be a sliding average, a convolution decomposition, or a multi-scale decomposition based on wavelet transform. The down-sampling layer can be selected from (2, 4, 6, 8).

[0040] Step 3: To avoid the loss of periodic information caused by down-sampling, a learnable periodic compensation module is designed. This module introduces an offset vector for each time position by training a periodic matrix, achieving phase alignment and feature compensation of periodic signals.

[0041] Optionally, the periodic compensation method can be a matrix lookup table compensation, a position coding compensation, or a periodic embedding based on a sine function.

[0042] In step 3, the purpose of multi-scale feature decomposition and periodic compensation processing is to enhance the modeling ability of the model for different time structure information and improve the capture accuracy of periodic features. The process is as follows:

[0043] Step 3.1: First, Fourier decomposition combined with a learnable convolution is used to extract the periodic features of the input sequence; then, the global periodic component is obtained through frequency domain multiplication; finally, the component is fused with the original sequence to obtain a periodic reinforced sequence.

[0044] Step 3.2: The periodic reinforced sequence is input into multiple groups of average pooling channels, each group of channels is set with different down-sampling ratios to extract local dynamic features at different time granularities; then, the pooling results are linearly mapped to restore the original dimension, and are fused with the high-resolution path through residual connection to form a step-by-step progressive multi-scale compensation path.

[0045] Step 3.3: A learnable periodic compensation module is set to alleviate the loss of periodic information caused by multiple down-sampling; in this module, m trainable periodic matrices are defined, and a periodic offset vector is dynamically generated according to the current time index at each prediction, and compensation features related to the periodic position are introduced through position cyclic value retrieval, thereby strengthening the model's expression ability for periodic patterns.

[0046]

[0047] wherein T is the input period length, N is the number of channels, t is the current time, H is the number of channels, and p is the time step.

[0048] Step 3.4: By looking up the table, it is ensured that different time steps always correspond to the same phase index, and in the training process, the periodic matrix is used to learn the deviation signal between the real data and the predicted data at each periodic position, and the offset vector is used to adaptively adjust the local time difference.

[0049] Step 4: For the dependency between multiple variables, a channel interaction mechanism with a star topology is adopted. This mechanism realizes the efficient fusion between multiple variables through the interaction path of the center channel and the edge channel, while maintaining low computational complexity.

[0050] Optionally, the channel interaction mode can be an attention mechanism, a graph neural network modeling, or a convolution fusion.

[0051] In step 4, the purpose is to solve the problem of complex dependency between multiple meteorological variables in the channel dimension in photovoltaic power generation prediction, and to improve the expression ability and generalization ability of the model. The process is as follows:

[0052] Step 4.1: In photovoltaic power generation prediction, multiple meteorological variables such as solar radiation, temperature, humidity, etc. are taken as input; linear transformation is performed on the features in each channel to obtain the corresponding channel feature representation.

[0053] Step 4.2: A center channel node is established, and an interaction path is established with each edge channel, so that each channel can obtain global channel context information through dynamic interaction with the center node as follows:

[0054] R j =Linear2(Linear1(X j )) (3);

[0055] Where R j is the feature extracted from the jth channel, X j ∈R B×D is the feature of the jth channel, and Linear is a linear layer.

[0056] Step 4.3: Calculate the attention weight for the feature of each channel, and generate the center embedding vector based on the weight; combine the representations corresponding to all channels to form a channel embedding matrix, which is used to represent the global dependency between channels.

[0057]

[0058] Where a j is the attention weight for the jth channel, obtained by Softmax calculation.

[0059] Step 4.4: Concatenate the center embedding vector with the original features of each channel; then complete the fusion through a multi-layer perceptron (MLP) to obtain the output representation of each channel. This fusion method can accurately model the complex dependency between variables and improve the expression ability and generalization ability of the model in multi-variable photovoltaic power prediction.

[0060] Step 5: The multi-scale and multi-variable fused features are input into a dual-branch structure, one branch of which is a Kolmogorov-Arnold network (KAN) for nonlinear function fitting, and the other branch is a multi-layer perceptron (MLP) for efficient feature combination. Finally, the outputs of the two branches are adaptively weighted and fused through an attention mechanism.

[0061] In step 5, the purpose is to coordinate the different features learned by the MLP network and the KAN network branches, and to improve the model's expression and generalization capabilities. The process is as follows:

[0062] Step 5.1: Use the MLP branch and the KAN branch to model the input features respectively: the MLP network realizes efficient high-dimensional feature combination and abstract expression through a simple multi-layer structure; the KAN network replaces the fixed activation function with a learnable function base (such as spline function, orthogonal polynomial, etc.), realizing flexible expression of complex segmented nonlinear structure.

[0063] Step 5.2: Concatenate the MLP output features and the KAN output features in the feature dimension to obtain the joint feature representation.

[0064] Step 5.3: Input the joint feature representation into a lightweight attention network; the attention network consists of two fully connected layers and an activation function, which is used to generate attention weights for the two branches; then, through normalization operation to ensure that the weights are dynamically allocated between the two branches.

[0065] w = Softmax(φ([X MLP ; X KAN ])) (5); where φ(·) represents a subnetwork consisting of two fully connected layers and a ReLU activation function, which is used to generate attention weights for the two branches. The Softmax layer ensures that the weights are normalized between the two branches, X MLP ∈ R B×T×d represents the MLP output features, and X KAN ∈ R B×T×d represents the KAN output features.

[0066] Step 5.4: According to the attention weights, the MLP output features and the KAN output features are weighted and fused to obtain the fused feature representation; the fused feature retains the advantages of both types of representation, realizing feature complementation.

[0067] Step 5.5: The fused features and the periodic terms output by the explicit periodic modeling branch are weighted and fused, and input into the output head module to generate the final prediction result. This can be represented by the following formula:

[0068]

[0069] where fhead (·) represents the output head module, X cycle represents the periodic term of the explicit periodic modeling branch output.

[0070] Step 6: Loss function improvement In the prediction stage, a trend error awareness term is introduced to modify the traditional mean square error loss, so that the model pays more attention to the consistency of the trend term and the periodic term in the training, thereby improving the overall prediction accuracy and robustness.

[0071] In step 6, the loss function is improved in the prediction stage. The purpose is to retain the overall fitting ability of the mean square error (MSE) while introducing a trend error awareness term to improve the modeling accuracy and robustness of the model for periodic and trend characteristics. The process is as follows:

[0072] Step 6.1: For the overall difference between the predicted value and the true value, the mean square error (MSE) is used as the basic loss function to measure the global error between the predicted sequence and the actual sequence.

[0073] Step 6.2: To enhance the modeling ability of the model for periodic and trend characteristics, a trend error awareness term is introduced based on the mean square error. The trend error awareness term calculates the difference between the predicted sequence and the actual sequence in the frequency domain by performing a fast Fourier transform (FFT) on them, thereby constraining the consistency of the model in the trend and periodic components.

[0074] Step 6.3: The mean square error and the trend error awareness term are combined by weighting to form the improved loss function. The new loss function is as follows:

[0075]

[0076] Where α ∈ [0, 1]; C represents the number of samples; y k represents the kth true value; that is, the actual data point in the time series; is the kth predicted value; FFT(y k ) represents the frequency domain after performing a fast Fourier transform on the sequence y k .

[0077] In this embodiment, the new prediction method takes into account periodic and trend modeling, multi-scale feature extraction, multi-variable efficient fusion, and has strong generalization ability, which can meet the demand of modern photovoltaic systems for high-precision prediction, and is expected to play an important role in ensuring the stability of power grid operation, optimizing dispatching plan, reducing light rejection rate and improving energy storage utilization efficiency, etc.

[0078] The embodiments of the present specification are merely illustrative of the implementation forms of the inventive concept, and are only used for the purpose of description. The protection scope of the present application should not be regarded as being limited to the specific forms shown in the embodiments, and the protection scope of the present application also includes the equivalent technical means that can be thought of by those skilled in the art according to the inventive concept.

Claims

1. A multi-scale photovoltaic power generation prediction method fusing explicit periodic modeling and KAN-MLP hybrid structure, characterized in that, The method comprises the following steps: Step 1: Normalizing the multivariate input data collected by the photovoltaic power station to automatically eliminate the numerical scale difference of different variables and alleviate the volatility of time series over a long time span; Step 2: Decomposing the input time series data into trend items, periodic items and disturbance items through Fourier decomposition and hierarchical downsampling to capture dynamic characteristics at different time scales; Step 3: Designing a learnable periodic compensation module that introduces an offset vector for each time position by training a periodic matrix to achieve phase alignment and feature compensation of periodic signals; Step 4: For the dependency relationship between multiple variables, a channel interaction mechanism with a star topology is adopted; Step 5: The multi-scale and multi-variable fused features are input into a double-branch structure, one branch of which is a Kolmogorov-Arnold network for nonlinear function fitting, and the other branch is a multilayer perceptron for efficient feature combination, and finally the two branches are adaptively weighted and fused through an attention mechanism; Step 6: In the prediction stage, the loss function is improved by introducing a trend error perception term to modify the traditional mean square error loss, so that the model pays more attention to the consistency of trend items and periodic items during training.

2. The method of claim 1, wherein the method is a multi-scale photovoltaic power generation prediction method of fusion explicit periodic modeling and KAN-MLP hybrid structure. The process of step 1 is as follows: Step 1.1: Obtain the continuously collected photovoltaic power generation related time series data, which includes meteorological features and power generation features; Step 1.2: Identify the volatility and non-stationarity characteristics of the time series, including variance fluctuations caused by the decrease in component efficiency due to rising temperature or significant fluctuations in power generation caused by changes in solar radiation intensity; Step 1.3: For the input time series data, calculate the mean and standard deviation along the time dimension, and normalize the time series based on the mean and standard deviation. The normalized sequence data distribution is more stable, which is conducive to capturing periodic and trend characteristics in the subsequent model, and the model can better capture local dynamic characteristics, thereby improving the model's anti-interference ability; Step 1.4: The normalized time series is used as input data for the subsequent modeling stage.

3. The multi-scale photovoltaic power generation prediction method of fusing explicit periodic modeling and KAN-MLP hybrid structure according to claim 1 or 2, characterized in that, The process of step 3 is as follows: Step 3.1: First, use Fourier decomposition combined with learnable convolution to extract the periodic characteristics of the input sequence; then, obtain the global periodic component through frequency domain multiplication; finally, fuse the component with the original sequence to obtain a periodic reinforced sequence; Step 3.2: Input the periodic reinforced sequence into multiple average pooling channels, each channel with a different downsampling ratio to extract local dynamic characteristics at different time granularities; Then, linearly map the pooling results back to the original dimension and fuse them with the high-resolution path through residual connection to form a step-by-step multi-scale compensation path; Step 3.3: Set up a learnable periodic compensation module to alleviate the loss of periodic information caused by multiple downsampling; In this module, m trainable periodic matrices are defined, and a periodic offset vector is dynamically generated based on the current time index at each prediction, which introduces compensation features related to the periodic position through position-cyclic value retrieval, thereby enhancing the model's ability to express periodic patterns; Step 3.4: Ensure that different time steps always correspond to the same phase index by looking up the table, and then learn the bias signal between the true data and the predicted data for each cycle position during the training process using the periodic matrix, and adaptively adjust the local time difference through the offset vector.

4. The multi-scale photovoltaic power generation prediction method of fusion explicit periodic modeling and KAN-MLP hybrid structure according to claim 1 or 2, characterized in that, The process of step 4 is as follows: Step 4.1: In photovoltaic power generation prediction, meteorological variables such as solar radiation, temperature, and humidity are used as input; linear transformation is performed on the features in each channel to obtain the corresponding channel feature representation; Step 4.2: Establish a central channel node and establish an interaction path with each edge channel, so that each channel can obtain global channel context information through dynamic interaction with the central node; Step 4.3: Calculate the attention weight for the features of each channel, and generate a central embedding vector based on the weight; combine the representations corresponding to all channels to form a channel embedding matrix, which is used to represent the global dependency relationship between channels; Step 4.4: Concatenate the central embedding vector with the original features of each channel; Then complete the fusion through multi-layer perception MLP to obtain the output representation of each channel.

5. The multi-scale photovoltaic power generation prediction method of fusing explicit periodic modeling and KAN-MLP hybrid structure according to claim 1 or 2, characterized in that, The process of step 5 is as follows: Step 5.1: Model the input features using the MLP branch and the KAN branch respectively: the MLP network realizes efficient high-dimensional feature combination and abstract expression through a simple multi-layer structure; the KAN network realizes flexible expression of complex segmented nonlinear structure by replacing the fixed activation function with a learnable function basis; Step 5.2: Concatenate the MLP output features and the KAN output features in the feature dimension to obtain a joint feature representation; Step 5.3: Input the joint feature representation into the lightweight attention network; the attention network consists of two fully connected layers and an activation function, which is used to generate attention weights for the two branches; Then, through normalization operation to ensure that the weights are dynamically allocated between the two branches; Step 5.4: According to the attention weight, the MLP output feature and the KAN output feature are weighted and fused to obtain a fusion feature representation; the The fusion feature retains the advantages of both types of representation and realizes feature complementation; Step 5.5: Weighted fusion of the fusion feature and the period term output by the explicit period modeling branch, input into the output head module to generate the final prediction result.

6. The multi-scale photovoltaic power generation prediction method of fusing explicit periodic modeling and KAN-MLP hybrid structure according to claim 1 or 2, characterized in that, The process of step 6 is as follows: Step 6.1: For the overall difference between the predicted value and the true value, the mean square error MSE is used as the basic loss function to measure the global error between the predicted sequence and the actual sequence; Step 6.2: To enhance the modeling ability of the model for periodic and trend features, a trend error perception term is introduced based on the mean square error, which calculates the difference between the predicted sequence and the actual sequence in the frequency domain representation by performing fast Fourier transform FFT, thereby constraining the consistency of the model before and after the trend and cycle components; Step 6.3: Weighted combination of mean square error and trend error perception term to form an improved loss function.

Images