Photovoltaic generating capacity prediction method based on AI

Abstract

The invention discloses an AI-based photovoltaic generating capacity prediction method. The method comprises the following steps: S1, acquiring meteorological and photovoltaic system data of different regions; s2, inputting the processed meteorological and photovoltaic system data of different regions into an initial prediction model based on an MLP (Multilayer Perceptron Neural Network); and S3, establishing constraint conditions defined by combining the data with single-variable and cross-variable rules, and carrying out multi-dimensional judgment on the data rationality. And S4, sorting data, reports and verification results to form a final photovoltaic power generation capacity prediction result report. The method mainly solves key problems existing in an existing photovoltaic generating capacity prediction method, achieves remarkable improvement of prediction precision and reliability, and has important practical significance and wide application prospects. The method not only improves the prediction precision and reduces the prediction error rate, but also enhances the generalization ability of the model, so that the model can better adapt to different environment and climate changes.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and photovoltaic power generation technology, and in particular to an AI-based method for predicting photovoltaic power generation. Background Technology Solar energy, as a clean energy source, is playing an increasingly important role in energy supply. Photovoltaic power generation is widely used due to its environmental friendliness and renewability; however, its fluctuating and intermittent nature, influenced by meteorological factors, poses challenges to power system dispatch management, stability control, and photovoltaic site selection. While existing forecasting methods have made some progress, many problems remain to be solved. Photovoltaic power generation forecasting methods can be mainly divided into physical methods, statistical methods, and artificial intelligence algorithms. Physical methods, based on the physical characteristics of photovoltaic cell modules, predict power through mathematical formulas (such as calculating theoretical output using temperature and irradiance). This method is intuitive but sensitive to meteorological forecast errors. Statistical methods include time series models (such as ARIMA and grey models) and probabilistic predictions (such as Bayesian methods and support vector regression), utilizing historical data trends and seasonal information for prediction, but they struggle to handle complex nonlinear relationships and large-scale data. Artificial intelligence algorithms, such as traditional machine learning methods like Support Vector Machine (SVM), k-Nearest Neighbors (KNN), and Random Forest (RF), as well as neural network models like Backpropagation Neural Network (BP), LSTM, GRU, CNN-LSTM, and Transformer, make predictions by constructing a mapping function between inputs (such as temperature, irradiance, and humidity) and output power. They have strong nonlinear modeling capabilities and automatic feature extraction capabilities. However, a single model often struggles to capture complex nonlinear relationships and is prone to overfitting or underfitting, resulting in insufficient generalization ability of the model.

[0002] Currently, many studies are dedicated to improving and optimizing photovoltaic (PV) power generation forecasting methods. For example, a distributed PV short-term power forecasting method based on weather-change adaptive classification and matching, through steps such as feature selection and dimensionality reduction, power curve smoothing, historical power data clustering, and weather process matching, combined with an improved multiverse algorithm-optimized convolutional neural network (IMVO-CNN) model, effectively improves prediction accuracy, especially demonstrating good adaptability under different weather events. Furthermore, a PV power forecasting method based on an XGBoost-LSTM combined model, by analyzing the influencing factors of PV power generation, constructing relevant features, and combining XGBoost and LSTM models for prediction, has also achieved good forecasting results. However, existing methods still have shortcomings in handling the uncertainty of meteorological factors, data preprocessing, feature extraction, and model optimization. Developing an efficient, accurate, and reliable PV power generation forecasting method is of great significance for promoting the large-scale application of PV power generation, optimizing grid operation, and improving energy economic efficiency.

[0003] This invention proposes an intelligent prediction method for photovoltaic power generation. Through the deep integration of physical rule constraints and multimodal data-driven modeling, it achieves multi-task collaborative prediction. Based on multi-source data of meteorological and photovoltaic system parameters, the method employs statistical verification techniques and constructs a high-quality training set (1080 sets) containing 14 dynamic parameters such as irradiance, ambient temperature, and module tilt angle through joint statistical distribution verification. On this basis, a multi-task hierarchical attention network architecture (MLP) is designed, introducing a dynamic feature selection mechanism and hierarchical attention modules. A cross-task gradient coupling algorithm synchronously decouples the prediction tasks of four indicators: effective output energy of the array, grid-connected power, and daily power generation of the array and system, significantly improving the model's generalization ability to complex operating conditions. Validated on 100 independent test sets, the prediction error rate is reduced to 1.17%~2.03%, and the goodness of fit (R²) of key indicators is 98.4%~99.7%, overcoming the overfitting and task conflict bottlenecks of traditional single-task models. This provides highly reliable decision support for the refined operation of photovoltaic power plants and dynamic grid scheduling. Summary of the Invention To address the problems mentioned above, this invention proposes an AI-based photovoltaic power generation prediction method, aiming to improve the accuracy and reliability of photovoltaic power generation prediction.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] An AI-based method for predicting photovoltaic power generation includes the following steps:

[0006] S1: Acquire meteorological and photovoltaic system data from different regions and construct a dataset containing 1080 samples. Statistically validate the data, clean and remove outliers, and standardize the data to process the meteorological and photovoltaic system data from different regions. Output 14-dimensional variable parameters, covering meteorological variables (total horizontal radiation, ambient temperature) and photovoltaic system variables (PV module power, number of modules, area of ​​a single module, power of the module array, inverter power, number of inverters, system efficiency, array loss, system loss, tilt angle of the light-receiving surface, azimuth angle, and usable roof area) from six cities across the country, including Beijing, Guangzhou, Shanghai, Changchun, Chongqing, and Zhengzhou.

[0007] S2: Processed meteorological and photovoltaic system data from different regions are input into the initial prediction model based on a multilayer perceptron neural network (MLP). Features are fused to integrate different features, weights are dynamically adjusted, and the loss function is optimized. Specifically, when constructing the initial prediction model, the number of neurons in the input layer corresponds to 14 dynamic parameters, the hidden layer uses the hyperbolic tangent activation function, and the output layer uses the identity activation function and the sum of squares error function. Next, collected meteorological and photovoltaic system data from different regions are input into the model. Feature fusion technology is used to integrate different features, and the weights are dynamically adjusted to ensure the model focuses on features important to the prediction target. Finally, the mean squared error or mean absolute error is defined as the loss function, and optimization algorithms such as stochastic gradient descent are used to optimize the loss function, adjusting model parameters to reduce the loss function value and thus improve the model's prediction accuracy. A prediction data report is also output.

[0008] S3: Establish constraints based on univariate and intervariate rule definitions for the data, and make multi-dimensional judgments on the rationality of the data. Prepare 100 new data samples as independent test sets to verify the model performance, output prediction results and performance indicators, including RMAE, MAE, and R², to evaluate the model's generalization ability and prediction accuracy.

[0009] S4: Compile data, reports, and verification results to form the final photovoltaic power generation forecast report.

[0010] Preferably, step S1 specifically includes:

[0011] S11: Conduct multi-source data collection to obtain meteorological data and photovoltaic system data from different regions, and construct a dataset containing 1080 sets of samples. Each set of samples contains meteorological data and photovoltaic system data from different regions within a day.

[0012] S12: Perform statistical verification and data cleaning on the samples in the dataset. Use statistical methods to fully verify the sample data and use SPSS statistical software to identify and remove obvious erroneous data and outliers, such as removing temperature or power values ​​that are outside the physical range.

[0013] S13: Perform data standardization. Standardize the cleaned sample data to transform the sample data from different dimensions to the same scale. Use the Z-Score standardization method to make the mean of the data 0 and the standard deviation 1. The formula is as follows: Where x is a value in the sample data. The mean, The z-value represents the distance between a given value in the sample data and the mean, expressed in standard deviations.

[0014] S14: Output dynamic parameters. Extract and output 14-dimensional dynamic parameters, including key indicators such as irradiance, ambient temperature, component power, and system efficiency. These parameters can comprehensively characterize the operating status and external environmental conditions of the photovoltaic power generation system.

[0015] Preferably, in step S2, the processed data is input into the initial prediction model based on the MLP (Multilayer Perceptron Neural Network). Operations such as feature fusion integrating different features, dynamic adjustment of weight allocation, and loss function optimization are performed, while a prediction data report is output.

[0016] S21: Build an initial prediction model based on an MLP (Multilayer Perceptron) neural network, design the network structure, including an input layer, multiple hidden layers, and an output layer, and determine the number of neurons and activation functions for each layer.

[0017] S22: During model training, feature fusion technology is used to integrate different features, explore the correlation and complementary information between features, and dynamically adjust the weight allocation so that the model pays more attention to features that have an important impact on the prediction target, thereby improving the model's sensitivity to key information and learning ability.

[0018] S23: Define a suitable loss function, including sum of squared errors (SSE) or mean total relative error (MAPE), to measure the difference between the model's predicted values ​​and the actual values. Then, use an optimization algorithm, namely stochastic gradient descent, to optimize the loss function, continuously adjusting the model's parameters to gradually reduce the loss function value and improve the model's prediction accuracy.

[0019] Preferably, the specific content of building the initial prediction model in step S21 is as follows:

[0020] S211: In the input layer, the number of neurons in the input layer corresponds to 14-dimensional dynamic parameters. The input data is represented as a vector x=[x1,x2,…xi,…,x14], where each xi represents a dynamic parameter.

[0021] S212: In the hidden layer, the hyperbolic tangent activation function is used, and the hidden layer has... There are neurons, and the weight matrix is... The bias vector is The output h of the hidden layer is calculated as follows: , where tanh(⋅) is the hyperbolic tangent activation function, applied element-wise to its input vector.

[0022] S213: In the output layer, the output layer selects the identity activation function and the sum of squares error function according to the requirements of the prediction task. The output layer has There are 10 neurons (determined based on the specific photovoltaic power generation prediction task), and the weight matrix is ​​as follows: The bias vector is ,in, Represents the set of real numbers. Output of the output layer. The calculation is as follows: The identity activation function is used, meaning the activation function of the output layer is a linear function. The sum of squares error function E is calculated as follows: in, These are model predictions. `i` is the actual value, and `i` is the counting index variable used to iterate through the output layer neurons. The sum of squares error function is used to measure the difference between the model's predicted value and the actual value.

[0023] Preferably, the specific content of step S22 is as follows:

[0024] S221: Introduces an attention mechanism to dynamically adjust feature weights, focusing on highly correlated parameters (such as irradiance and component temperature). The weight allocation formula is: ,in, For query vector, Let be the i-th dimensional feature vector; T represents the transpose, used to calculate the similarity between the query vector and the feature vector; j is an index variable used to iterate through the dimensions of all feature vectors, ranging from 1 to 14.

[0025] Preferably, step S3 specifically includes:

[0026] S31: Through model validation in SPSS statistical software, and combining univariate and intervariate rule definitions, a multi-dimensional judgment of the reasonableness of the predicted data is made. Univariate rule definition refers to setting reasonable value ranges and constraints for each predicted variable, such as the predicted power generation should not be negative; intervariate rule definition refers to considering the interrelationships and constraints between different variables, such as the relationship between power generation and irradiance should conform to certain physical relationships. If the predicted results violate these rules, the data is judged to be unreasonable.

[0027] S32: Prepare 100 new data samples as an independent test set. These new data samples were not used in the model training process, allowing for objective verification of the model's performance. Input the independent test set into the trained prediction model to obtain prediction results. Calculate performance metrics between the predicted results and the actual values, including RMAE, MAE, and R², to evaluate the model's generalization ability and prediction accuracy.

[0028] Preferably, the specific content of the data reasonableness judgment in step S31 is as follows:

[0029] S311: Univariate rules involve setting reasonable ranges of values ​​and constraints for each predictor variable.

[0030] Number of photovoltaic modules ≥1;

[0031] Number of inverters ≥1;

[0032] Inclination of light-transmitting surface ∈[0°,90°];

[0033] Azimuth ∈[-90°, 90°];

[0034] Usable roof area ≥2;

[0035] System efficiency ∈(0,1);

[0036] S312: Intervariate rules consider the relationships and constraints between different variables.

[0037] Photovoltaic array power ≤Number of photovoltaic modules × Power of a single component ;

[0038] Photovoltaic array power ≤Inverter power ;

[0039] Inverter power ≤Number of inverters × Single inverter capacity ;

[0040] Usable roof area ≥Number of photovoltaic modules ×Area of ​​a single component ;

[0041] Preferably, the specific content of the new sample data verification in step S32 is as follows:

[0042] S321: Prepare 100 new data samples as an independent test set. These samples were not involved in the model training process and can objectively verify the model's performance.

[0043] S322: Input the independent test set into the trained prediction model to obtain the prediction results.

[0044] S323: Calculate performance metrics between predicted and actual values, including RMAE, MAE, R², etc., to evaluate the model's generalization ability and prediction accuracy. Relative Absolute Error (RMAE) formula: ,in For predicted values, This is the actual value. The average of the actual values. For the sample size; the mean absolute error (MAE) formula is: ,in For predicted values, This is the actual value. For the sample size; the formula for the coefficient of determination (R²) is: ,in For predicted values, This is the actual value. The average of the actual values. This represents the number of samples.

[0045] The advantages of this invention compared to the prior art are:

[0046] Compared with existing technologies, this invention mainly solves the key problems existing in current photovoltaic power generation prediction methods, achieving a significant improvement in prediction accuracy and reliability, and has important practical significance and broad application prospects. Addressing the problems of physical methods requiring a large number of input parameters and having limited prediction accuracy, statistical methods struggling to handle complex nonlinear relationships and large-scale data, traditional machine learning methods having limited capacity for large-scale data processing, and deep learning methods having insufficient generalization ability, this invention proposes an AI-based photovoltaic power generation prediction method by combining multilayer perceptron neural networks, dynamic feature selection mechanisms, hierarchical attention networks, and cross-task gradient coupling techniques. This method not only improves prediction accuracy and reduces the prediction error rate, but also enhances the model's generalization ability, enabling it to better adapt to different environments and climate changes. Simultaneously, the rapid training efficiency and high-quality data processing capabilities allow the model to provide decision support more promptly, meeting the accuracy and reliability requirements of large-scale photovoltaic power generation applications. This helps optimize grid dispatch, improve power generation utilization, reduce curtailment, and promote improved energy economic efficiency. The method of this invention is mainly applied in the early stages of building design, providing more accurate photovoltaic power generation predictions for building design, thereby optimizing building energy utilization schemes. Attached Figure Description

[0047] Figure 1 This is a technical flowchart of the AI-based photovoltaic power generation prediction method of the present invention.

[0048] Figure 2 This is a framework diagram of travel demand prediction according to a preferred embodiment of the present invention.

[0049] Figure 3 This is a diagram of the neural network structure used for training the model of this invention.

[0050] Figure 4 This is a prediction-test graph showing the goodness of fit of the model training results of this invention.

[0051] The specific embodiments are described below with reference to the accompanying drawings and preferred embodiments.

[0052] like Figure 1 As shown, a preferred embodiment of the present invention discloses a fault prediction method for an engine, specifically for the fault prediction of a gas turbine engine, including the following steps: S101: Obtain 1080 sets of raw data samples;

[0053] The 14 variables collected are: total horizontal radiation, ambient temperature, photovoltaic module power, number of modules, area of ​​a single module, power of the module array, inverter power, number of inverters, system efficiency, array loss, system loss, tilt angle of the light-receiving surface, azimuth angle, and usable roof area. To make the technical problems, technical solutions, and beneficial effects to be solved by the embodiments of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0054] This invention relates to the field of photovoltaic power generation prediction technology, and more particularly to an AI-based photovoltaic power generation prediction method. With the escalating global energy crisis, solar energy, as a clean energy source, is playing an increasingly important role in energy supply. Photovoltaic power generation is widely used due to its environmental friendliness and renewability; however, due to the influence of meteorological factors, it exhibits volatility and intermittency, posing challenges to power system dispatch management, stability control, and photovoltaic site selection. Existing prediction methods are mainly classified into physical methods, statistical methods, and artificial intelligence algorithms. Physical methods predict power based on the physical characteristics of photovoltaic cell modules using mathematical formulas, but are sensitive to meteorological forecast errors. Statistical methods include time series models and probabilistic predictions, utilizing historical data trends and seasonal information for prediction, but struggle to handle complex nonlinear relationships and large-scale data. Artificial intelligence algorithms, such as traditional machine learning methods like support vector machines, k-nearest neighbors, and random forests, as well as neural network models like BP neural networks, LSTM, GRU, CNN-LSTM, and Transformer, predict power by constructing a mapping function between input and output power, possessing strong nonlinear modeling capabilities and automatic feature extraction capabilities. However, single models often struggle to capture complex nonlinear relationships, easily leading to overfitting or underfitting problems, resulting in insufficient model generalization ability.

[0055] In view of this, this invention proposes an AI-based photovoltaic (PV) power generation prediction method to improve the accuracy and reliability of predictions. While current research focuses on improving and optimizing PV power generation prediction methods, such as distributed PV short-term power prediction methods based on weather change adaptive classification and matching, which effectively improve prediction accuracy by combining feature selection and dimensionality reduction steps, and PV power prediction methods based on the XGBoost-LSTM combined model, existing methods still have shortcomings in handling the uncertainty of meteorological factors, data preprocessing, feature extraction, and model optimization. Therefore, developing an efficient, accurate, and reliable PV power generation prediction method is of great significance for promoting the large-scale application of PV power generation, optimizing grid operation, and improving energy economic efficiency. This invention proposes an AI-based PV power generation prediction method by combining multilayer perceptron neural networks, dynamic feature selection mechanisms, hierarchical attention networks, and cross-task gradient coupling techniques to solve the problems in existing technologies and improve the accuracy and reliability of PV power generation prediction.

[0056] like Figure 1 As shown, the photovoltaic power generation prediction method of the present invention includes the following steps:

[0057] S1 acquires meteorological and photovoltaic system data from different regions, constructing a dataset containing 1080 samples. The data is statistically validated, outliers are cleaned and removed, and the data is standardized. A 14-dimensional variable parameter is output, covering key indicators such as irradiance, ambient temperature, module power, and system efficiency.

[0058] S2 inputs the processed data into the initial prediction model based on a multilayer perceptron neural network (MLP). It performs operations such as feature fusion to integrate different features, dynamic adjustment of weight allocation, and loss function optimization, while simultaneously outputting a prediction data report.

[0059] S3 establishes constraints based on univariate and intervariate rule definitions to assess the data's reasonableness from multiple dimensions. It prepares 100 new data samples as independent test sets to validate model performance, outputting prediction results and performance metrics, including RMAE, MAE, R², etc., to evaluate the model's generalization ability and prediction accuracy.

[0060] S4 compiles data, reports, and verification results to form the final photovoltaic power generation forecast report.

[0061] Preferably, step S1 specifically includes:

[0062] S11 performs multi-source data collection, acquiring meteorological and photovoltaic system data from different regions, and constructs a dataset containing 1080 sets of samples, each set containing detailed data for one day.

[0063] S12 performs statistical verification and data cleaning, using statistical methods to comprehensively verify the collected data, and using SPSS statistical software to identify and remove obvious erroneous data and outliers, such as removing temperature or power values ​​that are outside the physical range.

[0064] S13 performs data standardization, standardizing the cleaned data to transform data from different dimensions to the same scale. For example, the Z-Score standardization method is used to make the mean of the data 0 and the standard deviation 1, as shown in the following formula: ,in, The mean, The standard deviation is denoted as .

[0065] S14 outputs dynamic parameters, extracting and outputting 14-dimensional dynamic parameters, including key indicators such as irradiance, ambient temperature, component power, and system efficiency. These parameters can comprehensively characterize the operating status and external environmental conditions of the photovoltaic power generation system.

[0066] Preferably, in step S2, the processed data is input into the initial prediction model based on a multilayer perceptron neural network (MLP). Operations such as feature fusion integrating different features, dynamic adjustment of weight allocation, and loss function optimization are performed, while a prediction data report is output.

[0067] S21 builds an initial prediction model based on a multilayer perceptron neural network (MLP), designs the network structure including an input layer, multiple hidden layers and an output layer, and determines the number of neurons and activation functions for each layer.

[0068] During model training, S22 uses feature fusion technology to integrate different features, explore the correlation and complementary information between features, and dynamically adjust the weight allocation to make the model pay more attention to features that have an important impact on the prediction target, thereby improving the model's sensitivity to key information and learning ability.

[0069] S23 defines an appropriate loss function (sum of squared errors (SSE) or mean total relative error (MAPE)) to measure the difference between the model's predicted values ​​and the actual values, and uses an optimization algorithm (stochastic gradient descent) to optimize the loss function, continuously adjusting the model's parameters to gradually reduce the loss function value and improve the model's prediction accuracy.

[0070] Preferably, the specific content of building the initial prediction model in step S21 is as follows:

[0071] In the input layer, the number of neurons corresponds to 14-dimensional dynamic parameters. Assume the input data is represented as a vector x=[x1,x2,…,x14], where each xi represents a dynamic parameter.

[0072] In S212, the hidden layer uses a hyperbolic tangent activation function. Assume the hidden layer has... There are neurons, and the weight matrix is... The bias vector is The output h of the hidden layer is calculated as follows: , where tanh(⋅) is the hyperbolic tangent activation function, which is applied element-wise to its input vector.

[0073] In the output layer (S213), the output layer selects the identity activation function and the sum of squares error function based on the requirements of the prediction task. Assume the output layer has... There are neurons (determined based on the specific prediction task), and the weight matrix is... The bias vector is Output of the output layer The calculation is as follows: Here, the identity activation function is used, meaning the activation function of the output layer is a linear function. The sum of squares error function E is calculated as follows: in, These are model predictions. This is the actual value. The sum of squares error function is used to measure the difference between the model's predicted value and the actual value.

[0074] Preferably, the specific content of step S22 is as follows:

[0075] S221 introduces an attention mechanism to dynamically adjust feature weights, focusing on highly correlated parameters (such as irradiance and module temperature). The weight allocation formula is as follows: ,in, For query vector, Let be the i-th eigenvector.

[0076] Preferably, step S3 specifically includes:

[0077] S31 uses SPSS statistical software for model validation, combining univariate and intervariate rule definitions to perform multi-dimensional data rationality assessment of the predicted data. Univariate rule definitions refer to setting reasonable value ranges and constraints for each predicted variable, such as the predicted power generation not being negative. Intervariate rule definitions refer to considering the interrelationships and constraints between different variables, such as the relationship between power generation and irradiance conforming to certain physical relationships. If the predicted results violate these rules, the data is deemed unreasonable.

[0078] S32 prepares 100 new data samples as an independent test set. These samples were not involved in the model training process and can objectively verify the model's performance. The independent test set is input into the trained prediction model to obtain prediction results, and performance metrics such as RMAE, MAE, and R² are calculated between the prediction results and the actual values ​​to evaluate the model's generalization ability and prediction accuracy.

[0079] Preferably, the specific content of the data reasonableness judgment in step S31 is as follows:

[0080] The S311 univariate rule sets reasonable value ranges and constraints for each predictor variable:

[0081] Number of photovoltaic modules ≥1

[0082] Number of inverters ≥1

[0083] Inclination of light-transmitting surface ∈[0°,90°]

[0084] Azimuth ∈[-90°, 90°]

[0085] Usable roof area ≥2

[0086] System efficiency ∈(0,1)

[0087] The S312 intervariate rule considers the relationships and constraints between different variables:

[0088] Photovoltaic array power ≤Number of photovoltaic modules × Power of a single component

[0089] Photovoltaic array power ≤Inverter power

[0090] Inverter power ≤Number of inverters × Single inverter capacity

[0091] Usable roof area ≥Number of photovoltaic modules ×Area of ​​a single component

[0092] Preferably, the specific content of the new sample data verification in step S32 is as follows:

[0093] S321 prepares 100 new data samples as an independent test set. These samples were not involved in the model training process and can objectively verify the model's performance.

[0094] S322 inputs the independent test set into the trained prediction model to obtain the prediction results.

[0095] S323 calculates performance metrics between predicted and actual values, including RMAE, MAE, and R², to evaluate the model's generalization ability and prediction accuracy. The formula for Relative Absolute Error (RMAE) is: ,in For predicted values, This is the actual value. The average of the actual values. For the sample size; the mean absolute error (MAE) formula is: ,in For predicted values, This is the actual value. For the sample size; the formula for the coefficient of determination (R²) is: ,in For predicted values, This is the actual value. The average of the actual values. This represents the number of samples.

[0096] The AI-based photovoltaic power generation prediction method proposed in the preferred embodiment of this invention has the following advantages:

[0097] The advantages of this invention compared to existing technologies are mainly reflected in solving key problems in existing photovoltaic power generation prediction methods, achieving a significant improvement in prediction accuracy and reliability, and possessing important practical significance and broad application prospects. This invention addresses the problems of physical methods requiring a large number of input parameters and having limited prediction accuracy, statistical methods struggling to handle complex nonlinear relationships and large-scale data, traditional machine learning methods having limited capacity for large-scale data processing, and deep learning methods having insufficient generalization ability. By combining multilayer perceptron neural networks, dynamic feature selection mechanisms, hierarchical attention networks, and cross-task gradient coupling techniques, this invention proposes an AI-based photovoltaic power generation prediction method. This method not only improves prediction accuracy and reduces the prediction error rate but also enhances the model's generalization ability, enabling it to better adapt to different environments and climate changes. Simultaneously, the rapid training efficiency and high-quality data processing capabilities allow the model to provide decision support more promptly, meeting the accuracy and reliability requirements of large-scale photovoltaic power generation applications. This helps optimize grid dispatch, improve power generation utilization, reduce curtailment, and promote improved energy economic efficiency. The method of this invention is mainly applied in the early stages of building design, providing more accurate photovoltaic power generation predictions for building design, thereby optimizing building energy utilization schemes.

[0098] The following describes the route planning method for cooperative buses proposed in the preferred embodiment of the present invention in conjunction with specific examples. (1) Obtain 100 sets of data from different cities (70% training, 30% validation) and compare the prediction performance of the traditional RBF model with that of the present invention.

[0099] The results show that the traditional RBF model has an RMAE of 5.7%, an R² of 0.923, and a training time of 0.53 seconds; while the model of this invention reduces the RMAE to 1.17%~2.03%, increases the R² to 0.997, and requires only 0.07 seconds for training. The data demonstrates that the prediction error of the model of this invention is reduced by 84.7% compared to PVsyst, and the training efficiency is improved by 86.8% compared to the RBF model, verifying its high accuracy and strong robustness.

[0100] Table 2 Comparison of Results

[0101] RMAE R² Training time (s) The method proposed in this invention 0.00117 0.923 0.07 Traditional RBF model 0.00570 0.997 0.53

[0102] The background section of this invention may include background information about the problems or environment of the invention, rather than a description of prior art by others. Therefore, the content included in the background section is not an admission of prior art by the applicant. The above description is a further detailed description of the invention in conjunction with specific / preferred embodiments, and it should not be considered that the specific implementation of the invention is limited to these descriptions. For those skilled in the art, several substitutions or modifications can be made to these described embodiments without departing from the concept of the invention, and these substitutions or modifications should be considered to fall within the protection scope of the invention. In the description of this specification, the reference to terms such as "an embodiment," "some embodiments," "preferred embodiment," "example," "specific example," or "some examples," etc., means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and combine different embodiments or examples and features of different embodiments or examples described in this specification. Although embodiments of the invention and their advantages have been described in detail, it should be understood that various changes, substitutions and alterations may be made herein without departing from the scope defined by the appended claims.

Claims

1. An AI-based method for predicting photovoltaic power generation, characterized in that, Includes the following steps: S1: Acquire meteorological and photovoltaic system data from different regions, construct a dataset containing 1080 samples; statistically validate the data, clean and remove outliers, and standardize the data to complete the processing of meteorological and photovoltaic system data from different regions, outputting 14-dimensional variable parameters, including two-dimensional meteorological variables and twelve-dimensional photovoltaic system variables; the two-dimensional meteorological variables are total horizontal radiation and ambient temperature; the twelve-dimensional photovoltaic system variables are photovoltaic module power, number of modules, area of ​​a single module, power of module array, inverter power, number of inverters, system efficiency, array loss, system loss, tilt angle of the light-receiving surface, azimuth angle, and usable roof area; S2: Input the processed meteorological and photovoltaic system data from different regions into the initial prediction model based on the MLP (Multilayer Perceptron) neural network; perform feature fusion to integrate different features, dynamically adjust weight allocation, and optimize the loss function. Specifically, when constructing the initial prediction model, the number of neurons in its input layer corresponds to 14-dimensional dynamic parameters, the hidden layer uses the hyperbolic tangent activation function, and the output layer uses the identity activation function and the sum of squares error function. Next, input the collected meteorological and photovoltaic system data from different regions into the model, use feature fusion technology to integrate different features, and dynamically adjust the weight allocation so that the model focuses on features important to the prediction target; finally, define the mean squared error or mean absolute error as the loss function, use the stochastic gradient descent method to optimize the loss function, and output the prediction data report. S3: Establish constraints for data combining univariate and cross-variate rule definitions, and make multi-dimensional judgments on the rationality of the data; prepare 100 sets of new data samples as independent test sets to verify the model performance, output prediction results and performance indicators, including RMAE, MAE, R², and evaluate the model's generalization ability and prediction accuracy. S4: Compile data, reports, and verification results to form the final photovoltaic power generation forecast report.

2. The AI-based photovoltaic power generation prediction method according to claim 1, characterized in that, Step S1 specifically includes: S11: Conduct multi-source data collection to obtain meteorological data and photovoltaic system data from different regions, and construct a dataset containing 1080 sets of samples, each set of samples containing meteorological data and photovoltaic system data from different regions within a day; S12: Perform statistical verification and data cleaning on the samples in the dataset. Use statistical methods to fully verify the sample data. Use SPSS statistical software to identify and remove obvious erroneous data and outliers, and remove temperature or power values ​​that are outside the physical range. S13: Perform data standardization. Standardize the cleaned sample data to transform the sample data from different dimensions to the same scale. Use the Z-Score standardization method to make the mean of the data 0 and the standard deviation 1. The formula is as follows: Where x is a value in the sample data. The mean, The z-value represents the standard deviation; the calculated z-value is a measure of the distance between a value in the sample data and the mean, expressed in standard deviations. S14: Output dynamic parameters. Extract and output 14-dimensional dynamic parameters, including key indicators such as irradiance, ambient temperature, component power, and system efficiency. These parameters can comprehensively characterize the operating status and external environmental conditions of the photovoltaic power generation system.

3. The AI-based photovoltaic power generation prediction method according to claim 1, characterized in that, In step S2, the processed data is input into the initial prediction model based on the MLP multilayer perceptron neural network; feature fusion is performed to integrate different features, weight allocation is dynamically adjusted, and loss function is optimized, while a prediction data report is output. S21: Build an initial prediction model based on MLP multilayer perceptron neural network, design the network structure, including input layer, multiple hidden layers and output layer, and determine the number of neurons and activation function of each layer; S22: During model training, feature fusion technology is used to integrate different features, explore the correlation and complementary information between features, and dynamically adjust the weight allocation so that the prediction model pays more attention to features that have an important impact on the prediction target, thereby improving the prediction model's sensitivity to key information and learning ability. S23: Define a suitable loss function, including the sum of squared errors (SSE) or the mean total relative error (MAPE), to measure the difference between the predicted values ​​and the actual values ​​of the prediction model. Then, use an optimization algorithm, namely stochastic gradient descent, to optimize the loss function, continuously adjust the parameters of the prediction model, and gradually reduce the value of the loss function to improve the prediction accuracy of the prediction model.

4. The AI-based photovoltaic power generation prediction method according to claim 3, characterized in that, The specific details of building the initial prediction model in step S21 are as follows: S211: In the input layer, the number of neurons in the input layer corresponds to 14-dimensional dynamic parameters. The input data is represented as a vector x=[x1,x2,…xi,…,x14], where each xi represents a dynamic parameter. S212: In the hidden layer, the hyperbolic tangent activation function is used, and the hidden layer has... There are neurons, and the weight matrix is... The bias vector is The output h of the hidden layer is calculated as follows: , where tanh(⋅) is the hyperbolic tangent activation function, applied element-wise to its input vector; S213: In the output layer, the output layer selects the identity activation function and the sum of squares error function according to the requirements of the prediction task; the output layer has There are 10 neurons, the number of which is determined based on the specific photovoltaic power generation prediction task, and the weight matrix is ​​as follows: The bias vector is ,in, Represents the set of real numbers; output of the output layer The calculation is as follows: The identity activation function is used, meaning the activation function of the output layer is a linear function; the sum of squares error function E is calculated as follows: in, These are model predictions. is the actual value, while i is the counting index variable used to traverse the output layer neurons; the sum of squares error function is used to measure the difference between the model's predicted value and the actual value.

5. The AI-based photovoltaic power generation prediction method according to claim 4, characterized in that, The specific details of step S22 are as follows: S221: Introduces an attention mechanism to dynamically adjust feature weights, focusing on highly correlated parameters; the weight allocation formula is: ,in, For query vector, Let be the i-th dimensional feature vector; T represents the transpose, used to calculate the similarity between the query vector and the feature vector; j is an index variable used to iterate through the dimensions of all feature vectors, ranging from 1 to 14.

6. The AI-based photovoltaic power generation prediction method according to claim 1, characterized in that, Step S3 specifically includes: S31: Through model validation in SPSS statistical software, combined with the definitions of univariate and intervariate rules, a multi-dimensional judgment of the reasonableness of the predicted data is made. The definition of univariate rules refers to setting a reasonable range of values ​​and constraints for each predicted variable, such as the predicted power generation should not be negative. The definition of intervariate rules refers to considering the interrelationships and constraints between different variables, such as the relationship between power generation and irradiance should conform to a certain physical relationship. If the prediction results violate the rules, the data is judged to be unreasonable. S32: Prepare 100 new data samples as independent test sets. The new data samples are not involved in the training process of the prediction model, so as to objectively verify the performance of the model. Input the independent test sets into the trained prediction model to obtain the prediction results, and calculate the performance indicators between the prediction results and the actual values, including RMAE, MAE, and R², to evaluate the generalization ability and prediction accuracy of the prediction model.

7. The AI-based photovoltaic power generation prediction method according to claim 6, characterized in that, The specific content of the data reasonableness judgment in step S31 is as follows: S311: Univariate rules involve setting reasonable ranges of values ​​and constraints for each predictor variable. Number of photovoltaic modules ≥1; Number of inverters ≥1; Inclination of light-transmitting surface ∈[0°,90°]; Azimuth ∈[-90°, 90°]; Usable roof area ≥2; System efficiency ∈(0,1); S312: Intervariate rules consider the relationships and constraints between different variables. Photovoltaic array power ≤Number of photovoltaic modules × Power of a single component ; Photovoltaic array power ≤Inverter power ; Inverter power ≤Number of inverters × Single inverter capacity ; Usable roof area ≥Number of photovoltaic modules ×Area of ​​a single component .

8. The AI-based photovoltaic power generation prediction method according to claim 6, characterized in that, The specific details of the new sample data validation in step S32 are as follows: S321: Prepare 100 new data samples as an independent test set. These samples were not involved in the model training process and can objectively verify the model's performance. S322: Input the independent test set into the trained prediction model to obtain the prediction results; S323: Calculate performance metrics between predicted results and actual values, including RMAE, MAE, R², etc., to evaluate the model's generalization ability and prediction accuracy; Relative Absolute Error RMAE: ,in For predicted values, This is the actual value. The average of the actual values. Sample size; Mean Absolute Error (MAE): ,in For predicted values, This is the actual value. Where is the sample size; coefficient of determination R²: ,in For predicted values, This is the actual value. The average of the actual values. This represents the number of samples.

Images