High-precision prediction method and system for photovoltaic power generation power

Abstract

The invention relates to the technical field of photovoltaic power generation power prediction, in particular to a photovoltaic power generation power high-precision prediction method and system. According to the technical scheme, the method comprises the following steps: data acquisition and preprocessing: historical photovoltaic power generation power data are widely collected, the time span covers at least one year in the past so as to obtain power generation power change information under different seasons, weather conditions and time periods, and meanwhile, meteorological data corresponding to the power generation power change information are collected; and cleaning the collected data, and removing missing values through data integrity check. The power generation efficiency and economy are improved, the operation and maintenance plan can be optimized, and the energy storage utilization efficiency is improved; the power grid stability and reliability are enhanced, scheduling is assisted, resource configuration is optimized, power market operation is greatly improved, support is provided for transaction decision making, renewable energy consumption is promoted, the system adaptability and the intelligent level are improved, real-time tracking and updating can be achieved, and intelligent management of a power station is assisted.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power prediction, and in particular to a method and system for high-precision prediction of photovoltaic power. Background Art

[0002] As the global demand for clean energy continues to increase, photovoltaic power generation has developed rapidly as a sustainable and pollution-free energy production method. Large-scale photovoltaic power stations continue to emerge, and their share in the energy supply system is increasing. However, photovoltaic power generation is affected by many factors and is characterized by intermittency and volatility. Accurate prediction of photovoltaic power generation is essential for the efficient operation and management of photovoltaic power stations, the stable dispatch of power grids, and the effective operation of the power market. For example, in the operation of photovoltaic power stations, reasonable power forecasting can help power station operators arrange maintenance plans in advance and optimize the charging and discharging strategies of energy storage systems; for the power grid, accurate forecasting results can enable the dispatching department to better arrange the power generation plans of other power sources, ensure the balance and stability of power supply, and reduce the potential safety hazards of the power grid caused by fluctuations in photovoltaic power generation.

[0003] Traditional statistical methods: Some early prediction methods based on statistical analysis, such as the autoregressive moving average model (ARMA) in time series analysis, are effective in processing some stable data, but their prediction accuracy is often low for data such as photovoltaic power generation that is affected by complex meteorological factors and nonlinear relationships.

[0004] In recent years, machine learning and artificial intelligence technologies have been widely used in photovoltaic power generation prediction. However, a single machine learning model such as support vector machine (SVM) may face problems such as high computational complexity and difficulty in selecting kernel functions when processing large-scale, high-dimensional data; although random forest (RF) has a certain degree of robustness to noisy data, it may not be ideal when processing data with long-term time dependencies. For neural network models, such as multi-layer perceptron (MLP), the order of data may be ignored when processing time series data, and simple recurrent neural network (RNN) has the problem of gradient disappearance or explosion, which limits its application in long-term prediction.

[0005] In photovoltaic power generation prediction, data quality and feature selection have a key impact on the prediction results. On the one hand, during the data collection process, sensors may have measurement errors, data loss, or outliers. Different types of sensors have different accuracy and stability under different environmental conditions, which increases the difficulty of obtaining accurate data. On the other hand, extracting effective features from a large amount of raw data is a complex process. Existing feature engineering methods may not be able to fully mine features that are highly correlated with power generation, or may introduce too many irrelevant features, resulting in inefficient model training and reduced prediction accuracy.

[0006] During the long-term operation of photovoltaic power stations, their power generation is affected by component aging and environmental changes (such as climate change, changes in the surrounding environment, etc.), and the change pattern of power generation may change. Therefore, the prediction model needs to be able to be updated in time to adapt to these changes. However, some existing prediction systems are not flexible enough in terms of model update mechanisms and cannot optimize the models with new data in a timely manner. At the same time, in terms of real-time prediction, due to factors such as data transmission delays and model calculation time, some systems cannot meet the requirements for fast and accurate real-time prediction of power generation, affecting their value in practical applications.

[0007] In summary, the existing photovoltaic power prediction technology has certain limitations in terms of accuracy, data processing, model adaptability and real-time performance, and a new high-precision prediction method and system are urgently needed to meet the needs of the growing photovoltaic power generation industry. This application proposes a high-precision photovoltaic power prediction method and system. Summary of the invention

[0008] The purpose of the present invention is to propose a high-precision photovoltaic power prediction method and system to address the problem that the existing photovoltaic power prediction technology in the background technology has certain limitations in terms of accuracy, data processing, model adaptability and real-time performance.

[0009] In one aspect, the present invention provides a method for high-precision prediction of photovoltaic power generation, comprising the following steps: Data collection and preprocessing: Collect historical photovoltaic power generation data, covering at least the past year, to obtain information on power generation changes under different seasons, weather conditions and time periods, and collect corresponding meteorological data; The collected data were cleaned and the missing values ​​were removed through data integrity check. For data segments with continuous missing values ​​of more than 3 hours, linear interpolation was used to supplement them. The method based on the 3σ principle was used to remove outliers, that is, if the value of a data point deviated from the mean by more than 3 times the standard deviation, it was regarded as an outlier and removed. The mean and standard deviation were obtained by statistical calculation of historical data. Perform data normalization and use linear normalization method to map the data to the [0,1] interval. The specific formula is: ,in is the original data, is the minimum value in the data set, is the maximum value in the data set, is the normalized data; Feature Engineering: Extracting features related to photovoltaic power generation from the preprocessed data; The extracted features are combined and transformed, the ratio of irradiance to temperature and the product of humidity and wind speed are used as new combined features, the time features are processed by one-hot encoding, and the categorical variables of different seasons and 0-23h are converted into binary vectors; By calculating the correlation coefficient between meteorological features, a subset of meteorological features with a high correlation with photovoltaic power generation is screened out. The calculation formula of the correlation coefficient is: , in and The two features are values, and are the means of the two features, is the sample size, is the correlation coefficient, and the correlation threshold is set to 0.5. Features above this threshold are retained, and features below the threshold are removed; Model building and training: A hybrid model based on deep learning is used for construction, which includes a convolutional neural network and a long short-term memory network: The CNN part uses a structure in which multiple convolutional layers and pooling layers are connected alternately. First, an input layer is set, whose shape matches the dimension of the feature vector; The first convolution layer uses 32 convolution kernels of size 3×3 and ReLU activation function for feature extraction and preliminary nonlinear transformation; The maximum pooling layer with a pooling window size of 2×2 is used to reduce the data dimension and extract the main features; Connect two more convolutional layers, using 64 and 128 3×3 kernels respectively, also using the ReLU activation function, and follow the corresponding max pooling layer; The multi-dimensional feature vector output by the convolutional layer is converted into a one-dimensional vector through a flat layer and input into the LSTM part; The LSTM part contains multiple LSTM units, and the number of neurons is adjusted according to the size and feature complexity of the data set: the optimal number of neurons is determined by cross-validation, and the search is performed in the range of [64, 256]. Each LSTM unit receives the output result of CNN as input to perform time series feature learning and power prediction. After the LSTM layer, a fully connected layer is added, and its number of neurons is the same as the dimension of the prediction target. The prediction target is photovoltaic power generation, and a linear activation function is used to output the final prediction value; During the training process, an adaptive learning rate adjustment algorithm and early stopping method are used to prevent overfitting. The Adam algorithm dynamically adjusts the learning rate based on the first-order moment estimate and the second-order moment estimate of the gradient. The calculation formula is: , in, For the The model parameters at the iteration, is the initial learning rate, set to 0.001, is the first moment estimate of the gradient, is the second-order moment estimate of the gradient, is a very small constant, set to 1e-8, to prevent the denominator from being 0; the early stopping method sets the patience value to 5, that is, when the loss function value on the validation set no longer decreases after 5 consecutive iterations, the training is stopped; the mean square error is used as the loss function during the training process, and the model parameters are updated through the back propagation algorithm; Model evaluation and optimization: Using Root Mean Square Error , mean absolute error , coefficient of determination and mean absolute percentage error Evaluation indicators evaluate the trained model; The calculation formula is: ; The calculation formula is: ; The calculation formula is: ; The calculation formula is: ; in is the actual value, is the predicted value, is the actual mean value, is the sample size; According to the evaluation results, the model parameters are adjusted and the structure is optimized. and , increase the complexity of the model, including adding convolutional layers or The number of units, adjust the size and number of convolution kernels; if or , optimize feature engineering, adjust feature combinations or filter more relevant features; Through repeated trials and adjustments, the model achieves optimal performance on the validation set, that is, each evaluation indicator reaches the optimal value or meets the preset threshold requirements; Real-time prediction: The real-time collected meteorological data and equipment operation data are preprocessed to make their format and features consistent with the training data, and the real-time data are extracted and transformed according to the steps of feature engineering.

[0010] Optionally, in the data collection and preprocessing steps, for the supplement of missing data values, when the continuous missing time is between 1-2 hours, linear interpolation is used; when the continuous missing time is between 2-3 hours, linear interpolation is first used to supplement, and then the supplemented data is smoothed using a moving average method with a window size of 3 to reduce the error fluctuation caused by interpolation; for data segments with missing time exceeding 3 hours, linear interpolation is used to supplement, and correction is made based on the mean of similar day data. The selection of similar days is based on the similarity judgment of weather type, season and date type factors. By calculating the similarity index, the top 5 dates in similarity are selected as similar days, and the data mean of these similar days corresponding to the missing time period is taken to correct the missing values ​​after linear interpolation.

[0011] Calculate the similarity index and use the Euclidean distance algorithm. Suppose the meteorological feature vector of the data day to be processed is , Indicates the day The meteorological feature value of the candidate similar day is , the Euclidean distance between the two is: , Suppose the comprehensive weight coefficients of date type and seasonal factors are , , ,Pick =0.4, =0.6, calculate the comprehensive similarity: , in, is the Euclidean distance based on meteorological characteristics, The maximum value of the Euclidean distance between the meteorological characteristics of all candidate days and the day to be processed; It is the quantified distance based on date type and seasonal factors. Weekdays, weekends, statutory holidays are assigned values ​​respectively, and different seasons are also assigned values ​​respectively. The absolute value of the difference is the distance. is the maximum value of the corresponding distance; Based on comprehensive similarity , select the top 5 dates in similarity from the past historical data as similar days, and take the data mean of the missing time period corresponding to these similar days to correct the missing values ​​after linear interpolation.

[0012] Optionally, in the data collection and preprocessing step, the data normalization process adopts a linear normalization method, and the normalized data is standardized so that its mean is 0 and its standard deviation is 1. The standardization process formula is: , in is the data after linear normalization, is the mean of the normalized data, is the standard deviation of the normalized data, is the standardized data.

[0013] Optionally, in the feature engineering, after the time features are one-hot encoded, a principal component analysis method is used to reduce the dimension of the encoded time features, and the principal components with a cumulative contribution rate of more than 90% are retained.

[0014] Optionally, in the model construction and training, in the convolution layer of the CNN part, a batch normalization layer is added after each convolution layer, and the calculation formula of the batch normalization layer is: , , in is the input data, and are the mean and variance of the batch data, is a very small constant, set to 1e-5 to prevent the denominator from being 0. is the normalized data, and is a learnable parameter, adjusted through training, is the output of the batch normalization layer.

[0015] Optionally, in the model construction and training, the number of neurons in the LSTM part is determined by a combination of grid search and cross-validation. First, a search range for the number of neurons is set, the range is [32, 64, 128, 256, 512], and then the data set is divided into a training set, a validation set, and a test set; for each neuron number setting, the model is trained on the training set, the model performance is evaluated on the validation set, and the corresponding evaluation index values ​​are recorded; finally, the number of neurons that performs best on the validation set is selected as the final parameters of the model.

[0016] Optionally, in the model construction and training, the adaptive learning rate adjustment algorithm adopts the Adam algorithm, combined with the learning rate decay strategy, the initial learning rate is set to 0.001, and after every 10 iterations, the learning rate decays to 0.9 times the original, that is, ,in For the The learning rate for each epoch.

[0017] Optionally, in the model evaluation and optimization, use , , and In addition to the evaluation indicators, the relative root mean square error is also introduced As an evaluation indicator, the calculation formula is: , in, is the root mean square error, is the mean of the actual values.

[0018] Optionally, in the real-time prediction, abnormal data points caused by meteorological sensor failure adopt an outlier detection and replacement strategy, and determine whether they are outliers by calculating the degree of deviation between the real-time data points and the mean and standard deviation of the historical data in the sliding window. If the degree of deviation exceeds 3 times the standard deviation, it is regarded as an outlier; for outliers, the median of the historical data in the sliding window is used for replacement, and the processed real-time data is then input into the model for prediction.

[0019] In another aspect, the present invention provides a photovoltaic power generation high-precision prediction system, the system comprising: Data acquisition and preprocessing module, the data acquisition and preprocessing module includes: The data collection unit collects extensive historical photovoltaic power data, with a time span of no less than one year, covering power generation change information under different seasons, weather conditions and time periods, and simultaneously collects corresponding meteorological data, including hourly irradiance, temperature, humidity, wind speed, and air pressure, and obtains relevant equipment operation data, including the model of photovoltaic panels, installation angle, aging degree, equipment failure rate, and maintenance records; The data cleaning unit identifies and processes missing values ​​through data integrity checks. For data segments that are missing for 1-2 hours in a row, linear interpolation is used to supplement them. For data segments that are missing for 2-3 hours, linear interpolation is used first and then moving average smoothing is used. The moving average window size is 3. For data segments that are missing for more than 3 hours, in addition to linear interpolation, the mean value of similar daily data is used for correction. Outliers are removed based on the 3σ principle, that is, if a data point deviates from the mean by more than 3 times the standard deviation, it is considered an outlier and removed. The mean and standard deviation are calculated based on historical data statistics. Data normalization and standardization unit, use linear normalization method to map data to the [0,1] interval, the formula is ,in is the original data, is the minimum value in the data set, is the maximum value in the data set, The normalized data is standardized so that its mean is 0 and its standard deviation is 1. The formula is: , in, is the mean, is the standard deviation, is the standardized data; Feature engineering module, the feature engineering module includes: Feature extraction and combination unit, extracts features from preprocessed data, including time features, meteorological features, and equipment features; constructs combined features such as the ratio of irradiance to temperature, and the product of humidity and wind speed; The feature processing and screening unit performs one-hot encoding on the time features, and then uses the principal component analysis method to reduce the dimension, retaining the principal components with a cumulative contribution rate of more than 90%; by calculating the correlation coefficients between meteorological features and between combinations, the feature subset with high correlation and significant impact on photovoltaic power generation is screened out; Model building and training module, the model building and training module includes: The model architecture design unit uses a hybrid model based on deep learning, including convolutional neural networks and long short-term memory networks; Training mechanism unit, the training process uses the Adam algorithm as the adaptive learning rate adjustment algorithm, the initial learning rate is set to 0.001, the formula is: , in are model parameters, is the learning rate, and are the first-order and second-order moment estimates of the gradient, The denominator is set to 1e-8 to prevent it from being 0. At the same time, combined with the learning rate decay strategy, the learning rate decays to 0.9 times of the original value every 10 epochs. The early stopping method is used to prevent overfitting. The patience value is set to 5. When the loss function value of the validation set does not decrease for 5 consecutive epochs, the training is stopped. The mean square error (MSE) is used as the loss function, and the model parameters are updated through the back propagation algorithm. Model evaluation and optimization module, the model evaluation and optimization module includes: The evaluation index system unit uses root mean square error, mean absolute error, determination coefficient, mean absolute percentage error and relative root mean square error multi-dimensional evaluation indicators to comprehensively evaluate the model performance; The calculation formula is: ; The calculation formula is: ; The calculation formula is: ; The calculation formula is: ; The calculation formula is: ; in, is the actual value, is the predicted value, is the actual mean value, is the sample size; The optimization strategy unit adjusts the parameters and optimizes the structure of the model according to the evaluation results; Real-time prediction module, the real-time prediction module includes: The data processing and prediction unit performs preprocessing and feature extraction transformation on the real-time collected meteorological data and equipment operation data, uses a sliding window to process real-time data, and the window size is set to 5 time points. The data in the window is input into the optimized model for prediction, and the average value of the predicted value in the window is taken as the final real-time prediction result; The exception handling and model updating unit detects outliers by calculating the degree of deviation between the real-time data points and the mean and standard deviation of the historical data in the sliding window, and then replaces the outliers with the median of the historical data in the window before making predictions; the latest real-time data is added to the historical data set every week to retrain and optimize the model.

[0020] Compared with the prior art, the present invention has at least one of the following beneficial technical effects: 1. Through high-precision power forecasting, the present invention enables power station operators to predict changes in power generation in advance, which can effectively improve the overall power generation efficiency of photovoltaic power stations and increase power generation revenue. For photovoltaic power stations equipped with energy storage systems, accurate power forecasting helps to optimize the charging and discharging strategies of energy storage systems.

[0021] 2. When the photovoltaic power generation power fluctuates greatly, the output power of other power sources can be adjusted in time to ensure the balance of supply and demand of the power system. It helps to reduce the negative impact of the intermittent and volatile photovoltaic power generation on the stability of the power grid, reduce the risk of power grid failure, and enhance the reliability of the power grid. Accurate prediction enables the power grid to better plan and allocate power transmission resources.

[0022] 3. Accurate photovoltaic power generation forecast provides a strong decision-making basis for power station operators to participate in power transactions. Operators can formulate power sales strategies more reasonably based on the forecast results. For example, when the power generation power is expected to be high, they can strive for more power sales share in the power market; when the power is low, they can make power purchase plans in advance or adjust trading strategies to reduce market risks and improve competitiveness in the power market.

[0023] 4. High-precision forecasts help the power market better accept photovoltaic power generation. By understanding the changes in photovoltaic power generation in advance, the power market can better coordinate the transactions and consumption of different energy sources, reduce market transaction barriers caused by the uncertainty of photovoltaic power generation, and promote the proportion of renewable energy in the entire energy market.

[0024] 5. The system can collect data and make predictions in real time, and update the model regularly so that it can adapt well to various changes in the operation of photovoltaic power plants. It provides support for the intelligent management of photovoltaic power plants. Through integrated data collection, processing and prediction functions, the entire system can realize automated data-driven decision-making.

[0025] In summary, the invention improves power generation efficiency and economy, can optimize operation and maintenance plans, and improve energy storage utilization efficiency; enhance power grid stability and reliability, assist in scheduling, optimize resource allocation, greatly improve power market operations, provide support for trading decisions, promote renewable energy consumption, improve system adaptability and intelligence, and can track and update in real time to assist in intelligent management of power stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flow chart of a method for high-precision prediction of photovoltaic power generation; Figure 2 This is a principle block diagram of a high-precision photovoltaic power prediction system. DETAILED DESCRIPTION

[0027] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0028] Embodiment 1:

[0029] like Figure 1 As shown, this embodiment discloses a method for high-precision prediction of photovoltaic power generation, comprising the following steps: Data collection and preprocessing steps: Collect historical photovoltaic power data extensively, covering at least the past year, to obtain information on power generation changes in different seasons, weather conditions and time periods. At the same time, collect corresponding meteorological data, including but not limited to hourly irradiance, temperature, humidity, wind speed, air pressure, and related equipment operation data, such as photovoltaic panel model, installation angle, aging degree, equipment failure rate, and maintenance records.

[0030] The collected data are cleaned and the missing values ​​are removed through data integrity check. For data segments that are missing for more than a certain period of time (such as 3 hours), linear interpolation is used to supplement them. For the removal of outliers, a method based on the 3σ principle is used, that is, if the value of a data point deviates from the mean by more than 3 times the standard deviation, it is regarded as an outlier and removed, where the mean and standard deviation are obtained by statistical calculation of historical data. For the supplement of missing data values, when the continuous missing time is between 1-2 hours, linear interpolation is used; when the continuous missing time is between 2-3 hours, linear interpolation is first used to supplement, and then the supplemented data is smoothed, using the moving average method with a window size of 3 to reduce the error fluctuation caused by interpolation. For data segments with missing time exceeding 3 hours, in addition to using linear interpolation to supplement, corrections are also made based on the mean of similar day data. The selection of similar days is based on the similarity judgment of weather type, season and date type (working day or holiday). By calculating the similarity index, the top 5 dates in similarity are selected as similar days. The mean of the data corresponding to the missing time period of these similar days is taken to correct the missing values ​​after linear interpolation.

[0031] Calculate the similarity index and use the Euclidean distance algorithm. Suppose the meteorological feature vector of the data day to be processed is , Indicates the day The meteorological feature value of the candidate similar day is , the Euclidean distance between the two is: , Suppose the comprehensive weight coefficients of date type and seasonal factors are , , ,Pick =0.4, =0.6, calculate the comprehensive similarity: , in, is the Euclidean distance based on meteorological characteristics, The maximum value of the Euclidean distance between the meteorological characteristics of all candidate days and the day to be processed; It is the quantified distance based on date type and seasonal factors. Weekdays, weekends, statutory holidays are assigned values ​​respectively, and different seasons are also assigned values ​​respectively. The absolute value of the difference is the distance. is the maximum value of the corresponding distance; Based on comprehensive similarity , select the top 5 dates in similarity from the past historical data as similar days, and take the data mean of the missing time period corresponding to these similar days to correct the missing values ​​after linear interpolation.

[0032] Perform data normalization and use linear normalization method to map the data to the [0,1] interval. The specific formula is: ,in is the original data, is the minimum value in the data set, is the maximum value in the data set, The normalized data is used to eliminate the influence of different data dimensions and magnitudes, and improve the efficiency and stability of model training. In addition to the linear normalization method, the normalized data is also standardized to make its mean 0 and standard deviation 1, which further improves the stability of the data and the convergence speed of the model. The standardization formula is: , in is the data after linear normalization, is the mean of the normalized data, is the standard deviation of the normalized data, is the standardized data.

[0033] Feature Engineering Steps: Extract features related to photovoltaic power generation from preprocessed data. In terms of time features, it not only includes basic information such as hours, seasons, and dates, but also further refines to different time periods of the day (such as morning, afternoon, evening, etc.), as well as whether it is a special time factor for holidays. In addition to conventional irradiance, temperature, humidity, and wind speed, meteorological features also consider meteorological change trends, such as the rate of change of irradiance and the range of temperature rise and fall. Equipment features cover the material, brand, and production batch details of photovoltaic panels, as well as the cumulative operating time of the equipment and the time interval between the last maintenance.

[0034] The extracted features are combined and transformed, and the ratio of irradiance to temperature and the product of humidity and wind speed are used as new combined features. The time features are processed by One-Hot Encoding, and the categorical variables of season (spring, summer, autumn, winter) and hour (0-23) are converted into binary vectors to better represent the nonlinear relationship between time features and photovoltaic power generation.

[0035] By calculating the correlation coefficient between meteorological features, a subset of meteorological features with a high correlation with photovoltaic power generation is screened out to reduce feature dimensions, reduce model complexity, and improve model training efficiency and prediction accuracy. The calculation formula for the correlation coefficient is: , in and The two features are values, and are the means of the two features, is the sample size, is the correlation coefficient, and the correlation threshold is set to 0.5. Features above this threshold are retained, and features below the threshold are removed. In the feature engineering step, after the time feature is one-hot encoded, in order to further reduce the feature dimension and avoid the problem of feature sparsity, the principal component analysis (PCA) method is used to reduce the dimension of the encoded time feature, retaining the principal components with a cumulative contribution rate of more than 90%, so as to extract the main information of the time feature and reduce the impact of redundant information on the model.

[0036] Model building and training steps: A hybrid model based on deep learning is used for construction, which includes a convolutional neural network (CNN) and a long short-term memory network (LSTM). The CNN part uses a structure with multiple convolutional layers and pooling layers connected alternately. First, an input layer is set, whose shape matches the dimension of the feature vector. Then, the first convolutional layer uses 32 convolution kernels of size 3×3 and uses the ReLU activation function to perform feature extraction and preliminary nonlinear transformation. This is followed by a maximum pooling layer with a pooling window size of 2×2 to reduce the data dimension and extract the main features. After that, two more convolutional layers are connected, using 64 and 128 3×3 convolution kernels respectively, also using the ReLU activation function, and followed by the corresponding maximum pooling layer. Finally, a flatten layer is used to convert the multi-dimensional feature vector output by the convolutional layer into a one-dimensional vector for input to the LSTM part.

[0037] The LSTM part contains multiple LSTM units, and the number of neurons is adjusted according to the size and feature complexity of the data set. The optimal number of neurons is determined by cross-validation, and is generally searched in the range of [64, 256]. Each LSTM unit receives the output of the CNN as input for time series feature learning and power prediction. After the LSTM layer, a fully connected layer (Dense layer) is added, whose number of neurons is the same as the dimension of the prediction target (photovoltaic power generation), and a linear activation function is used to output the final prediction value.

[0038] During the training process, an adaptive learning rate adjustment algorithm (Adam algorithm) and early stopping method are used to prevent overfitting. The Adam algorithm dynamically adjusts the learning rate based on the first-order moment estimate and the second-order moment estimate of the gradient. The calculation formula is: , in, For the The model parameters at the iteration, is the initial learning rate (set to 0.001), is the first moment estimate of the gradient, is the second-order moment estimate of the gradient, is a very small constant (set to 1e-8) to prevent the denominator from being 0. The early stopping method sets the patience value to 5, that is, when the loss function value on the validation set no longer decreases for 5 consecutive epochs (iteration cycles), the training is stopped to avoid overtraining of the model. The mean square error (MSE) is used as the loss function during training. The parameters of the model are updated through the back propagation algorithm to optimize the performance of the model to improve the prediction accuracy and generalization ability of the model. In the model construction and training steps, in the convolution layer of the CNN part, a batch normalization layer is added after each convolution layer to accelerate the training convergence speed of the model and reduce the gradient disappearance and explosion problems. The calculation formula of the batch normalization layer is: , , in is the input data, and are the mean and variance of the batch data, is a very small constant (set to 1e-5) to prevent the denominator from being zero. is the normalized data, and is a learnable parameter, adjusted through training, is the output of the batch normalization layer.

[0039] In the model construction and training steps, the number of neurons in the LSTM part is determined by combining grid search and cross-validation. First, a search range for the number of neurons is set, such as [32, 64, 128, 256, 512], and then the data set is divided into training set, validation set, and test set. For each setting of the number of neurons, the model is trained on the training set, the model performance is evaluated on the validation set, and the corresponding evaluation index values ​​are recorded. Finally, the number of neurons that performs best on the validation set is selected as the final parameter of the model to improve the model's learning ability and prediction accuracy for time series data, while avoiding overfitting and underfitting problems.

[0040] In the model construction and training steps, the adaptive learning rate adjustment algorithm not only uses the Adam algorithm, but also combines the learning rate decay strategy. The initial learning rate is set to 0.001, and after every 10 epochs (iteration cycles), the learning rate decays to 0.9 times the original, that is, ,in For the In this way, a larger learning rate can be used in the early stage of training to speed up the convergence of the model. In the later stage of training, as the model gradually approaches the optimal solution, a smaller learning rate can be used for fine-tuning to further improve the training effect and stability of the model.

[0041] Model evaluation and optimization: The trained models were evaluated using evaluation metrics such as root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R²), and mean absolute percentage error (MAPE).

[0042] The calculation formula is: ; The calculation formula is: ; The calculation formula is: ; The calculation formula is: ; in is the actual value, is the predicted value, is the actual mean value, is the sample size.

[0043] According to the evaluation results, the model parameters are adjusted and the structure is optimized. and , increase the complexity of the model, including adding convolutional layers or The number of units, or adjust the size and number of convolution kernels. or , optimize feature engineering, including adjusting feature combinations or screening more relevant features. Through repeated trials and adjustments, until the model achieves the best performance on the validation set, that is, each evaluation indicator reaches the optimal value or meets the preset threshold requirements. In model evaluation and optimization, in addition to using , , and In addition to the evaluation indicators such as , the relative root mean square error ( ) is used as the evaluation index, and the calculation formula is: , in, is the root mean square error, is the mean of the actual values. By comprehensively considering multiple evaluation indicators of different dimensions, the performance of the model is evaluated more comprehensively, especially the prediction accuracy under different magnitudes of photovoltaic power data. According to the performance of each evaluation indicator, a weighted combination method is used to determine the comprehensive performance score of the model. For example, , , , and The weights are assigned to 0.3, 0.3, 0.2, 0.1 and 0.1 respectively, and then the comprehensive score is calculated. According to the comprehensive score, the model parameters are adjusted and the structure is optimized, such as adjusting the model's layers, number of neurons, learning rate and other hyperparameters, or trying different feature combinations and data preprocessing methods until the model achieves the best performance.

[0044] Real-time prediction: The real-time meteorological data and equipment operation data are preprocessed to make their format and features consistent with the training data. Then, feature extraction and transformation of real-time data are performed according to the steps of feature engineering.

[0045] In order to improve the accuracy and stability of real-time prediction, the real-time collected data is processed by sliding window, and the window size is set to 5 time points (which can be adjusted according to actual conditions), that is, the data of 5 consecutive time points is taken as a window and input into the optimized model for prediction. Then the average value of the predicted value in the window is taken as the final real-time prediction result to smooth the fluctuation of the predicted value and reduce the impact of noise and abnormal data on the prediction result.

[0046] In real-time prediction, in order to deal with sudden abnormal situations in real-time data, such as abnormal data points caused by meteorological sensor failure, an outlier detection and replacement strategy is adopted. First, the deviation degree between the real-time data point and the mean and standard deviation of the historical data in the sliding window is calculated to determine whether it is an outlier. If the deviation degree exceeds 3 times the standard deviation, it is considered an outlier. For outliers, the median of the historical data in the sliding window is used for replacement. Then, the processed real-time data is input into the model for prediction to improve the accuracy and reliability of real-time prediction. At the same time, in order to update the model in time to adapt to the dynamic changes of the photovoltaic power generation system, the latest real-time data is added to the historical data set at regular intervals (such as one week), and the model is retrained and optimized to ensure that the model always maintains good prediction performance.

[0047] Embodiment 2:

[0048] like Figure 2 As shown, this embodiment discloses a photovoltaic power generation high-precision prediction system, which includes a data acquisition and preprocessing module, a feature engineering module, a model building and training module, a model evaluation and optimization module, and a real-time prediction module. The following is a detailed description of each module.

[0049] 1. Data acquisition and preprocessing module Data Collection Extensively collect historical photovoltaic power generation data with a time span of no less than one year, covering information on power generation changes under different seasons, weather conditions and time periods.

[0050] The corresponding meteorological data are collected synchronously, including hourly irradiance, temperature, humidity, wind speed, air pressure and other comprehensive meteorological elements.

[0051] Obtain relevant equipment operation data, such as the model of the photovoltaic panel, installation angle, aging degree, equipment failure rate, maintenance records and other detailed information.

[0052] Data cleaning Through data integrity checks, missing values ​​are identified and processed. For data segments with continuous missing data of 1-2 hours, linear interpolation is used to supplement; for data segments with missing data of 2-3 hours, linear interpolation is used first and then smoothed by moving average method (window size is 3); for data segments with missing data of more than 3 hours, in addition to linear interpolation, similar daily data average (selected and calculated based on weather, season, and date type similarity) is used for correction.

[0053] Outliers are removed based on the 3σ principle, that is, if a data point deviates from the mean by more than 3 times the standard deviation, it is considered an outlier and removed. The mean and standard deviation are calculated based on historical data statistics.

[0054] Data normalization and standardization Use linear normalization method to map the data to the [0,1] interval, the formula is ,in is the original data, is the minimum value in the data set, is the maximum value in the data set, is the normalized data.

[0055] The normalized data is standardized so that its mean is 0 and its standard deviation is 1. The formula is: , in, is the mean, is the standard deviation, The data is standardized to improve data stability and model convergence speed.

[0056] 2. Feature Engineering Module Feature extraction and combination Rich features are extracted from the preprocessed data, including time features (such as hours, seasons, dates, time periods, holidays, etc.), meteorological features (conventional meteorological elements and changing trends), and equipment features (photovoltaic panel details and equipment operating time, maintenance intervals, etc.).

[0057] Construct new combined features such as the ratio of irradiance to temperature and the product of humidity and wind speed to enhance feature expression capabilities.

[0058] Feature processing and screening After one-hot encoding the time features, the principal component analysis (PCA) method is used to reduce the dimension, retaining the principal components with a cumulative contribution rate of more than 90%, reducing the feature dimension and sparsity problems.

[0059] By calculating the correlation coefficients between meteorological features and their combinations (such as constructing a multivariate linear regression model to calculate the regression coefficient and significance level), a feature subset with high correlation and significant impact on photovoltaic power generation is screened out to improve feature validity and prediction accuracy.

[0060] 3. Model building and training module Model architecture design A hybrid model based on deep learning is adopted, which includes convolutional neural network (CNN) and long short-term memory network (LSTM). The CNN part sets the input layer to match the feature vector dimension, connects multiple convolution layers in sequence (for example, the first convolution layer uses 32 3×3 convolution kernels, ReLU activation function, followed by a maximum pooling layer, pooling window 2×2, and the number of convolution kernels is increased in the subsequent convolution layers) and pooling layers, and finally converts the multi-dimensional feature vector into a one-dimensional vector through a flat layer. The LSTM part adjusts the number of neurons according to the data set and feature complexity (determined in the range of [32,512] by grid search and cross-validation), receives the CNN output for time series learning and power prediction, and then connects to a fully connected layer to output the final prediction value.

[0061] Training Mechanism The training process uses the Adam algorithm as the adaptive learning rate adjustment algorithm, and the initial learning rate is set to 0.001. The formula is: , in are model parameters, is the learning rate, and are the first-order and second-order moment estimates of the gradient, The value is 1e-8 to prevent the denominator from being 0. At the same time, combined with the learning rate decay strategy, the learning rate decays to 0.9 times the original value every 10 epochs.

[0062] The early stopping method is used to prevent overfitting, and the patience value is set to 5. When the validation set loss function value does not decrease for 5 consecutive epochs, the training is stopped. The mean square error (MSE) is used as the loss function, and the model parameters are updated through the back propagation algorithm.

[0063] 4. Model evaluation and optimization module Evaluation Index System Multi-dimensional evaluation indicators such as root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R²), mean absolute percentage error (MAPE) and relative root mean square error (RRMSE) are used to comprehensively evaluate the model performance.

[0064] The calculation formula is: ; The calculation formula is: ; The calculation formula is: ; The calculation formula is: ; The calculation formula is: ; in, is the actual value, is the predicted value, is the actual mean value, is the sample size.

[0065] Optimization strategy Adjust the model parameters and optimize the structure according to the evaluation results. If RMSE and MAE are large, consider increasing the model complexity (such as adjusting the number of convolutional layers or LSTM units, the size of the convolution kernel, etc.) or optimizing feature engineering; if R² is low or MAPE and RRMSE are high, further optimize feature selection and combination. Determine the comprehensive performance of the model by weighted combination of evaluation indicators (such as RMSE, MAE, R², MAPE, and RRMSE are weighted by 0.3, 0.3, 0.2, 0.1, and 0.1 to calculate the comprehensive score), and repeat the test and adjustment until the model achieves the best performance in the validation set.

[0066] 5. Real-time prediction module Data processing and prediction The real-time collected meteorological data and equipment operation data are preprocessed (the format and features are consistent with the training data) and feature extracted and transformed.

[0067] A sliding window is used to process real-time data. The window size is set to 5 time points (adjustable). The data in the window is input into the optimized model for prediction. The average of the predicted values ​​in the window is taken as the final real-time prediction result to smooth out fluctuations and improve accuracy and stability.

[0068] Exception handling and model updating By calculating the deviation between the real-time data point and the mean and standard deviation of the historical data in the sliding window (more than 3 times the standard deviation is considered abnormal), outliers are detected and forecasted after replacing the outliers with the median of the historical data in the window. The latest real-time data is added to the historical data set every week, and the model is retrained and optimized to adapt to the dynamic changes of the photovoltaic power generation system and maintain good forecasting performance.

[0069] The above specific embodiments are only several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A high-precision prediction method for photovoltaic power generation, characterized in that: The following steps are involved: Data collection and preprocessing: Collect historical photovoltaic power data, obtain power generation change information under different seasons, weather conditions and time periods, and collect corresponding meteorological data; The collected data were cleaned and the missing values ​​were removed through data integrity check. For data segments with continuous missing values ​​of more than 3 hours, linear interpolation was used to supplement them. The method based on the 3σ principle was used to remove outliers, that is, if the value of a data point deviated from the mean by more than 3 times the standard deviation, it was regarded as an outlier and removed. The mean and standard deviation were obtained by statistical calculation of historical data. Perform data normalization and use linear normalization method to map the data to the [0,1] interval. The specific formula is: ,in is the original data, is the minimum value in the data set, is the maximum value in the data set, is the normalized data; feature project: Extracting features related to photovoltaic power generation from the preprocessed data; The extracted features are combined and transformed, the ratio of irradiance to temperature and the product of humidity and wind speed are used as new combined features, the time features are processed by one-hot encoding, and the categorical variables of different seasons and 0-23h are converted into binary vectors; By calculating the correlation coefficient between meteorological features, a subset of meteorological features with a high correlation with photovoltaic power generation is screened out. The calculation formula of the correlation coefficient is: ， in and The two features are values, and are the means of the two features, is the sample size, is the correlation coefficient, and the correlation threshold is set to 0.

5. Features above this threshold are retained, and features below the threshold are removed; Model building and training: A hybrid model based on deep learning is used for construction, wherein the hybrid model includes a convolutional neural network and a long short-term memory network; The CNN part uses a structure in which multiple convolutional layers and pooling layers are connected alternately. First, an input layer is set, whose shape matches the dimension of the feature vector; The first convolution layer uses 32 convolution kernels of size 3×3 and ReLU activation function for feature extraction and preliminary nonlinear transformation; The maximum pooling layer with a pooling window size of 2×2 is used to reduce the data dimension and extract the main features; Connect two more convolutional layers, using 64 and 128 3×3 kernels respectively, also using the ReLU activation function, and follow the corresponding max pooling layer; The multi-dimensional feature vector output by the convolutional layer is converted into a one-dimensional vector through a flat layer and input into the LSTM part; The LSTM part contains multiple LSTM units, and the number of neurons is adjusted according to the size and feature complexity of the data set: the optimal number of neurons is determined by cross-validation, and the search is performed in the range of [64, 256]. Each LSTM unit receives the output result of CNN as input to perform time series feature learning and power prediction. After the LSTM layer, a fully connected layer is added, and its number of neurons is the same as the dimension of the prediction target. The prediction target is photovoltaic power generation, and a linear activation function is used to output the final prediction value; During the training process, an adaptive learning rate adjustment algorithm and early stopping method are used to prevent overfitting. The Adam algorithm dynamically adjusts the learning rate based on the first-order moment estimate and the second-order moment estimate of the gradient. The calculation formula is: ， in, For the The model parameters at the iteration, is the initial learning rate, set to 0.001, is the first moment estimate of the gradient, is the second-order moment estimate of the gradient, is a very small constant, set to 1e-8, to prevent the denominator from being 0; the early stopping method sets the patience value to 5, that is, when the loss function value on the validation set no longer decreases after 5 consecutive iterations, the training is stopped; the mean square error is used as the loss function during the training process, and the model parameters are updated through the back propagation algorithm; Model evaluation and optimization: Using Root Mean Square Error , mean absolute error , coefficient of determination and mean absolute percentage error Evaluation indicators evaluate the trained model; The calculation formula is: ; The calculation formula is: ; The calculation formula is: ; The calculation formula is: ; in is the actual value, is the predicted value, is the actual mean value, is the sample size; According to the evaluation results, the model parameters are adjusted and the structure is optimized. and , increase the complexity of the model, including adding convolutional layers or The number of units, adjust the size and number of convolution kernels; if or , optimize feature engineering, adjust feature combinations or filter more relevant features; Through repeated trials and adjustments, the model achieves optimal performance on the validation set, that is, each evaluation indicator reaches the optimal value or meets the preset threshold requirements; Real-time prediction: The real-time collected meteorological data and equipment operation data are preprocessed to make their format and features consistent with the training data, and the real-time data are extracted and transformed according to the steps of feature engineering.

2. A high-precision photovoltaic power prediction method according to claim 1, characterized in that: In the data collection and preprocessing steps, for the supplementation of missing data values, when the continuous missing time is between 1-2 hours, the linear interpolation method is used; when the continuous missing time is between 2-3 hours, the linear interpolation method is first used to supplement, and then the supplemented data is smoothed, using the moving average method with a window size of 3 to reduce the error fluctuation caused by interpolation; for data segments with missing time exceeding 3 hours, the linear interpolation method is used to supplement, and correction is performed in combination with the mean of similar day data. The selection of similar days is based on the similarity judgment of weather type, season and date type factors. By calculating the similarity index, the top 5 dates ranked in similarity are selected as similar days, and the data mean of the missing time period corresponding to these similar days is taken to correct the missing values ​​after linear interpolation.

3. A high-precision photovoltaic power prediction method according to claim 2, characterized in that: In the data collection and preprocessing steps, the data normalization process adopts a linear normalization method, and the normalized data is standardized so that its mean is 0 and its standard deviation is 1. The standardization formula is: ， in is the data after linear normalization, is the mean of the normalized data, is the standard deviation of the normalized data, is the standardized data.

4. A high-precision photovoltaic power prediction method according to claim 1, characterized in that: In the feature engineering, after the time features are processed by one-hot encoding, the principal component analysis method is used to reduce the dimension of the encoded time features, and the principal components with a cumulative contribution rate of more than 90% are retained.

5. A high-precision photovoltaic power prediction method according to claim 1, characterized in that: In the model construction and training, in the convolution layer of the CNN part, a batch normalization layer is added after each convolution layer. The calculation formula of the batch normalization layer is: ， ， in is the input data, and are the mean and variance of the batch data, is a very small constant, set to 1e-5 to prevent the denominator from being 0. is the normalized data, and is a learnable parameter, adjusted through training, is the output of the batch normalization layer.

6. A method for high-precision prediction of photovoltaic power generation according to claim 1, characterized in that: In the model construction and training, the number of neurons in the LSTM part is determined by combining grid search and cross-validation. First, a search range for the number of neurons is set, which is [32, 64, 128, 256, 512]. Then, the data set is divided into a training set, a validation set, and a test set. For each setting of the number of neurons, the model is trained on the training set, the model performance is evaluated on the validation set, and the corresponding evaluation index values ​​are recorded. Finally, the number of neurons with the best performance on the validation set is selected as the final parameter of the model.

7. A method for high-precision prediction of photovoltaic power generation according to claim 1, characterized in that: In the model construction and training, the adaptive learning rate adjustment algorithm adopts the Adam algorithm, combined with the learning rate decay strategy, the initial learning rate is set to 0.001, and after every 10 iterations, the learning rate decays to 0.9 times the original, that is, ,in For the The learning rate for each epoch.

8. A method for high-precision prediction of photovoltaic power generation according to claim 1, characterized in that: In the model evaluation and optimization, the , , and In addition to the evaluation indicators, the relative root mean square error is also introduced As an evaluation indicator, the calculation formula is: ， in, is the root mean square error, is the mean of the actual values.

9. A method for high-precision prediction of photovoltaic power generation according to claim 1, characterized in that: In the real-time prediction, abnormal data points caused by meteorological sensor failure adopt an outlier detection and replacement strategy. The degree of deviation between the real-time data point and the mean and standard deviation of the historical data in the sliding window is calculated to determine whether it is an outlier. If the degree of deviation exceeds 3 times the standard deviation, it is regarded as an outlier. For outliers, the median of the historical data in the sliding window is used for replacement, and the processed real-time data is then input into the model for prediction.

10. A high-precision photovoltaic power prediction system, characterized in that: The system includes: Data acquisition and preprocessing module, the data acquisition and preprocessing module includes: The data collection unit collects extensive historical photovoltaic power data, with a time span of no less than one year, covering power generation change information under different seasons, weather conditions and time periods, and simultaneously collects corresponding meteorological data, including hourly irradiance, temperature, humidity, wind speed, and air pressure, and obtains relevant equipment operation data, including the model of photovoltaic panels, installation angle, aging degree, equipment failure rate, and maintenance records; The data cleaning unit identifies and processes missing values ​​through data integrity checks. For data segments that are missing for 1-2 hours in a row, linear interpolation is used to supplement them. For data segments that are missing for 2-3 hours, linear interpolation is used first and then moving average smoothing is used. The moving average window size is 3. For data segments that are missing for more than 3 hours, in addition to linear interpolation, the mean value of similar daily data is used for correction. Outliers are removed based on the 3σ principle, that is, if a data point deviates from the mean by more than 3 times the standard deviation, it is considered an outlier and removed. The mean and standard deviation are calculated based on historical data statistics. Data normalization and standardization unit, use linear normalization method to map data to the [0,1] interval, the formula is ,in is the original data, is the minimum value in the data set, is the maximum value in the data set, The normalized data is standardized so that its mean is 0 and its standard deviation is 1. The formula is: ， in, is the mean, is the standard deviation, is the standardized data; Feature engineering module, the feature engineering module includes: Feature extraction and combination unit, extracts features from preprocessed data, including time features, meteorological features, and equipment features; constructs combined features such as the ratio of irradiance to temperature, and the product of humidity and wind speed; The feature processing and screening unit performs one-hot encoding on the time features, and then uses the principal component analysis method to reduce the dimension, retaining the principal components with a cumulative contribution rate of more than 90%; by calculating the correlation coefficients between meteorological features and between combinations, the feature subset with high correlation and significant impact on photovoltaic power generation is screened out; Model building and training module, the model building and training module includes: The model architecture design unit uses a hybrid model based on deep learning, including convolutional neural networks and long short-term memory networks; Training mechanism unit, the training process uses the Adam algorithm as the adaptive learning rate adjustment algorithm, the initial learning rate is set to 0.001, the formula is: ， in are model parameters, is the learning rate, and are the first-order and second-order moment estimates of the gradient, The denominator is set to 1e-8 to prevent it from being 0. At the same time, combined with the learning rate decay strategy, the learning rate decays to 0.9 times of the original value every 10 epochs. The early stopping method is used to prevent overfitting. The patience value is set to 5. When the loss function value of the validation set does not decrease for 5 consecutive epochs, the training is stopped. The mean square error (MSE) is used as the loss function, and the model parameters are updated through the back propagation algorithm. Model evaluation and optimization module, the model evaluation and optimization module includes: The evaluation index system unit uses root mean square error, mean absolute error, determination coefficient, mean absolute percentage error and relative root mean square error multi-dimensional evaluation indicators to comprehensively evaluate the model performance; The calculation formula is: ; The calculation formula is: ; The calculation formula is: ; The calculation formula is: ; The calculation formula is: ; in, is the actual value, is the predicted value, is the actual mean value, is the sample size; The optimization strategy unit adjusts the parameters and optimizes the structure of the model according to the evaluation results; Real-time prediction module, the real-time prediction module includes: The data processing and prediction unit performs preprocessing and feature extraction transformation on the real-time collected meteorological data and equipment operation data, uses a sliding window to process real-time data, and the window size is set to 5 time points. The data in the window is input into the optimized model for prediction, and the average value of the predicted value in the window is taken as the final real-time prediction result; The exception handling and model updating unit detects outliers by calculating the degree of deviation between the real-time data points and the mean and standard deviation of the historical data in the sliding window, and then replaces the outliers with the median of the historical data in the window before making predictions; the latest real-time data is added to the historical data set every week to retrain and optimize the model.

Images