Photovoltaic power prediction method and system based on artificial intelligence

Abstract

The invention discloses a photovoltaic power prediction method and system based on artificial intelligence. The method comprises the steps of photovoltaic data acquisition, preliminary data processing, power prediction model construction, model hyper-parameter optimization and photovoltaic power prediction. The invention relates to the technical field of photovoltaic data processing, in particular to a photovoltaic power prediction method and system based on artificial intelligence, and the method comprises the steps: obtaining original data through photovoltaic data; a primary processing method of missing value processing, spatio-temporal data alignment, data standardization and data set segmentation is adopted; an improved graph convolutional network model is adopted as a power prediction model, nonlinear correlation is effectively captured through adaptive feature extraction and an uncertainty perception mechanism, and meanwhile, the interpretability and the anti-interference capability of the model are enhanced by combining physical rule constraint output; a resonance evolutionary optimization algorithm is adopted to carry out model hyper-parameter optimization, and a dynamic adjustment strategy is combined with frequency domain analysis and hierarchical optimization, so that the stability of the model in a variable environment is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic data processing, and specifically refers to a photovoltaic power prediction method and system based on artificial intelligence. Background Art

[0002] Photovoltaic power prediction refers to predicting the power output of a photovoltaic power generation system by analyzing historical meteorological data, solar radiation, temperature, cloud cover and other factors, using mathematical models and algorithms; it can improve the scheduling efficiency of photovoltaic power generation and the stability of the power grid, reduce the impact of uncertainty on the power system, optimize the allocation of energy resources, ensure the reliability of the power grid operation, help reduce the demand for power storage and reserve capacity, and promote the sustainable development of green energy.

[0003] However, traditional photovoltaic power prediction methods have the technical problems of relying too much on single-time series features or static spatial correlations, lacking the dynamic response ability to multi-dimensional environmental factors such as meteorological mutations and terrain shadows, and at the same time, their fixed topological structures cannot adapt to complex spatial relationships and are difficult to meet the real-time decision-making needs of photovoltaic systems in high-uncertainty environments; traditional photovoltaic power prediction methods have the technical problems of being limited by fixed search steps and static update rules when performing hyperparameter optimization, being difficult to adapt to the non-linear characteristics of high-dimensional parameter spaces, being prone to falling into local optima or computational redundancy, and lacking the ability to perceive the dynamic characteristics of the optimization process and being unable to effectively use historical search information to guide the direction, resulting in slow convergence speed and unstable results. Summary of the Invention

[0004] In view of the above situation, to overcome the defects of the prior art, the present invention provides a photovoltaic power prediction method and system based on artificial intelligence. Aiming at the technical problems that traditional photovoltaic power prediction methods mostly rely on single temporal features or static spatial associations, lack the dynamic response ability to multi-dimensional environmental factors such as meteorological mutations and terrain shadows, and at the same time their fixed topological structures cannot adapt to complex spatial relationships and are difficult to meet the real-time decision-making needs of photovoltaic systems in high-uncertainty environments, this solution creatively uses an improved graph convolutional network model as the power prediction model. Through an adaptive feature extraction and uncertainty perception mechanism, it effectively captures the non-linear associations in complex environments. At the same time, combined with physical law constraints for output, it enhances the interpretability and anti-interference ability of the model, and improves the prediction accuracy and adaptability. Aiming at the technical problems that traditional photovoltaic power prediction methods are limited by fixed search steps and static update rules when performing hyperparameter optimization, are difficult to adapt to the non-linear characteristics of high-dimensional parameter spaces, are prone to falling into local optima or computational redundancy, and lack the perception ability of the dynamic characteristics of the optimization process and cannot effectively use historical search information to guide the direction, resulting in slow convergence speed and unstable results, this solution creatively uses a resonant evolutionary optimization algorithm for model hyperparameter optimization. Through an intelligent optimization mechanism, it realizes the efficient search and global convergence of hyperparameters. Its dynamic adjustment strategy combines frequency domain analysis and hierarchical optimization, and improves the stability of the model in a changing environment.

[0005] The technical solution adopted by the present invention is as follows: The photovoltaic power prediction method based on artificial intelligence provided by the present invention includes the following steps:

[0006] Step S1: Photovoltaic data acquisition;

[0007] Step S2: Preliminary data processing;

[0008] Step S3: Power prediction model construction;

[0009] Step S4: Model hyperparameter optimization;

[0010] Step S5: Photovoltaic power prediction.

[0011] Further, in step S1, the photovoltaic data acquisition is used to obtain the original data required for photovoltaic power prediction. Specifically, data is obtained from the photovoltaic power station monitoring and data acquisition system to obtain a power prediction original data set. The power prediction original data set specifically includes a historical photovoltaic original data set and a to-be-predicted original data set. The historical photovoltaic original data set and the to-be-predicted original data set specifically include meteorological observation data, spatial geographical data, and photovoltaic device data. The historical photovoltaic original data set further includes historical photovoltaic power data.

[0012] Further, in step S2, the preliminary data processing is used to perform preliminary processing on the originally collected power prediction data, and specifically includes the following steps:

[0013] Step S21: Missing value processing, which is used to remove missing values, specifically to remove the missing values in the historical photovoltaic original data set and the to-be-predicted original data set, and obtain a roughly processed historical data set and a roughly processed to-be-predicted data set;

[0014] Step S22: Spatiotemporal data alignment, which is used to align spatiotemporal data. Specifically, the spatial data in the roughly processed historical data set and the roughly processed to-be-predicted data set are converted into a unified coordinate system, and the time data is synchronized in clock, to obtain an aligned historical data set and an aligned to-be-predicted data set;

[0015] Step S23: Data standardization, which is used to standardize the aligned data. Specifically, the minimum-maximum standardization method is used to process the aligned historical data set and the aligned to-be-predicted data set, and a standardized historical data set and a standardized to-be-predicted data set are obtained;

[0016] Step S24: Data set segmentation, which is used to segment the data set. Specifically, the standardized historical data set is segmented into a model training set and a model test set.

[0017] Further, in step S3, the power prediction model construction is used to construct the model required for photovoltaic power prediction. Specifically, an improved graph convolutional network model is constructed as the power prediction model. The improved graph convolutional network model specifically includes a multi-modal feature extraction module, a galvanometer attention module, an improved graph convolutional module, and an output module;

[0018] The power prediction model construction specifically includes the following steps:

[0019] Step S31: Construct a multi-modal feature extraction module, which is used to extract multi-modal features. The steps include:

[0020] Step S311: Temporal modality extraction, which is used to extract joint time-domain and frequency-domain features from temporal data. Specifically, long short-term memory network is used to capture time dependence, and wavelet transform is combined to obtain frequency-domain features for feature entanglement;

[0021] Step S312: Spatial modality extraction, which is used to extract spatial features. Specifically, deformable convolution is used to extract spatial features;

[0022] Step S313: Meteorological modality extraction, which is used to extract multi-scale meteorological features. Specifically, multi-scale meteorological features are fused through pooling and gating mechanisms at different scales;

[0023] Step S314: Multi-modal feature aggregation, which is used to adaptively fuse multi-modal features;

[0024] Step S32: Construct a galvanometer attention module. Specifically, construct the galvanometer attention module based on the galvanometer attention mechanism. The steps include:

[0025] Step S321: Calculate the scanning basis for generating the scanning basis in eight directions.

[0026] Step S322: Obtain the output of the galvanometer attention module. Specifically, obtain the output of the galvanometer attention module based on the galvanometer attention mechanism.

[0027] Step S33: Construct an improved graph convolution module. Specifically, select a photovoltaic power station as a node, construct an uncertainty-aware adjacency matrix, and design a graph convolution network based on the uncertainty-aware adjacency matrix to construct the improved graph convolution module. The steps include:

[0028] Step S331: Generate a similarity measurement factor. Specifically, use a multi-layer perceptron to generate a normal distribution and sample based on the geographical distance and the dynamic time warping distance of meteorological modal features respectively to obtain the similarity measurement factor.

[0029] Step S332: Sample the uncertainty-aware adjacency matrix. Specifically, calculate the probability of connection between each pair of nodes and obtain the uncertainty-aware adjacency matrix through Monte Carlo sampling.

[0030] Step S333: Perform graph convolution. Specifically, perform graph convolution on the output features of the galvanometer attention module based on the uncertainty-aware adjacency matrix to obtain the output features of the improved graph convolution module.

[0031] Step S34: Construct an output module for obtaining the model prediction output. Specifically, use a physical prediction to correct the model output and generate a chaotic signal to modulate the model output through the Lorenz system to obtain the final model prediction result. The steps include:

[0032] Step S341: Obtain the initial output of the model. Specifically, obtain the initial output of the model through a multi-layer perceptron.

[0033] Step S342: Obtain the corrected output of the model. Specifically, calculate the theoretical photovoltaic power and correct the initial output of the model based on the theoretical photovoltaic power to obtain the corrected output of the model.

[0034] Step S343: Obtain the final model prediction result. Specifically, generate a chaotic signal to modulate the model output through the Lorenz system to obtain the final model prediction result.

[0035] Step S35: Construct and train a model. Specifically, construct an improved graph convolutional network model by constructing the multi-modal feature extraction module, the galvanometer attention module, the improved graph convolutional module, and the output module, train the model based on the model training set, verify the model performance based on the model test set, and obtain the improved graph convolutional network model, which is used as the power prediction model.

[0036] Further, in step S4, the model hyperparameter optimization is used to optimize the model hyperparameters of the power prediction model. Specifically, the resonant evolution optimization algorithm is used to optimize the model hyperparameters of the power prediction model to obtain an optimized power prediction model.

[0037] The model hyperparameter optimization specifically includes the following steps:

[0038] Step S41: Algorithm initialization. Specifically, initialize the individual unit set using the tent chaotic map. The individual units in the individual unit set are used to represent the combination of model hyperparameters to be optimized for the power prediction model, and the prediction accuracy of the power prediction model is selected as the fitness value of the individual unit.

[0039] Step S42: Main frequency analysis. Specifically, form a sequence with the fitness values of the individual units, and perform spectral analysis processing on the fitness value sequence using the fast Fourier transform to obtain the main frequency of the fitness.

[0040] Step S43: Adjust the position update step size, which is used to dynamically adjust the search step size.

[0041] Step S44: Update the positions of the individual units. Specifically, during each iteration update, sort the individual units in descending order of fitness value, select the top 20% of the individual units as elite units, and the remaining individual units as non-elite units, and update their positions respectively. The steps include:

[0042] Step S441: Update the positions of the elite units.

[0043] Step S442: Calculate the acceptance probability of the positions of the non-elite units, which is used to calculate the probability of the non-elite units accepting the updated positions.

[0044] Step S443: Update the positions of the non-elite units.

[0045] Step S45: Optimize the model hyperparameters. Specifically, design a resonant evolution optimization algorithm based on the algorithm initialization, the main frequency analysis, the position update step size adjustment, and the individual unit position update, and optimize the model hyperparameters of the power prediction model to obtain an optimized power prediction model.

[0046] Further, in step S5, the photovoltaic power prediction specifically uses the optimized power prediction model to perform photovoltaic power prediction based on the standardized dataset to be predicted, and obtain a reference result of photovoltaic power prediction.

[0047] The photovoltaic power prediction system based on artificial intelligence provided by the present invention includes a photovoltaic data acquisition module, a preliminary data processing module, a power prediction model construction module, a model hyperparameter optimization module, and a photovoltaic power prediction module.

[0048] The photovoltaic data acquisition module is used for acquiring photovoltaic data. By collecting photovoltaic data, an original dataset for power prediction is obtained, and the original dataset for power prediction is sent to the preliminary data processing module.

[0049] The preliminary data processing module is used for preliminary data processing. Through preliminary data processing, a standardized dataset to be predicted, a model training set, and a model test set are obtained. The standardized dataset to be predicted is sent to the photovoltaic power prediction module, and the model training set and the model test set are sent to the power prediction model construction module.

[0050] The power prediction model construction module is used for constructing a power prediction model. By constructing an improved graph convolutional network model, a power prediction model is obtained, and the power prediction model is sent to the model hyperparameter optimization module.

[0051] The model hyperparameter optimization module is used for optimizing model hyperparameters. By using a resonant evolution optimization algorithm to optimize model hyperparameters, an optimized power prediction model is obtained, and the optimized power prediction model is sent to the photovoltaic power prediction module.

[0052] The photovoltaic power prediction module is used for photovoltaic power prediction. By using the optimized power prediction model to perform photovoltaic power prediction, a reference result of photovoltaic power prediction is obtained.

[0053] The beneficial effects achieved by the present invention using the above solution are as follows:

[0054] (1) Aiming at the technical problems that traditional photovoltaic power prediction methods rely too much on single temporal features or static spatial associations, lack the dynamic response ability to multi-dimensional environmental factors such as meteorological mutations and terrain shadows, and their fixed topological structures cannot adapt to complex spatial relationships, making it difficult to meet the real-time decision-making needs of photovoltaic systems in a highly uncertain environment, this solution creatively uses an improved graph convolutional network model as the power prediction model. Through an adaptive feature extraction and uncertainty perception mechanism, it effectively captures non-linear associations in complex environments. At the same time, by combining physical law constraints for output, it enhances the interpretability and anti-interference ability of the model, and improves the prediction accuracy and adaptability.

[0055] (2)Regarding the technical problems that traditional photovoltaic power prediction methods are limited by fixed search step sizes and static update rules when performing hyperparameter optimization, difficult to adapt to the non-linear characteristics of high-dimensional parameter spaces, prone to falling into local optima or computational redundancy, lacking the ability to perceive the dynamic characteristics of the optimization process, and unable to effectively utilize historical search information to guide the direction, resulting in slow convergence speed and unstable results, this solution creatively adopts a resonant evolution optimization algorithm for model hyperparameter optimization, achieving efficient search and global convergence of hyperparameters through an intelligent optimization mechanism. Its dynamic adjustment strategy combines frequency domain analysis and hierarchical optimization, enhancing the stability of the model in a changing environment. Description of the Drawings

[0056] Figure 1 It is a schematic flow chart of the photovoltaic power prediction method based on artificial intelligence provided by the present invention;

[0057] Figure 2 It is a schematic module diagram of the photovoltaic power prediction system based on artificial intelligence provided by the present invention;

[0058] Figure 3 It is a schematic flow chart of the preliminary data processing in step S2;

[0059] Figure 4 It is a schematic flow chart of the power prediction model construction in step S3;

[0060] Figure 5 It is a schematic flow chart of the model hyperparameter optimization in step S4.

[0061] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. Detailed Embodiments

[0062] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0063] In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc. indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation to the present invention.

[0064] Example 1, refer to Figure 1 , the technical solution adopted by the present invention is as follows: The photovoltaic power prediction method based on artificial intelligence provided by the present invention includes the following steps:

[0065] Step S1: Photovoltaic data acquisition;

[0066] Step S2: Preliminary data processing;

[0067] Step S3: Construction of power prediction model;

[0068] Step S4: Optimization of model hyperparameters;

[0069] Step S5: Photovoltaic power prediction.

[0070] Example 2, refer to Figure 1 and Figure 2 , in step S1, the photovoltaic data acquisition is used to obtain the original data required for photovoltaic power prediction. Specifically, data is obtained from the photovoltaic power station monitoring and data acquisition system to obtain the original power prediction dataset. The original power prediction dataset specifically includes the historical photovoltaic original dataset and the original dataset to be predicted. The historical photovoltaic original dataset and the original dataset to be predicted specifically include meteorological observation data, spatial geographical data, and photovoltaic equipment data. The historical photovoltaic original dataset also includes historical photovoltaic power data. The meteorological observation data specifically includes solar irradiance data, ambient temperature and humidity data, wind speed data, wind direction data, and cloud amount data. The spatial geographical data specifically includes the longitude and latitude coordinates of the photovoltaic power station, the altitude of the photovoltaic power station, terrain undulation data, and surrounding shadow distribution data. The photovoltaic equipment data specifically includes photovoltaic equipment parameter data, photovoltaic module tilt angle data, and temperature data when the photovoltaic module is working.

[0071] Example 3, refer to Figure 1 , Figure 2 and Figure 3 , based on the above example, in step S2, the preliminary data processing is used to perform preliminary processing on the collected original power prediction data, and specifically includes the following steps:

[0072] Step S21: Missing value processing, which is used to remove missing values. Specifically, the missing values in the historical photovoltaic original dataset and the original dataset to be predicted are removed to obtain the roughly processed historical dataset and the roughly processed dataset to be predicted;

[0073] Step S22: Spatiotemporal data alignment, which is used to align spatiotemporal data. Specifically, the spatial data in the roughly processed historical dataset and the roughly processed dataset to be predicted is converted into a unified coordinate system, and the time data is clock synchronized to obtain the aligned historical dataset and the aligned dataset to be predicted;

[0074] Step S23: Data standardization, which is used to standardize the aligned data. Specifically, the min-max standardization method is adopted to process the aligned historical data set and the aligned data set to be predicted, so as to obtain a standardized historical data set and a standardized data set to be predicted.

[0075] Step S24: Data set segmentation, which is used to segment the data set. Specifically, the standardized historical data set is segmented into a model training set and a model test set.

[0076] Example 4, refer to Figure 1 、 Figure 2 and Figure 4 Based on the above example, in step S3, the power prediction model construction is used to construct the model required for photovoltaic power prediction. Specifically, an improved graph convolutional network model is constructed as the power prediction model. The improved graph convolutional network model specifically includes a multi-modal feature extraction module, a galvanometer attention module, an improved graph convolutional module, and an output module.

[0077] The power prediction model construction specifically includes the following steps:

[0078] Step S31: Construct a multi-modal feature extraction module, which is used to extract multi-modal features. The steps include:

[0079] Step S311: Temporal modality extraction, which is used to extract the joint features of the time domain and the frequency domain from the temporal data. Specifically, the long short-term memory network is used to capture the time dependence, and the wavelet transform is combined to obtain the frequency domain features, and feature entanglement is performed. The formula used is as follows:

[0080] ;

[0081] In the formula, represents the temporal domain feature, represents the frequency domain feature, represents the input temporal data, represents the long short-term memory network function, represents the discrete wavelet transform function, represents the temporal modality feature, represents the ReLU activation function, represents the real part taking function, represents the tensor product;

[0082] Step S312: Spatial modality extraction, which is used to extract spatial features. Specifically, deformable convolution is adopted to extract spatial features. The formula used is as follows:

[0083] ;

[0084] In the formula, Represents the offset of the deformable convolution kernel, Represents the hyperbolic tangent function, Represents the convolution operation function, Represents the input spatial feature, Represents the spatial modal feature, Represents the deformable convolution function;

[0085] Step S313: Meteorological modal extraction, which is used to extract multi-scale meteorological features. Specifically, multi-scale meteorological features are fused through pooling and gating mechanisms at different scales. The formula used is as follows:

[0086] ;

[0087] In the formula, Represents the average pooling result of the meteorological feature scale s, Represents the average pooling function of scale s, Represents the input meteorological feature, s represents the scale index, Represents the meteorological gating weight of scale s, Represents the sigmoid function, Represents the multi-layer perceptron function used to generate the meteorological gating weight, Represents the meteorological modal feature, Represents element-wise multiplication;

[0088] Step S314: Multi-modal feature aggregation, which is used to adaptively fuse multi-modal features. The formula used is as follows:

[0089] ;

[0090] In the formula, Represents the modal attention, a, b, and c represent the modal feature indices, Represents the LeakyReLU activation function, Represents the learnable vector for calculating the modal attention, T represents the transpose operation, Represents the modal aggregation feature;

[0091] Step S32: Construct the galvanometer attention module. Specifically, the galvanometer attention module is constructed based on the galvanometer attention mechanism. The steps include:

[0092] Step S321: Calculate the scanning base, which is used to generate the scanning bases in eight directions. The formula used is as follows:

[0093] ;

[0094] In the formula, Represents the direction angle of the d-th scanning base, Denote the d-th scanning substrate;

[0095] Step S322: Obtain the output of the galvanometer attention module, specifically, obtain the output of the galvanometer attention module based on the galvanometer attention mechanism, and the formula used is as follows:

[0096] ;

[0097] In the formula, denotes the galvanometer attention score, denotes the d-th learnable mapping matrix, denotes the e-th learnable mapping matrix, denotes the output feature of the galvanometer attention module;

[0098] Step S33: Construct an improved graph convolutional module, specifically, select a photovoltaic power station as a node, construct an uncertainty-aware adjacency matrix, and design a graph convolutional network based on the uncertainty-aware adjacency matrix to construct the improved graph convolutional module. The steps include:

[0099] Step S331: Generate a similarity measurement factor, specifically, use a multi-layer perceptron to generate a normal distribution and sample based on the geographical distance and the dynamic time warping distance of meteorological modal features respectively to obtain the similarity measurement factor. The formula used is as follows:

[0100] ;

[0101] In the formula, denotes the mean of the similarity measurement factor based on the geographical distance, denotes the multi-layer perceptron function for generating the mean of the similarity measurement factor based on the geographical distance, denotes the geographical distance between node p and node q, denotes the mean of the similarity measurement factor based on the dynamic time warping distance of meteorological modal features, denotes the multi-layer perceptron function for generating the mean of the similarity measurement factor based on the dynamic time warping distance of meteorological modal features, denotes the dynamic time warping distance calculation function, denotes the meteorological modal feature of node p, denotes the meteorological modal feature of node q, Ds denotes the similarity measurement factor based on the geographical distance, Ws denotes the similarity measurement factor based on the dynamic time warping distance of meteorological modal features, denotes the learnable standard deviation of the similarity measurement factor based on the geographical distance, denotes the learnable standard deviation of the similarity measurement factor based on the dynamic time warping distance of meteorological modal features;

[0102] Step S332: Uncertainty-aware adjacency matrix sampling, specifically calculating the probability of connection between each pair of nodes and obtaining the uncertainty-aware adjacency matrix through Monte Carlo sampling. The formula used is as follows:

[0103] ;

[0104] In the formula, represents the uncertainty-aware adjacency matrix, represents the element of the uncertainty-aware adjacency matrix. Its value of 1 indicates the existence of a connection between node p and node q, and its value of 0 indicates the non-existence of a connection between node p and node q. represents the probability of the existence of a connection between node p and node q, represents the learnable weight of the geographical distance, represents the learnable weight of the dynamic time warping distance of the meteorological modal feature, represents the uncertainty-aware adjacency matrix of the v-th sampling, represents the Bernoulli distribution, and V represents the total number of samplings;

[0105] Step S333: Perform graph convolution, specifically performing graph convolution on the output features of the galvanometer attention module based on the uncertainty-aware adjacency matrix to obtain the output features of the improved graph convolution module. The formula used is as follows:

[0106] ;

[0107] In the formula, represents the output features of the improved graph convolution module, represents the graph convolution function;

[0108] Step S34: Construct an output module for obtaining the model prediction output. Specifically, the model output is corrected by the physical prediction, and the model output is modulated by the chaotic signal generated by the Lorenz system to obtain the final model prediction result. The steps include:

[0109] Step S341: Obtain the initial model output, specifically obtaining the initial model output through a multi-layer perceptron. The formula used is as follows:

[0110] ;

[0111] In the formula, represents the initial model output, represents the multi-layer perceptron function used to generate the initial model output;

[0112] Step S342: Obtain the corrected model output, specifically calculating the theoretical photovoltaic power and correcting the initial model output based on the theoretical photovoltaic power to obtain the corrected model output. The formula used is as follows:

[0113] ;

[0114] Wherein, represents the theoretical photovoltaic power at time t, represents the photovoltaic conversion efficiency, PA represents the surface area of the photovoltaic module, represents the solar irradiance at time t, represents the temperature attenuation coefficient, represents the temperature of the photovoltaic module during operation at time t, represents the output of the model correction at time t, w represents the learnable compensation coefficient, represents the moving average of the theoretical photovoltaic power;

[0115] Step S343: Obtain the final model prediction result, specifically by generating a chaotic signal to modulate the model output through the Lorenz system to obtain the final model prediction result. The formula used is as follows:

[0116] ;

[0117] Wherein, represents the chaotic state, represents the first state variable, represents the second state variable, represents the third state variable, represents the Prandtl number, represents the normalized Rayleigh number, represents the geometric constraint coefficient, represents the final model prediction result at time t;

[0118] Step S35: Construct and train the model, specifically by constructing the multi-modal feature extraction module, the galvanometer attention module, the improved graph convolutional module, and the output module to construct an improved graph convolutional network model, training the model based on the model training set, and verifying the model performance based on the model test set to obtain an improved graph convolutional network model, which is used as the power prediction model.

[0119] By performing the above operations, aiming at the technical problems of the traditional photovoltaic power prediction method that has multiple dependencies on single-time series features or static spatial associations, lacks the dynamic response ability to multi-dimensional environmental factors such as meteorological mutations and terrain shadows, and its fixed topological structure cannot adapt to complex spatial relationships and is difficult to meet the real-time decision-making needs of photovoltaic systems in high-uncertainty environments, this solution creatively uses an improved graph convolutional network model as the power prediction model. Through the adaptive feature extraction and uncertainty perception mechanism, it effectively captures the non-linear associations in complex environments, and at the same time combines physical law constraints to output, enhancing the interpretability and anti-interference ability of the model, and improving the prediction accuracy and adaptability.

[0120] Example 5, refer to Figure 1 、 Figure 2 and Figure 5 , based on the above example, in step S4, the model hyperparameter optimization is used to optimize the model hyperparameters of the power prediction model. Specifically, the harmonic evolution optimization algorithm is used to optimize the model hyperparameters of the power prediction model to obtain an optimized power prediction model;

[0121] The model hyperparameter optimization specifically includes the following steps:

[0122] Step S41: Algorithm initialization. Specifically, the tent chaos mapping is used to initialize the individual unit set. The individual units in the individual unit set are used to represent the combination of model hyperparameters to be optimized for the power prediction model. The prediction accuracy of the power prediction model is selected as the fitness value of the individual unit;

[0123] Step S42: Main frequency analysis. Specifically, the fitness values of the individual units are formed into a sequence, and the fast Fourier transform is used to perform spectral analysis on the fitness value sequence to obtain the main frequency of the fitness;

[0124] Step S43: Position update step size adjustment, which is used to dynamically adjust the search step size. The formula used is as follows:

[0125] ;

[0126] In the formula, represents the time-varying gain coefficient at the dt-th iteration, dt represents the current iteration number, represents the maximum iteration number, represents the update step size of the i-th individual unit at the (dt + 1)-th iteration, represents the reference update step size of the i-th individual unit;

[0127] Step S44: Individual unit position update. Specifically, during each iteration update, the individual units are sorted in descending order of fitness value. The top 20% of the individual units are selected as elite units, and the remaining individual units are used as non-elite units, and their positions are updated respectively. The steps include:

[0128] Step S441: Elite unit position update. The formula used is as follows:

[0129] ;

[0130] In the formula, represents the position of the j-th elite unit at the (dt + 1)-th iteration, represents the historical optimal position of the j-th elite unit, represents the mutation intensity attenuation coefficient, represents the standard multivariate normal distribution, which is used to introduce Gaussian perturbations, and I represents the identity matrix;

[0131] Step S442: Calculation of the acceptance probability of the non-elite unit position, which is used to calculate the probability that the non-elite unit accepts the updated position. The formula used is as follows:

[0132] ;

[0133] In the formula, represents the time-varying temperature attenuation parameter, represents the initial time-varying temperature attenuation parameter, represents the acceptance probability of the position of the k-th non-elite unit at the (dt + 1)-th iteration, represents the fitness value of the k-th non-elite unit at the dt-th iteration, represents the fitness value of the k-th non-elite unit at the (dt + 1)-th iteration;

[0134] Step S443: Update of the non-elite unit position. The formula used is as follows:

[0135] ;

[0136] In the formula, represents the direction angle of the k-th non-elite unit at the (dt + 1)-th iteration, represents the direction angle of the k-th non-elite unit at the dt-th iteration, represents the phase update rate, represents the main frequency of fitness, represents the position of the k-th non-elite unit at the (dt + 1)-th iteration, represents the position of the k-th non-elite unit at the dt-th iteration, and Cy represents the Cauchy perturbation term;

[0137] Step S45: Optimize the hyperparameters of the model. Specifically, based on the algorithm initialization, the main frequency analysis, the adjustment of the position update step size, and the update of the individual unit position, design a resonant evolution optimization algorithm, and optimize the hyperparameters of the power prediction model to obtain an optimized power prediction model.

[0138] By performing the above operations, for the traditional photovoltaic power prediction method, when performing hyperparameter optimization, it is limited by the fixed search step size and static update rules, and it is difficult to adapt to the non-linear characteristics of the high-dimensional parameter space. It is easy to fall into local optima or computational redundancy, and it lacks the ability to perceive the dynamic characteristics of the optimization process and cannot effectively use historical search information to guide the direction, resulting in slow convergence speed and unstable results. The present solution creatively uses the resonant evolution optimization algorithm to optimize the model hyperparameters, and realizes the efficient search and global convergence of hyperparameters through an intelligent optimization mechanism. Its dynamic adjustment strategy combines frequency domain analysis and hierarchical optimization, which improves the stability of the model in a changing environment.

[0139] Example Six. Refer to Figure 1 and Figure 2 . Based on the above example, in step S5, the photovoltaic power prediction specifically uses the optimized power prediction model to perform photovoltaic power prediction based on the standardized dataset to be predicted, and obtains a reference result of photovoltaic power prediction.

[0140] Example Seven. Refer to Figure 1 and Figure 2 . Based on the above example, the photovoltaic power prediction system based on artificial intelligence provided by the present invention includes a photovoltaic data acquisition module, a preliminary data processing module, a power prediction model construction module, a model hyperparameter optimization module, and a photovoltaic power prediction module;

[0141] The photovoltaic data acquisition module is used for photovoltaic data acquisition. By collecting photovoltaic data, a raw dataset for power prediction is obtained, and the raw dataset for power prediction is sent to the preliminary data processing module;

[0142] The preliminary data processing module is used for preliminary data processing. Through preliminary data processing, a standardized dataset to be predicted, a model training set, and a model test set are obtained, and the standardized dataset to be predicted is sent to the photovoltaic power prediction module, and the model training set and the model test set are sent to the power prediction model construction module;

[0143] The power prediction model construction module is used for power prediction model construction. By constructing an improved graph convolutional network model, a power prediction model is obtained, and the power prediction model is sent to the model hyperparameter optimization module;

[0144] The model hyperparameter optimization module is used for model hyperparameter optimization. By using the resonant evolution optimization algorithm to optimize the model hyperparameters, an optimized power prediction model is obtained, and the optimized power prediction model is sent to the photovoltaic power prediction module;

[0145] The photovoltaic power prediction module is used for photovoltaic power prediction. By using the optimized power prediction model for photovoltaic power prediction, a reference result of photovoltaic power prediction is obtained.

[0146] It should be noted that in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0147] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention.

[0148] The above description of the present invention and its embodiments is not restrictive. What is shown in the drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. In summary, if those of ordinary skill in the art are inspired by it and, without departing from the gist of the present invention, design similar structural modes and embodiments to this technical solution without creative efforts, they shall fall within the protection scope of the present invention.

Claims

1. A photovoltaic power prediction method based on artificial intelligence, characterized in that: The method includes the following steps: Step S1: Photovoltaic data acquisition. By obtaining data from the photovoltaic power station monitoring and data acquisition system, an original power prediction data set is obtained. The original power prediction data set specifically includes a historical photovoltaic original data set and a to-be-predicted original data set. Step S2: Preliminary data processing, which is used to preliminarily process the collected original power prediction data to obtain a standardized to-be-predicted data set, a model training set, and a model test set. Step S3: Power prediction model construction, which is used to construct the model required for photovoltaic power prediction. Specifically, an improved graph convolutional network model is constructed as the power prediction model. The improved graph convolutional network model specifically includes a multi-modal feature extraction module, a galvanometer attention module, an improved graph convolutional module, and an output module. The improved graph convolutional module specifically selects a photovoltaic power station as a node, constructs an uncertainty-aware adjacency matrix, and designs a graph convolutional network based on the uncertainty-aware adjacency matrix to construct the improved graph convolutional module. Step S4: Model hyperparameter optimization, which is used to optimize the model hyperparameters of the power prediction model. Specifically, a harmonic evolution optimization algorithm is used to optimize the model hyperparameters of the power prediction model to obtain an optimized power prediction model. Step S5: Photovoltaic power prediction. Specifically, the optimized power prediction model is used to perform photovoltaic power prediction based on the standardized to-be-predicted data set to obtain a reference result for photovoltaic power prediction.

2. The photovoltaic power prediction method based on artificial intelligence according to claim 1, wherein: The power prediction model construction specifically includes the following steps: Step S31: Construct a multi-modal feature extraction module, which is used to extract multi-modal features. The steps include: Step S311: Temporal modality extraction, which is used to extract joint time-domain and frequency-domain features from temporal data. Specifically, a long short-term memory network is used to capture time dependencies, and wavelet transform is combined to obtain frequency-domain features for feature entanglement. Step S312: Spatial modality extraction, which is used to extract spatial features. Specifically, deformable convolution is used to extract spatial features. Step S313: Meteorological modality extraction, which is used to extract multi-scale meteorological features. Specifically, multi-scale meteorological features are fused through pooling and gating mechanisms at different scales. Step S314: Multi-modal feature aggregation, which is used to adaptively fuse multi-modal features. Step S32: Construct a galvanometer attention module. Specifically, the galvanometer attention module is constructed based on the galvanometer attention mechanism. The steps include: Step S321: Calculate the scanning base, which is used to generate scanning bases in eight directions. Step S322: Obtain the output of the galvanometer attention module. Specifically, the output of the galvanometer attention module is obtained based on the galvanometer attention mechanism. Step S33: Construct an improved graph convolutional module. The steps include: Step S331: Generate a similarity measurement factor. Specifically, a multi-layer perceptron is used to generate a normal distribution and sample based on the geographical distance and the dynamic time warping distance of the meteorological modality features respectively to obtain a similarity measurement factor. Step S332: Uncertainty-aware adjacency matrix sampling. Specifically, the probability of connection between each pair of nodes is calculated, and the uncertainty-aware adjacency matrix is obtained through Monte Carlo sampling. Step S333: Perform graph convolution, specifically, perform graph convolution on the output features of the galvanometer attention module based on the uncertainty-aware adjacency matrix to obtain the output features of the improved graph convolution module; Step S34: Construct an output module for obtaining the model prediction output. Specifically, correct the model output using physical prediction and modulate the model output with a chaotic signal generated by the Lorenz system to obtain the final model prediction result. The steps include: Step S341: Obtain the initial model output, specifically, obtain the initial model output through a multi-layer perceptron; Step S342: Obtain the corrected model output, specifically, calculate the theoretical photovoltaic power and correct the initial model output based on the theoretical photovoltaic power to obtain the corrected model output; Step S343: Obtain the final model prediction result, specifically, modulate the model output with a chaotic signal generated by the Lorenz system to obtain the final model prediction result; Step S35: Construct and train the model. Specifically, construct an improved graph convolution network model through the constructed multi-modal feature extraction module, the constructed galvanometer attention module, the constructed improved graph convolution module, and the constructed output module, train the model based on the model training set, and verify the model performance based on the model test set to obtain an improved graph convolution network model, which is used as the power prediction model.

3. The photovoltaic power prediction method based on artificial intelligence according to claim 1, wherein: The optimization of the model hyperparameters specifically includes the following steps: Step S41: Algorithm initialization. Specifically, initialize the individual unit set using the tent chaotic map. The individual units in the individual unit set are used to represent the combination of model hyperparameters to be optimized for the power prediction model. Select the prediction accuracy of the power prediction model as the fitness value of the individual unit; Step S42: Main frequency analysis. Specifically, form a sequence of the fitness values of the individual units and perform spectral analysis processing on the fitness value sequence using the fast Fourier transform to obtain the main frequency of the fitness; Step S43: Adjust the position update step size for dynamically adjusting the search step size; Step S44: Update the positions of the individual units. Specifically, in each iteration update, sort the individual units in descending order of the fitness value, select the top 20% of the individual units as elite units, and the remaining individual units as non-elite units, and update their positions respectively. The steps include: Step S441: Update the positions of the elite units; Step S442: Calculate the acceptance probability of the non-elite unit positions for calculating the probability of the non-elite units accepting the updated positions; Step S443: Update the positions of the non-elite units; Step S45: Optimize the model hyperparameters. Specifically, design a resonant evolutionary optimization algorithm based on the algorithm initialization, the main frequency analysis, the position update step size adjustment, and the individual unit position update, and optimize the model hyperparameters of the power prediction model to obtain an optimized power prediction model.

4. The photovoltaic power prediction method based on artificial intelligence according to claim 1, wherein: The historical photovoltaic original dataset and the original dataset to be predicted specifically include meteorological observation data, spatial geographical data, and photovoltaic device data. The historical photovoltaic original dataset also includes historical photovoltaic power data.

5. The photovoltaic power prediction method based on artificial intelligence according to claim 1, characterized in that: The preliminary data processing specifically includes the following steps: Step S21: Missing value processing, which is used to remove missing values. Specifically, it removes the missing values in the historical photovoltaic original dataset and the to-be-predicted original dataset to obtain a roughly processed historical dataset and a roughly processed to-be-predicted dataset. Step S22: Spatiotemporal data alignment, which is used to align spatiotemporal data. Specifically, it converts the spatial data in the roughly processed historical dataset and the roughly processed to-be-predicted dataset into a unified coordinate system and synchronizes the time data to obtain an aligned historical dataset and an aligned to-be-predicted dataset. Step S23: Data standardization, which is used to standardize the aligned data. Specifically, it processes the aligned historical dataset and the aligned to-be-predicted dataset using the min-max standardization method to obtain a standardized historical dataset and a standardized to-be-predicted dataset. Step S24: Dataset segmentation, which is used to segment the dataset. Specifically, it segments the standardized historical dataset into a model training set and a model test set.

6. An artificial intelligence-based photovoltaic power prediction system for implementing the artificial intelligence-based photovoltaic power prediction method according to any one of claims 1-5, characterized in that: It includes a photovoltaic data acquisition module, a preliminary data processing module, a power prediction model construction module, a model hyperparameter optimization module, and a photovoltaic power prediction module.

7. The photovoltaic power prediction system based on artificial intelligence according to claim 6, characterized in that: The photovoltaic data acquisition module is used for photovoltaic data acquisition. By collecting photovoltaic data, it obtains a power prediction original dataset and sends the power prediction original dataset to the preliminary data processing module. The preliminary data processing module is used for preliminary data processing. By preliminary data processing, it obtains a standardized to-be-predicted dataset, a model training set, and a model test set, and sends the standardized to-be-predicted dataset to the photovoltaic power prediction module, and sends the model training set and the model test set to the power prediction model construction module. The power prediction model construction module is used for power prediction model construction. By constructing an improved graph convolutional network model, it obtains a power prediction model and sends the power prediction model to the model hyperparameter optimization module. The model hyperparameter optimization module is used for model hyperparameter optimization. By using the resonant evolution optimization algorithm for model hyperparameter optimization, it obtains an optimized power prediction model and sends the optimized power prediction model to the photovoltaic power prediction module. The photovoltaic power prediction module is used for photovoltaic power prediction. By using the optimized power prediction model for photovoltaic power prediction, it obtains a photovoltaic power prediction reference result.

Images