Short-term photovoltaic power combined prediction method based on feature screening and weight optimization

Abstract

The invention discloses a short-term photovoltaic power combined prediction method based on feature screening and weight optimization, and the method comprises the steps: enabling a photovoltaic power prediction model to comprise three photovoltaic power primary prediction models, and carrying out the training of a training set, the input features of each photovoltaic power primary prediction model are a first type of features and a second type of features, and determining the first type of features and the second type of features of each photovoltaic power primary prediction model through the training set; and inputting the first type of features and the second type of features corresponding to the test sample into each photovoltaic power primary prediction model in the trained photovoltaic power prediction models, so that the photovoltaic power prediction models perform prediction to obtain a final photovoltaic power prediction value. According to the method, feature screening is carried out by using the reconstructed data, input features of different photovoltaic power primary prediction models are obtained, the effectiveness and pertinence of the input features are ensured, the problems of feature redundancy, noise interference and the like are reduced, and the accuracy and robustness of prediction are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power prediction in electric power systems, and in particular relates to a short-term photovoltaic power combination prediction method based on feature screening and weight optimization. Background Art

[0002] Against the backdrop of global energy transformation and sustainable development, photovoltaic power generation, as a clean, renewable energy source, is becoming a vital component of power systems. The output characteristics of photovoltaic power are complex and variable, susceptible to factors such as weather conditions, irradiance, and temperature. These characteristics exhibit intermittent and random fluctuations, making the grid's supply and demand balance unstable and increasing the difficulty of regulation. Therefore, accurate photovoltaic power forecasting is crucial for optimizing energy allocation and ensuring stable grid operation.

[0003] Currently, the commonly used method in photovoltaic power prediction technology is the combined prediction method. The combined prediction method was proposed to cope with the complex fluctuation characteristics of photovoltaic power. The combined prediction method combines the advantages of data preprocessing technology, multiple prediction models and feature screening methods, aiming to give full play to their respective advantages and improve prediction accuracy. In order to reduce the random volatility of photovoltaic power, the combined prediction method usually uses signal decomposition methods such as wavelet decomposition (WT), variational mode decomposition (VMD), and complete ensemble empirical mode decomposition of adaptive white noise (CEEMDAN) to decompose the original sequence into several intrinsic mode functions (IMFs), predict them separately and superimpose the prediction results.

[0004] "Chen Qingbin, Yang Genghuang, Geng Liqing, et al. Short-term PV power portfolio forecasting based on similar day selection and data reconstruction [J / OL]. China Electric Power, 1-11 [2024-11-25]." This paper clusters PV power using KFCM, constructs similar day sets based on cooperative game theory, and decomposes and reconstructs PV power using VMD. High-frequency terms are predicted using a CNN-BiLSTM-Attention model, while trend terms and low-frequency terms are predicted using an LSTM model. Experimental results show that under clear weather conditions, the model achieves a MAE of 1.768, an RMSE of 2.357, and a MAPE (mean absolute percentage error) of 0.073; under cloudy weather conditions, the MAE is 4.380, the RMSE is 5.271, and the MAPE is 0.215; and under rainy weather conditions, the proposed model achieves a MAE of 4.837, an RMSE of 6.084, and a MAPE of 0.313. The above evaluation indicators clearly show that the prediction accuracy for sunny days is the best, followed by cloudy days, and the lowest for rainy days. This is because MAPE focuses on reflecting the average level of relative error between the predicted value and the true value, which can intuitively reflect the degree of deviation of the predicted value from the true value.

[0005] The existing combined forecasting methods have greatly improved the forecasting accuracy, but the following key issues still exist:

[0006] (1) Insufficient correlation between signal decomposition and meteorological characteristics: Most existing methods directly decompose photovoltaic power, but rarely consider the correlation between meteorological characteristics and the wavelet components of photovoltaic power. As a result, the correlation between the reconstructed high-frequency items and meteorological characteristics is significantly reduced, and the prediction error of high-frequency items increases, which in turn affects the overall prediction accuracy.

[0007] (2) The feature screening does not match the correlation between each frequency item: When screening input features, existing methods usually perform the screening before photovoltaic power decomposition, so that the selected features have a high correlation with the photovoltaic power before decomposition, but cannot guarantee the same correlation with the reconstructed high-frequency items, medium-frequency items and low-frequency items. As a result, the corresponding input features of each reconstructed frequency item contain some features with weak correlation or irrelevant features, which interferes with the accuracy and stability of the prediction model.

[0008] (3) Irrational model selection and weight distribution: When allocating the weights of the prediction results of each frequency item, the existing methods are often based on simple addition operations, lacking comprehensive consideration of the prediction errors, volatility and contribution of each frequency item to the overall prediction, resulting in irrational weight distribution, causing the frequency items with larger prediction errors to have an excessive impact on the final results. Summary of the Invention

[0009] In view of the shortcomings of the prior art, the present invention aims to provide a photovoltaic power prediction model.

[0010] The purpose of the present invention is to provide a training method for a photovoltaic power prediction model.

[0011] Another object of the present invention is to provide a short-term photovoltaic power combination prediction method based on feature screening and weight optimization.

[0012] The present invention is achieved through the following technical solutions.

[0013] A photovoltaic power prediction model includes: three photovoltaic power primary prediction models, the photovoltaic power prediction models are trained using a training set, the training samples in the training set belong to the same type of weather, the input features of each photovoltaic power primary prediction model are first-category features and second-category features, and the first-category features and second-category features of each photovoltaic power primary prediction model are determined using the training set;

[0014] The selection rules for the first type of features include 1-1) to 1-3)

[0015] 1-1) Obtaining reconstructed data. Methods for obtaining reconstructed data include 1-1-1) to 1-1-3):

[0016] 1-1-1) Taking the photovoltaic power as the sample true value, arranging the sample true values ​​corresponding to all training samples in the training set according to the acquisition time of the sample true value, obtaining the photovoltaic power curve corresponding to the training set, and decomposing the photovoltaic power curve to obtain M intrinsic mode functions;

[0017] 1-1-2) Calculate the sample entropy value corresponding to each intrinsic mode function, cluster the M sample entropy values, and obtain three clusters as clustering results. The three clusters are divided into high-frequency items, medium-frequency items, and low-frequency items according to the average entropy values ​​of the samples in the clusters from high to low;

[0018] 1-1-3) superimposing the intrinsic mode functions corresponding to the sample entropy values ​​belonging to the high-frequency items to obtain the reconstructed data of the high-frequency items; superimposing the intrinsic mode functions corresponding to the sample entropy values ​​belonging to the medium-frequency items to obtain the reconstructed data of the medium-frequency items; superimposing the intrinsic mode functions corresponding to the sample entropy values ​​belonging to the low-frequency items to obtain the reconstructed data of the low-frequency items; using the high-frequency item reconstructed data, the medium-frequency item reconstructed data and the low-frequency item reconstructed data as the reconstructed data respectively; making the high-frequency item reconstructed data correspond to the first photovoltaic power primary prediction model, the medium-frequency item reconstructed data correspond to the second photovoltaic power primary prediction model, and the low-frequency item reconstructed data correspond to the third photovoltaic power primary prediction model;

[0019] 1-2) Obtain a set of correlation coefficients E1 between N irradiance wavelet components and each set of reconstructed data;

[0020] 1-3) Take the largest correlation coefficient in the correlation coefficient set E1 as R max Calculate the correlation coefficient between the irradiance curve corresponding to the training set and each set of reconstructed data and use it as R i , if the R of the reconstructed data corresponding to the same photovoltaic power primary prediction model is i Greater than R max , the photovoltaic power primary prediction model selects irradiance as the first-category feature; otherwise, the photovoltaic power primary prediction model selects the frequency value of the irradiance wavelet component corresponding to the correlation coefficient with an absolute value greater than 0.2 in the corresponding correlation coefficient set E1 as the first-category feature;

[0021] The selection rules for the second type of features include: obtaining a set E2 of correlation coefficients between eight types of meteorological data and each set of reconstructed data; the photovoltaic power primary prediction model selects the meteorological data corresponding to the correlation coefficient with an absolute value greater than 0.2 in the set E2 of correlation coefficients corresponding to the reconstructed data as the second type of features.

[0022] In the above technical solution, the first photovoltaic power primary prediction model includes: a normalization module, a CNN-BiLSTM-Attention neural network and a denormalization module connected in sequence.

[0023] In the above technical solution, the CNN-BiLSTM-Attention neural network includes: a CNN feature extraction module, a Dropout layer, a BiLSTM information capture module, a Dropout layer, an attention mechanism, a fully connected layer, and an output layer connected in sequence.

[0024] In the above technical solution, the second photovoltaic power primary prediction model and the third photovoltaic power primary prediction model both include: a normalization module, an LSTM neural network and an anti-normalization module connected in sequence.

[0025] In the above technical solution, the photovoltaic power prediction model also includes: a weighting module, the three photovoltaic power primary prediction models are the first photovoltaic power primary prediction model, the second photovoltaic power primary prediction model and the third photovoltaic power primary prediction model, the output of the first photovoltaic power primary prediction model is the first photovoltaic power preliminary prediction value, the output of the second photovoltaic power primary prediction model is the second photovoltaic power preliminary prediction value, the output of the third photovoltaic power primary prediction model is the third photovoltaic power preliminary prediction value, the weights corresponding to the first photovoltaic power preliminary prediction value, the second photovoltaic power preliminary prediction value and the third photovoltaic power preliminary prediction value are weight A', weight B' and weight C' respectively; the input of the weighting module is the first photovoltaic power preliminary prediction value, the second photovoltaic power preliminary prediction value, the third photovoltaic power preliminary prediction value, weight A', weight B', weight C' and the constant term d', and the output of the weighting module is the final photovoltaic power prediction value.

[0026] In the above technical solution, weights A', B' and C' are optimized using the dung beetle algorithm.

[0027] In the above technical solution, the method for obtaining the correlation coefficient set E1 between N irradiance wavelet components and each group of reconstructed data is: arrange the irradiances in all training samples in the training set in sequence according to the acquisition time to obtain the irradiance curve corresponding to the training set; decompose the irradiance curve to obtain N irradiance wavelet components, calculate the correlation coefficient between each irradiance wavelet component and the group of reconstructed data, and obtain N correlation coefficients between the group of reconstructed data and the N irradiance wavelet components. The N correlation coefficients constitute the correlation coefficient set E1.

[0028] In the above technical solution, the eight types of meteorological data include: air temperature, relative humidity, ground air pressure, cloud opacity, dew point temperature, atmospheric precipitable water, wind direction at a height of 10 meters (expressed in angles), and wind speed at a height of 10 meters.

[0029] In the above technical solution, the method for obtaining the correlation coefficient set E2 between the eight types of meteorological data and each group of reconstructed data is: arrange each type of meteorological data of all training samples in the training set in sequence according to the collection time to obtain a meteorological data curve; a total of eight meteorological data curves are obtained for the eight types of meteorological data; calculate the correlation coefficients between the group of reconstructed data and the eight meteorological data curves respectively and form the correlation coefficient set E2.

[0030] In the above technical solution, the fitness function of the dung beetle algorithm is the mean absolute percentage error between the final predicted value of photovoltaic power and the true value of the sample.

[0031] A training method for a photovoltaic power prediction model comprises: determining first-category features and second-category features of each primary photovoltaic power prediction model according to a selection rule and a training set, inputting the first-category features and second-category features into the primary photovoltaic power prediction model, training the photovoltaic power prediction model, and obtaining a trained photovoltaic power prediction model.

[0032] A short-term photovoltaic power combination prediction method based on feature screening and weight optimization includes: inputting first-category features and second-category features corresponding to a test sample into each photovoltaic power primary prediction model in a trained photovoltaic power prediction model, enabling the photovoltaic power prediction model to perform prediction and obtain a final photovoltaic power prediction value.

[0033] The present invention has the following advantages due to the adoption of the above technical solution:

[0034] 1. The present invention uses a wavelet decomposition algorithm to decompose the irradiance curve. The irradiance wavelet component after the decomposition of the irradiance curve can reflect the instantaneous changes and fluctuations of the irradiance curve. The decomposed irradiance wavelet component is used as the input feature selection range of the photovoltaic power primary prediction model, providing more accurate and reasonable input features and improving the prediction accuracy.

[0035] 2. The present invention uses reconstructed data for feature screening to obtain input features for different photovoltaic power primary prediction models, ensuring the validity and pertinence of the input features, reducing problems such as feature redundancy and noise interference, and improving the accuracy and robustness of the prediction.

[0036] 3. The present invention uses the dung beetle optimization algorithm to find the optimal weight and calculates the final predicted value of photovoltaic power based on the optimal weight, realizing fine adjustment of the weights of different frequency items, solving the problem of unreasonable weight distribution in traditional methods, and significantly improving the prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a structural diagram of the photovoltaic power prediction model of the present invention;

[0038] Figure 2(a) shows the cluster samples corresponding to sunny days;

[0039] Figure 2(b) shows the cluster samples corresponding to cloudy conditions;

[0040] Figure 2(c) shows the cluster samples corresponding to rainy days;

[0041] Figure 3 is the irradiance curve and irradiance wavelet component of the training set in Example 4;

[0042] Figure 4 are the 11 eigenmode functions in Example 4;

[0043] Figure 5 The high-frequency item reconstructed data, the medium-frequency item reconstructed data, and the low-frequency item reconstructed data in Example 4;

[0044] FIG6( a ) is a graph showing the error between the final predicted values ​​of photovoltaic power and the true values ​​of samples for Comparative Examples 1-1, 2-1, 3-1, 4-1, and 5-1;

[0045] FIG6( b ) is a graph showing the error between the final predicted values ​​of photovoltaic power and the true values ​​of the samples for Comparative Examples 1-2, 2-2, 3-2, 4-2, and 5-2;

[0046] FIG6( c ) is a graph showing the error between the final predicted values ​​of photovoltaic power and the true values ​​of the samples for Comparative Examples 1-3, 2-3, 3-3, 4-3, and 5-3;

[0047] FIG7( a ) is a graph showing the error between the final predicted value of photovoltaic power and the true value of the samples in Example 7, Comparative Example 6-1, Comparative Example 7-1, Comparative Example 8-1, and Comparative Example 9-1;

[0048] FIG7( b ) is a graph showing the error between the final predicted value of photovoltaic power and the true value of the samples in Example 8, Comparative Example 6-2, Comparative Example 7-2, Comparative Example 8-2, and Comparative Example 9-2;

[0049] FIG7( c ) is a graph showing the error between the final predicted value of photovoltaic power and the true value of the samples in Example 9, Comparative Example 6-3, Comparative Example 7-3, Comparative Example 8-3, and Comparative Example 9-3; DETAILED DESCRIPTION

[0050] The short-term photovoltaic power combination prediction method based on feature screening and weight optimization of the present invention is described in detail below with reference to the accompanying drawings.

[0051] Example 1

[0052] like Figure 1As shown, a photovoltaic power prediction model includes: three photovoltaic power primary prediction models and a weighting module, the three photovoltaic power primary prediction models are a first photovoltaic power primary prediction model, a second photovoltaic power primary prediction model and a third photovoltaic power primary prediction model, the output of the first photovoltaic power primary prediction model is a first photovoltaic power preliminary prediction value, the output of the second photovoltaic power primary prediction model is a second photovoltaic power preliminary prediction value, and the output of the third photovoltaic power primary prediction model is a third photovoltaic power preliminary prediction value, and the input features of each photovoltaic power primary prediction model are first-category features and second-category features;

[0053] The weights corresponding to the first preliminary photovoltaic power prediction value, the second preliminary photovoltaic power prediction value and the third preliminary photovoltaic power prediction value are weight A', weight B' and weight C' respectively; the input of the weighting module is the first preliminary photovoltaic power prediction value, the second preliminary photovoltaic power prediction value, the third preliminary photovoltaic power prediction value, weight A', weight B', weight C' and the constant term d', the output of the weighting module is the final photovoltaic power prediction value, and the weighting module is used for weighting.

[0054] The photovoltaic power prediction model is trained through a training set. The training samples in the training set belong to the same type of weather. The first-class features and second-class features of each photovoltaic power primary prediction model are determined through the training set. The trained photovoltaic power prediction model can predict the test samples by inputting the first-class features and second-class features of the test samples of this type of weather.

[0055] The selection rules for the first type of features include 1-1) to 1-3);

[0056] 1-1) Obtaining reconstructed data. Methods for obtaining reconstructed data include 1-1-1) to 1-1-3):

[0057] 1-1-1) Taking photovoltaic power as the sample true value, arranging the sample true values ​​corresponding to all training samples in the training set according to the acquisition time of the sample true value, obtaining the photovoltaic power curve corresponding to the training set, and using the CEEMDAN algorithm (Wang Dewen, Jiao Tianyuan. Photovoltaic power prediction under different solar radiation based on quadratic decomposition [J]. Acta Energiae Solaris Sinica, 2024, 45(09): 360-368.) to decompose the photovoltaic power curve to obtain M intrinsic mode functions; using the CEEMDAN algorithm to decompose the sample true value effectively reduces the interference of residual noise in the intrinsic mode function, improves the problems of modal aliasing and pseudo-mode, overcomes the sensitivity of the VMD decomposition algorithm in the existing technology when selecting the number of modes, and accurately captures the high-frequency components and transient characteristics in the sample true value.

[0058] 1-1-2) Calculate the sample entropy value corresponding to each intrinsic mode function (Chen Shiqi, Wu Yan, Yang Dechang, et al. Short-term load forecasting of power system based on secondary load decomposition and feature processing [J / OL]. High Voltage Technology, 1-18 [2025-05-07]. https: / / doi.org / 10.13336 / j.1003-6520.hve.20241311). Use the K-means algorithm to cluster the M sample entropy values ​​and obtain three clusters as clustering results. The three clusters are divided into high-frequency items, medium-frequency items, and low-frequency items according to the average value of the sample entropy values ​​in the clusters from high to low.

[0059] 1-1-3) superimposing the intrinsic mode functions corresponding to the sample entropy values ​​belonging to the high-frequency items to obtain the reconstructed data of the high-frequency items; superimposing the intrinsic mode functions corresponding to the sample entropy values ​​belonging to the medium-frequency items to obtain the reconstructed data of the medium-frequency items; superimposing the intrinsic mode functions corresponding to the sample entropy values ​​belonging to the low-frequency items to obtain the reconstructed data of the low-frequency items; using the high-frequency item reconstructed data, the medium-frequency item reconstructed data and the low-frequency item reconstructed data as the reconstructed data respectively; making the high-frequency item reconstructed data correspond to the first photovoltaic power primary prediction model, the medium-frequency item reconstructed data correspond to the second photovoltaic power primary prediction model, and the low-frequency item reconstructed data correspond to the third photovoltaic power primary prediction model;

[0060] 1-2) Obtain a set E1 of correlation coefficients between N irradiance wavelet components and each set of reconstructed data. The method for obtaining the set E1 of correlation coefficients between N irradiance wavelet components and each set of reconstructed data is as follows: arrange the irradiances of all training samples in the training set in sequence according to the acquisition time to obtain the irradiance curve corresponding to the training set; use the wavelet decomposition algorithm (Wang Shibai, Wang Yuejiao, Wang Nan, et al. Photovoltaic power prediction based on wavelet decomposition and neural network [J]. Shandong Electric Power Technology, 2021, 48(05): 1-7.) to decompose the irradiance curve to obtain N irradiance wavelet components. The N irradiance wavelet components are irradiance wavelet component IMF1, irradiance wavelet component IMF2, ..., irradiance wavelet component IMF N ; The Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficient between each irradiance wavelet component and the set of reconstructed data. The set of reconstructed data and N irradiance wavelet components obtain N correlation coefficients, and the N correlation coefficients constitute the correlation coefficient set E1;

[0061] The wavelet decomposition algorithm not only retains the time correlation characteristics of the irradiance curve corresponding to the training set, but also converts the randomness and volatility of the irradiance of the training samples into high-frequency wavelet component data;

[0062] 1-3) Take the largest correlation coefficient in the correlation coefficient set E1 as R max Calculate the correlation coefficient between the irradiance curve corresponding to the training set and each set of reconstructed data and use it as Ri , if the R of the reconstructed data corresponding to the same photovoltaic power primary prediction model is i Greater than R max , the photovoltaic power primary prediction model selects irradiance as the first-category feature; otherwise, the photovoltaic power primary prediction model selects the frequency value of the irradiance wavelet component corresponding to the correlation coefficient with an absolute value greater than 0.2 in the corresponding correlation coefficient set E1 as the first-category feature;

[0063] The selection rules for the second type of features include 2-1) to 2-3);

[0064] 2-1) Obtain eight types of meteorological data corresponding to the training sample at time t, the eight meteorological data including: air temperature, relative humidity, ground pressure, cloud opacity, dew point temperature, atmospheric precipitable water, wind direction at 10 meters above sea level, and wind speed at 10 meters above sea level;

[0065] 2-2) Obtaining a set of correlation coefficients E2 between the eight types of meteorological data and each set of reconstructed data. The method for obtaining the set of correlation coefficients E2 between the eight types of meteorological data and each set of reconstructed data is as follows: arranging each type of meteorological data of all training samples in the training set in sequence according to the acquisition time to obtain a meteorological data curve; a total of eight meteorological data curves are obtained for the eight types of meteorological data; using the Spearman rank correlation coefficient algorithm to calculate the correlation coefficients between the set of reconstructed data and the eight meteorological data curves respectively and form the correlation coefficient set E2; photovoltaic power (sample true value) is affected by multiple nonlinear factors, and the Spearman rank correlation coefficient algorithm can measure the monotonic relationship between the ranks of two variables and has advantages in processing nonlinear data;

[0066] 2-3) The primary photovoltaic power prediction model selects the meteorological data corresponding to the correlation coefficient set E2 of the corresponding reconstructed data with an absolute value greater than 0.2 as the second type of features;

[0067] As a preference, weights A', B' and C' are optimized using the dung beetle algorithm (Yang Fan, Jiang Song, Huang Guangqiu, et al. VOCs concentration interval prediction based on data decomposition and weight optimization [J / OL]. Journal of Environmental Sciences, 1-12 [2024-11-27]). The input of the dung beetle algorithm is: the first photovoltaic power preliminary prediction value, the second photovoltaic power preliminary prediction value and the third photovoltaic power preliminary prediction value and the sample true value. The output of the dung beetle algorithm is the constant term d', weight A', weight B' and weight C'.

[0068] As a preferred method, the fitness function of the dung beetle algorithm is the mean absolute percentage error (MAPE) between the final predicted value of photovoltaic power and the true value of the sample, and the calculation formula is as follows:

[0069]

[0070] Where y(i) is the true value of the i-th training sample; i = 1, 2...n, n is the number of training samples in the training set; y′(i) is the weighted integration result of the first, second, and third photovoltaic power preliminary prediction values ​​in the dung beetle algorithm calculation process. The calculation formula of y′(i) is:

[0071] y′(i)=A×X1+B×X2+C×X3+d

[0072] Where A, B, and C are the weights corresponding to the first, second, and third photovoltaic power preliminary prediction values, respectively, in the calculation process of the dung beetle algorithm; X1, X2, and X3 are the first, second, and third photovoltaic power preliminary prediction values, respectively; and d is a constant term in the calculation process of the dung beetle algorithm.

[0073] In this embodiment, the first photovoltaic power primary prediction model includes: a normalization module, a CNN-BiLSTM-Attention neural network (Ouyang Fulian, Wang Jun, Zhou Hangxia. Short-term power load forecasting method based on improved transfer learning and multi-scale CNN-BiLSTM-Attention [J]. Power System Protection and Control, 2023, 51(02): 132-140.) and a denormalization module connected in sequence, and the CNN-BiLSTM-Attention neural network includes a CNN feature extraction module, a Dropout layer, a BiLSTM information capture module, a Dropout layer, an attention mechanism (Attention), a fully connected layer and an output layer connected in sequence, wherein the CNN feature extraction module adopts two convolutional layers, the convolution kernel size of each convolutional layer is 3, the number of channels of the two convolutional layers is 16 and 32 respectively, the number of hidden neurons of the BiLSTM information capture module is 25, the fully connected layer uses a Relu activation function, and the dropout rate of each Dropout layer is 0.1;

[0074] The second photovoltaic power primary prediction model and the third photovoltaic power primary prediction model both include: a normalization module, an LSTM neural network (Qin Yu, Xu Ye, Wang Xinpeng, et al. Research on short-term prediction of photovoltaic output based on improved FCM-LSTM [J]. Acta Energiae Solaris Sinica, 2024, 45(08): 304-313.) and an anti-normalization module connected in sequence. The LSTM layer uses the Relu activation function, and the number of hidden neurons in the LSTM layer is 25.

[0075] The normalization module performs normalization processing. The calculation formula for normalization processing is as follows:

[0076]

[0077] Where, X' is the value after normalization; X is the value before normalization; X max and X min These are the maximum and minimum values ​​before normalization. Normalization ensures that the dimensions of each feature input into the primary photovoltaic power prediction model are the same, improving prediction accuracy and accelerating gradient descent.

[0078] The weight search range of the dung beetle algorithm is set to [-20, 20], the population size is 20, and the number of iterations is 200.

[0079] Example 2

[0080] One method of obtaining a sample includes:

[0081] S1, collect eight types of meteorological data, irradiance (total horizontal radiation), and photovoltaic power of photovoltaic power stations in Ningxia during the mission period (mission period: May 1, 2020 - June 30, 2020) to obtain the original photovoltaic power data set. The daily collection period is from 8:30 am to 8:15 pm. Collection is carried out every 15 minutes during the daily collection period (a total of 48 moments per day). Each time, the meteorological data, irradiance, and photovoltaic power at the current moment are collected. The eight meteorological data are air temperature, relative humidity, ground pressure, cloud opacity, dew point temperature, atmospheric precipitable water, wind direction at 10 meters above sea level, and wind speed at 10 meters above sea level.

[0082] S2, remove outliers and fill missing values ​​in the meteorological data, irradiance, and sample true values ​​in the original photovoltaic power dataset to obtain the photovoltaic power dataset. The specific steps are as follows:

[0083] S2-1, the 3σ principle (Chen Qingbin, Yang Genghuang, Geng Liqing, et al. Short-term photovoltaic power combination forecast based on similar day selection and data reconstruction [J / OL]. China Electric Power, 1-11 [2024-11-25].) is used to filter outliers in the original photovoltaic power data set, delete outliers and treat them as missing values;

[0084] S2-2, fill the missing values ​​of the original photovoltaic power data set (Yang Yadong, Geng Liqing, Yang Genghuang, et al. Short-term power load forecasting method based on improved PSO-BP algorithm [J]. Journal of Tianjin Vocational and Technical Normal University, 2024, 34(03): 15-20.) to obtain the photovoltaic power data set.

[0085] S3, including S3-1 and S3-2, is as follows:

[0086] S3-1, in order to classify the weather of samples in the photovoltaic power dataset, all the true values ​​of the samples collected daily are taken as a cluster sample, and the K-means algorithm is used to perform sequence clustering on all cluster samples of the photovoltaic power dataset, obtaining three clusters as clustering results. The number of classification clusters k in the K-means algorithm is set to 3. The cluster with the smallest fluctuation of cluster samples among the three clusters is defined as sunny day, the cluster with the largest fluctuation of cluster samples among the three clusters is defined as rainy day, and the last remaining cluster among the three clusters is defined as cloudy. The cluster samples in the three clusters are respectively as follows: Figure 2(a) to Figure 2(c) As shown, Figure 2(a) to Figure 2(c) Each curve is composed of 48 true values ​​of samples collected in one day and arranged according to the collection time;

[0087] S3-2: In each cluster, the samples corresponding to the cluster samples with the latest collection date are used as test samples (i.e., 48 test samples each for sunny, cloudy, and rainy days), and the samples corresponding to the remaining cluster samples in the cluster are used as training samples (i.e., a total of 58 days * 48 training samples for sunny, cloudy, and rainy days). All training samples in each cluster constitute a training set, i.e., one training set each for sunny, cloudy, and rainy days.

[0088] Example 3

[0089] The training method of the photovoltaic power prediction model in Example 1 includes:

[0090] Determine the first and second features of each primary photovoltaic power prediction model according to the selection rule and a training set in Example 2, and input the first and second features into the primary photovoltaic power prediction model to train the photovoltaic power prediction model, thereby obtaining a trained photovoltaic power prediction model;

[0091] Among them, the training parameters of the first photovoltaic power primary prediction model (CNN-BiLSTM-Attention neural network), the second photovoltaic power primary prediction model and the third photovoltaic power primary prediction model (LSTM neural network) are all set as follows: the optimizer adopts the Adam optimizer, the maximum number of training times is 1000, the gradient threshold is 1, the initial learning rate is 0.001, the learning rate is automatically adjusted when the number of training times reaches 100, the learning rate adjustment factor is 0.001, and the regularization parameter is 0.001.

[0092] Example 4

[0093] The training set corresponds to sunny days, N = 5, the irradiance curves of all training samples in the training set and the five irradiance wavelet components after the irradiance curve is decomposed are as follows: Figure 3 As shown, the horizontal axis is the collection time of the training sample. Figure 3 The "raw data" in is the irradiance curve. Figure 3IMF1, IMF2, IMF3, IMF4 and IMF5 are the five irradiance wavelet components;

[0094] M=11, 11 eigenmode functions such as Figure 4 As shown, the horizontal axis is the acquisition time of the true value of the sample, Figure 4 The “raw data” in the figure is the photovoltaic power curve corresponding to the training set. Figure 4 "IMF1~IMF11" in the figure are 11 intrinsic mode functions;

[0095] The high-frequency item reconstruction data, the medium-frequency item reconstruction data and the low-frequency item reconstruction data are as follows Figure 5 As shown, Figure 5 The "high frequency items" in are the high frequency item reconstruction data. Figure 5 The "intermediate frequency term" is the intermediate frequency term reconstruction data, Figure 5 The "low-frequency item" is the reconstructed data of the low-frequency item.

[0096] Through the training set in this embodiment, it is determined that the first and second features of the first photovoltaic power primary prediction model are: cloud opacity, the frequency value of the irradiance wavelet component IMF2, and the frequency value of the irradiance wavelet component IMF3; the first and second features of the second photovoltaic power primary prediction model are: air temperature, cloud opacity, relative humidity, and irradiance; the first and second features of the third photovoltaic power primary prediction model are: air temperature, cloud opacity, relative humidity, and irradiance;

[0097] Example 5

[0098] When the training set corresponds to cloudy weather, M=11.

[0099] Through the training set in this embodiment, it is determined that the first and second characteristics of the first photovoltaic power primary prediction model are: the frequency value of the irradiance wavelet component IMF1, the frequency value of the irradiance wavelet component IMF2, and the cloud opacity; the first and second characteristics of the second photovoltaic power primary prediction model are: the frequency value of the irradiance wavelet component IMF1, the frequency value of the irradiance wavelet component IMF2, the frequency value of the irradiance wavelet component IMF3, the frequency value of the irradiance wavelet component IMF, and cloud opacity; the first and second characteristics of the third photovoltaic power primary prediction model are: irradiance, air temperature, and cloud opacity.

[0100] Example 6

[0101] The training set corresponds to rainy days, M=9.

[0102] The first and second characteristics of the first PV power primary prediction model are: cloud opacity, frequency value of irradiance wavelet component IMF1, frequency value of irradiance wavelet component IMF2, and frequency value of irradiance wavelet component IMF3; the first and second characteristics of the second PV power primary prediction model are: irradiance, air temperature, relative humidity, and cloud opacity; the first and second characteristics of the third PV power primary prediction model are: irradiance, air temperature, relative humidity, ground pressure, cloud opacity, atmospheric precipitable water, and wind speed at an altitude of 10 meters.

[0103] The sample entropy values ​​corresponding to the M intrinsic mode functions in Examples 4 to 6 and the clustering results in 1-1-2) are shown in Table 1;

[0104] Table 1

[0105]

[0106]

[0107] The N correlation coefficients obtained between each set of reconstructed data and the N irradiance wavelet components, the correlation coefficient between each set of reconstructed data and each meteorological data curve, and the correlation coefficient between each set of reconstructed data and the irradiance curve are shown in Table 2. As shown in Table 2, the correlation coefficient between each set of reconstructed data and the irradiance curve is each column corresponding to A1 in Table 2, the N correlation coefficients obtained between each set of reconstructed data and the N irradiance wavelet components are each column corresponding to A2 in Table 2, and the correlation coefficient between each set of reconstructed data and each meteorological data curve is each column corresponding to A3 in Table 2.

[0108] Table 2

[0109]

[0110]

[0111]

[0112] Example 7

[0113] A short-term photovoltaic power portfolio forecasting method based on feature screening and weight optimization includes:

[0114] The corresponding first-category features and second-category features are input into each photovoltaic power primary prediction model in the photovoltaic power prediction model trained in Example 4, so that the photovoltaic power prediction model performs prediction to obtain the final photovoltaic power prediction value; wherein the test sample corresponds to a sunny day.

[0115] Example 8

[0116] A short-term photovoltaic power portfolio forecasting method based on feature screening and weight optimization includes:

[0117] The corresponding first-category features and second-category features are input into each photovoltaic power primary prediction model in the photovoltaic power prediction model trained in Example 5, so that the photovoltaic power prediction model performs prediction to obtain the final photovoltaic power prediction value; wherein the test sample corresponds to cloudy weather.

[0118] Example 9

[0119] A short-term photovoltaic power portfolio forecasting method based on feature screening and weight optimization includes:

[0120] The corresponding first-category features and second-category features are input into each photovoltaic power primary prediction model in the photovoltaic power prediction model trained in Example 6, so that the photovoltaic power prediction model performs prediction to obtain the final photovoltaic power prediction value; wherein the test sample corresponds to a rainy day.

[0121] Comparative Example 1-1

[0122] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 7, with the only difference being:

[0123] 1. The input features of the three primary PV power prediction models are the same and are both first-category features and second-category features. The first-category feature is always irradiance. The second-category feature: The Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficients between the eight meteorological data curves and the PV power curves corresponding to the training set to form a correlation coefficient set E3. The meteorological data corresponding to the correlation coefficients with an absolute value greater than 0.2 in the correlation coefficient set E3 are selected as the second-category features.

[0124] 2. Set weight A', weight B', and weight C' to 1, and the constant term d to 0.

[0125] Comparative Example 1-2

[0126] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 8, with the only difference being:

[0127] 1. The input features of the three primary PV power prediction models are the same and are both first-category features and second-category features. The first-category feature is always irradiance. The second-category feature: The Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficients between the eight meteorological data curves and the PV power curves corresponding to the training set to form a correlation coefficient set E3. The meteorological data corresponding to the correlation coefficients with an absolute value greater than 0.2 in the correlation coefficient set E3 are selected as the second-category features.

[0128] 2. Set weight A', weight B', and weight C' to 1, and the constant term d to 0.

[0129] Comparative Examples 1-3

[0130] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 9, with the only difference being:

[0131] 1. The input features of the three primary PV power prediction models are the same and are both first-category features and second-category features. The first-category feature is always irradiance. The second-category feature: The Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficients between the eight meteorological data curves and the PV power curves corresponding to the training set to form a correlation coefficient set E3. The meteorological data corresponding to the correlation coefficients with an absolute value greater than 0.2 in the correlation coefficient set E3 are selected as the second-category features.

[0132] 2. Set weight A', weight B', and weight C' to 1, and the constant term d to 0.

[0133] Comparative Example 2-1

[0134] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 7, with the only difference being:

[0135] 1. The input features of the three primary PV power prediction models are the same and are all first-category features and second-category features;

[0136] The first type of feature: the irradiance of all training samples in the training set is arranged in sequence according to the acquisition time to obtain the irradiance curve corresponding to the training set; the irradiance curve is decomposed using the wavelet decomposition algorithm to obtain N irradiance wavelet components, and the Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficient between each irradiance wavelet component and the photovoltaic power curve corresponding to the training set. The photovoltaic power curve corresponding to the training set and the N irradiance wavelet components obtain N correlation coefficients, and the N correlation coefficients constitute the correlation coefficient set E4; the frequency value of the irradiance wavelet component corresponding to the correlation coefficient with an absolute value greater than 0.2 in the correlation coefficient set E4 is used as the first type of feature;

[0137] The second type of features: The Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficients between the eight meteorological data curves and the photovoltaic power curves corresponding to the training set and form a correlation coefficient set E3; the meteorological data corresponding to the correlation coefficients with absolute values ​​greater than 0.2 in the correlation coefficient set E3 are selected as the second type of features.

[0138] 2. Weight A', weight B', and weight C' are all 1, and the constant term d is 0.

[0139] Comparative Example 2-2

[0140] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 8, with the only difference being:

[0141] 1. The input features of the three primary PV power prediction models are the same and are all first-category features and second-category features;

[0142] The first type of feature: the irradiance of all training samples in the training set is arranged in sequence according to the acquisition time to obtain the irradiance curve corresponding to the training set; the irradiance curve is decomposed using the wavelet decomposition algorithm to obtain N irradiance wavelet components, and the Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficient between each irradiance wavelet component and the photovoltaic power curve corresponding to the training set. The photovoltaic power curve corresponding to the training set and the N irradiance wavelet components obtain N correlation coefficients, and the N correlation coefficients constitute the correlation coefficient set E4; the frequency value of the irradiance wavelet component corresponding to the correlation coefficient with an absolute value greater than 0.2 in the correlation coefficient set E4 is used as the first type of feature;

[0143] The second type of features: The Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficients between the eight meteorological data curves and the photovoltaic power curves corresponding to the training set and form a correlation coefficient set E3; the meteorological data corresponding to the correlation coefficients with absolute values ​​greater than 0.2 in the correlation coefficient set E3 are selected as the second type of features.

[0144] 2. Weight A', weight B', and weight C' are all 1, and the constant term d is 0.

[0145] Comparative Examples 2-3

[0146] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 9, with the only difference being:

[0147] 1. The input features of the three primary PV power prediction models are the same and are all first-category features and second-category features;

[0148] The first type of feature: the irradiance of all training samples in the training set is arranged in sequence according to the acquisition time to obtain the irradiance curve corresponding to the training set; the irradiance curve is decomposed using the wavelet decomposition algorithm to obtain N irradiance wavelet components, and the Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficient between each irradiance wavelet component and the photovoltaic power curve corresponding to the training set. The photovoltaic power curve corresponding to the training set and the N irradiance wavelet components obtain N correlation coefficients, and the N correlation coefficients constitute the correlation coefficient set E4; the frequency value of the irradiance wavelet component corresponding to the correlation coefficient with an absolute value greater than 0.2 in the correlation coefficient set E4 is used as the first type of feature;

[0149] The second type of features: The Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficients between the eight meteorological data curves and the photovoltaic power curves corresponding to the training set and form a correlation coefficient set E3; the meteorological data corresponding to the correlation coefficients with absolute values ​​greater than 0.2 in the correlation coefficient set E3 are selected as the second type of features.

[0150] 2. Weight A', weight B', and weight C' are all 1, and the constant term d is 0.

[0151] Comparative Example 3-1

[0152] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 7, with the only difference being:

[0153] 1. The first-category features of the three primary photovoltaic power prediction models are the same, and the first-category feature is always irradiance.

[0154] 2. Weight A', weight B', and weight C' are all 1, and the constant term d is 0.

[0155] Comparative Example 3-2

[0156] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 8, with the only difference being:

[0157] 1. The first-category features of the three primary photovoltaic power prediction models are the same, and the first-category feature is always irradiance.

[0158] 2. Weight A', weight B', and weight C' are all 1, and the constant term d is 0.

[0159] Comparative Example 3-3

[0160] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 9, with the only difference being:

[0161] 1. The first-category features of the three primary photovoltaic power prediction models are the same, and the first-category feature is always irradiance.

[0162] 2. Weight A', weight B', and weight C' are all 1, and the constant term d is 0.

[0163] Comparative Example 4-1

[0164] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 7, with the only difference being:

[0165] 1. Steps 1-3) are different. Steps 1-3) of this comparative example are: the photovoltaic power primary prediction model selects the frequency value of the irradiance wavelet component corresponding to the correlation coefficient with an absolute value greater than 0.2 in the corresponding correlation coefficient set E1 as the first type of feature.

[0166] 2. Weight A', weight B', and weight C' are all 1, and the constant term d is 0.

[0167] Comparative Example 4-2

[0168] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 8, with the only difference being:

[0169] 1. Steps 1-3) are different. Steps 1-3) of this comparative example are: the photovoltaic power primary prediction model selects the frequency value of the irradiance wavelet component corresponding to the correlation coefficient with an absolute value greater than 0.2 in the corresponding correlation coefficient set E1 as the first type of feature.

[0170] 2. Weight A', weight B', and weight C' are all 1, and the constant term d is 0.

[0171] Comparative Example 4-3

[0172] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 9, with the only difference being:

[0173] 1. Steps 1-3) are different. Steps 1-3) of this comparative example are: the photovoltaic power primary prediction model selects the frequency value of the irradiance wavelet component corresponding to the correlation coefficient with an absolute value greater than 0.2 in the corresponding correlation coefficient set E1 as the first type of feature.

[0174] 2. Weight A', weight B', and weight C' are all 1, and the constant term d is 0.

[0175] Comparative Example 5-1

[0176] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 7, the only difference being that weight A', weight B', and weight C' are all 1, and the constant term d is 0.

[0177] Comparative Example 5-2

[0178] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 8, with the only difference being that weight A', weight B', and weight C' are all 1, and the constant term d is 0.

[0179] Comparative Example 5-3

[0180] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 9, the only difference being that weight A', weight B', and weight C' are all 1, and the constant term d is 0.

[0181] FIG6( a ) is a graph showing the error between the final predicted value of the photovoltaic power (on a sunny day) and the sample true value (actual value) for Comparative Example 1-1 (M1), Comparative Example 2-1 (M2), Comparative Example 3-1 (M3), Comparative Example 4-1 (M4), and Comparative Example 5-1 (M5);

[0182] FIG6( b ) is a graph showing the error between the final predicted value of the (cloudy) photovoltaic power and the true value of the sample (actual value) for Comparative Example 1-2 (M1), Comparative Example 2-2 (M2), Comparative Example 3-2 (M3), Comparative Example 4-2 (M4), and Comparative Example 5-2 (M5);

[0183] FIG6( c ) is a graph showing the error between the final predicted value of the photovoltaic power (on rainy days) and the sample true value (actual value) for Comparative Example 1-3 (M1), Comparative Example 2-3 (M2), Comparative Example 3-3 (M3), Comparative Example 4-3 (M4), and Comparative Example 5-3 (M5);

[0184] The final predicted values ​​of photovoltaic power of Comparative Examples 1-1 to 5-3 were evaluated using RMSE, MAE, and MAPE evaluation indicators, as shown in Table 3.

[0185] Table 3

[0186]

[0187]

[0188] As shown in Table 3, the first-category features input in Comparative Example 1 and Comparative Example 3 are the same and the second-category features are both selected from eight types of meteorological data, but the prediction accuracy of Comparative Example 3 is better than that of Comparative Example 1.

[0189] By comparing Comparative Example 2 and Comparative Example 4, it can be seen that the prediction accuracy of Comparative Example 4 is better than that of Comparative Example 2. Therefore, the irradiance wavelet component can only improve the prediction accuracy of the high-frequency item reconstructed data after the photovoltaic power reconstruction, and will affect the prediction accuracy of the medium-frequency item reconstructed data after the photovoltaic power reconstruction and the low-frequency item reconstructed data after the photovoltaic power reconstruction, resulting in an increase in the error of the final predicted value of the photovoltaic power.

[0190] Among Comparative Examples 1 to 5, the prediction error of Comparative Example 5 is the smallest, and the prediction accuracy is greatly improved. This shows that the irradiance decomposition and feature screening in Comparative Example 5 can select the features with strong correlation corresponding to each frequency item as input features according to the characteristics of each frequency item of each type of reconstructed data, thereby improving the accuracy of the prediction results.

[0191] Comparative Example 6-1

[0192] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 7, with the only difference being that: the input features of the three photovoltaic power primary prediction models are the same and are all first-category features and second-category features, and the first-category feature always selects irradiance; the second-category feature: the Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficients between eight meteorological data curves and the photovoltaic power curves corresponding to the training set to form a correlation coefficient set E3; the meteorological data corresponding to the correlation coefficients with absolute values ​​greater than 0.2 in the correlation coefficient set E3 are selected as the second-category features.

[0193] Comparative Example 6-2

[0194] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 8, with the only difference being that: the input features of the three photovoltaic power primary prediction models are the same and are all first-category features and second-category features, and the first-category feature always selects irradiance; the second-category feature: the Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficients between eight meteorological data curves and the photovoltaic power curves corresponding to the training set to form a correlation coefficient set E3; the meteorological data corresponding to the correlation coefficients with absolute values ​​greater than 0.2 in the correlation coefficient set E3 are selected as the second-category features.

[0195] Comparative Example 6-3

[0196] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 9, with the only difference being that: the input features of the three photovoltaic power primary prediction models are the same and are all first-category features and second-category features, and the first-category feature always selects irradiance; the second-category feature: the Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficients between eight meteorological data curves and the photovoltaic power curves corresponding to the training set to form a correlation coefficient set E3; the meteorological data corresponding to the correlation coefficients with absolute values ​​greater than 0.2 in the correlation coefficient set E3 are selected as the second-category features.

[0197] Comparative Example 7-1

[0198] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 7, the only difference being that the input features of the three photovoltaic power primary prediction models are the same and are all first-category features and second-category features;

[0199] The first type of feature: the irradiance of all training samples in the training set is arranged in sequence according to the acquisition time to obtain the irradiance curve corresponding to the training set; the irradiance curve is decomposed using the wavelet decomposition algorithm to obtain N irradiance wavelet components, and the Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficient between each irradiance wavelet component and the photovoltaic power curve corresponding to the training set. The photovoltaic power curve corresponding to the training set and the N irradiance wavelet components obtain N correlation coefficients, and the N correlation coefficients constitute the correlation coefficient set E4; the frequency value of the irradiance wavelet component corresponding to the correlation coefficient with an absolute value greater than 0.2 in the correlation coefficient set E4 is used as the first type of feature;

[0200] The second type of features: The Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficients between the eight meteorological data curves and the photovoltaic power curves corresponding to the training set and form a correlation coefficient set E3; the meteorological data corresponding to the correlation coefficients with absolute values ​​greater than 0.2 in the correlation coefficient set E3 are selected as the second type of features.

[0201] Comparative Example 7-2

[0202] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 8, the only difference being that the input features of the three photovoltaic power primary prediction models are the same and are all first-category features and second-category features;

[0203] The first type of feature: the irradiance of all training samples in the training set is arranged in sequence according to the acquisition time to obtain the irradiance curve corresponding to the training set; the irradiance curve is decomposed using the wavelet decomposition algorithm to obtain N irradiance wavelet components, and the Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficient between each irradiance wavelet component and the photovoltaic power curve corresponding to the training set. The photovoltaic power curve corresponding to the training set and the N irradiance wavelet components obtain N correlation coefficients, and the N correlation coefficients constitute the correlation coefficient set E4; the frequency value of the irradiance wavelet component corresponding to the correlation coefficient with an absolute value greater than 0.2 in the correlation coefficient set E4 is used as the first type of feature;

[0204] The second type of features: The Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficients between the eight meteorological data curves and the photovoltaic power curves corresponding to the training set and form a correlation coefficient set E3; the meteorological data corresponding to the correlation coefficients with absolute values ​​greater than 0.2 in the correlation coefficient set E3 are selected as the second type of features.

[0205] Comparative Example 7-3

[0206] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 9, the only difference being that the input features of the three photovoltaic power primary prediction models are the same and are all first-category features and second-category features;

[0207] The first type of feature: the irradiance of all training samples in the training set is arranged in sequence according to the acquisition time to obtain the irradiance curve corresponding to the training set; the irradiance curve is decomposed using the wavelet decomposition algorithm to obtain N irradiance wavelet components, and the Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficient between each irradiance wavelet component and the photovoltaic power curve corresponding to the training set. The photovoltaic power curve corresponding to the training set and the N irradiance wavelet components obtain N correlation coefficients, and the N correlation coefficients constitute the correlation coefficient set E4; the frequency value of the irradiance wavelet component corresponding to the correlation coefficient with an absolute value greater than 0.2 in the correlation coefficient set E4 is used as the first type of feature;

[0208] The second type of features: The Spearman rank correlation coefficient algorithm is used to calculate the correlation coefficients between the eight meteorological data curves and the photovoltaic power curves corresponding to the training set and form a correlation coefficient set E3; the meteorological data corresponding to the correlation coefficients with absolute values ​​greater than 0.2 in the correlation coefficient set E3 are selected as the second type of features.

[0209] Comparative Example 8-1

[0210] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 7, with the only difference being that the first-category features input to the three photovoltaic power primary prediction models are the same, and the first-category feature is always irradiance.

[0211] Comparative Example 8-2

[0212] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 8, with the only difference being that the first-category features input to the three photovoltaic power primary prediction models are the same, and the first-category feature is always irradiance.

[0213] Comparative Example 8-3

[0214] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 9, with the only difference being that the first-category features input to the three photovoltaic power primary prediction models are the same, and the first-category feature is always irradiance.

[0215] Comparative Example 9-1

[0216] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 7, with the only difference being that steps 1-3) are different. Steps 1-3) of this comparative example are: the photovoltaic power primary prediction model selects the frequency value of the irradiance wavelet component corresponding to the correlation coefficient with an absolute value greater than 0.2 in its corresponding correlation coefficient set E1 as the first type of feature.

[0217] Comparative Example 9-2

[0218] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 8, with the only difference being that steps 1-3) are different. Steps 1-3) of this comparative example are: the photovoltaic power primary prediction model selects the frequency value of the irradiance wavelet component corresponding to the correlation coefficient with an absolute value greater than 0.2 in its corresponding correlation coefficient set E1 as the first type of feature.

[0219] Comparative Example 9-3

[0220] A short-term photovoltaic power combination prediction method based on feature screening is basically the same as Example 9, with the only difference being that steps 1-3) are different. Steps 1-3) of this comparative example are: the photovoltaic power primary prediction model selects the frequency value of the irradiance wavelet component corresponding to the correlation coefficient with an absolute value greater than 0.2 in its corresponding correlation coefficient set E1 as the first type of feature.

[0221] FIG7( a ) is a graph showing the error between the final predicted value of photovoltaic power (on a sunny day) and the sample true value (actual value) for Example 7 (M10), Comparative Example 6-1 (M6), Comparative Example 7-1 (M7), Comparative Example 8-1 (M8), and Comparative Example 9-1 (M9);

[0222] FIG7( b ) is a graph showing the error between the final predicted value of the (cloudy) photovoltaic power and the sample true value (actual value) of Example 8 (M10), Comparative Example 6-2 (M6), Comparative Example 7-2 (M7), Comparative Example 8-2 (M8), and Comparative Example 9-2 (M9);

[0223] FIG7( c ) is a graph showing the error between the final predicted value of photovoltaic power (on rainy days) and the sample true value (actual value) for Example 9 (M10), Comparative Example 6-3 (M6), Comparative Example 7-3 (M7), Comparative Example 8-3 (M8), and Comparative Example 9-3 (M9);

[0224] The weight A', weight B', weight C' and constant term d of Examples 7 to 9 and Comparative Examples 6 to 9 are shown in Table 4.

[0225] Table 4

[0226]

[0227]

[0228] The final predicted values ​​of photovoltaic power of Examples 7 to 9 and Comparative Examples 6 to 9 were evaluated using RMSE, MAE, and MAPE evaluation indicators, as shown in Table 5.

[0229] Table 5

[0230]

[0231] It can be seen from Tables 3 and 5 that the final predicted values ​​of photovoltaic power of Examples 7 to 9, Comparative Example 6 and Comparative Example 8 have better fit with the true value, and the fitting curves have more overlap on sunny days; under sunny weather, compared with Comparative Example 5-1, the MAPE of Example 7 is improved by 67.86%; under cloudy weather, the MAPE of Comparative Example 2-2 and Comparative Example 4-2 with poor prediction accuracy are improved by 13.27% and 10.95% respectively after weight optimization; under rainy weather, compared with Comparative Example 5-3 and Comparative Example 3-3, the MAPE of Example 9 and Comparative Example 8-3 are improved by 41.40% and 43.77% respectively.

[0232] In summary, whether it is a sunny day with relatively stable photovoltaic power or a rainy day with strong randomness, Examples 7 to 9 all show high prediction accuracy. The short-term photovoltaic power combination prediction method based on feature screening and weight optimization of the present invention fully considers the mutual influence between high-frequency items, medium-frequency items and low-frequency items, and realizes reasonable distribution of weights by comprehensively evaluating the contribution of each frequency item, thereby effectively improving the prediction accuracy.

[0233] Comparative Example 10

[0234] A short-term photovoltaic power portfolio forecast based on similar day selection and data reconstruction. See: Chen Qingbin, Yang Genghuang, Geng Liqing, et al. Short-term photovoltaic power portfolio forecast based on similar day selection and data reconstruction [J / OL]. China Electric Power, 1-11 [2024-11-25].

[0235] The RMSE, MAE, and MAPE evaluation indicators were used to evaluate the final photovoltaic power prediction values ​​on sunny days, cloudy days, and rainy days of Examples 7 to 9 and Comparative Example 10. As shown in Table 6, the prediction accuracy of Examples 7 to 9 was much higher than that of Comparative Example 10. In rainy weather, the MAPE of Comparative Example 10 was about 10% higher than that of Example 9, which shows that the short-term photovoltaic power combination prediction method based on feature screening and weight optimization still has certain advantages in rainy weather with strong randomness and large fluctuations.

[0236] Table 6

[0237]

[0238] The above is an exemplary description of the present invention. It should be noted that, without departing from the core of the present invention, any simple deformation, modification or other equivalent replacement that can be made by other skilled in the art without expending creative labor falls within the scope of protection of the present invention.

Claims

1. A photovoltaic power prediction model, characterized in that: include: Three primary photovoltaic power prediction models are trained using a training set. The training samples in the training set belong to the same type of weather. The input features of each primary photovoltaic power prediction model are first-class features and second-class features. The first-class features and second-class features of each primary photovoltaic power prediction model are determined using the training set. The selection rules for the first type of features include 1-1) to 1-3) 1-1) Obtaining reconstructed data. Methods for obtaining reconstructed data include 1-1-1) to 1-1-3): 1-1-1) Taking the photovoltaic power as the sample true value, arranging the sample true values ​​corresponding to all training samples in the training set according to the acquisition time of the sample true value, obtaining the photovoltaic power curve corresponding to the training set, and decomposing the photovoltaic power curve to obtain M intrinsic mode functions; 1-1-2) Calculate the sample entropy value corresponding to each intrinsic mode function, cluster the M sample entropy values, and obtain three clusters as clustering results. The three clusters are divided into high-frequency items, medium-frequency items, and low-frequency items according to the average entropy values ​​of the samples in the clusters from high to low; 1-1-3) superimposing the intrinsic mode functions corresponding to the sample entropy values ​​belonging to the high-frequency items to obtain the reconstructed data of the high-frequency items; superimposing the intrinsic mode functions corresponding to the sample entropy values ​​belonging to the medium-frequency items to obtain the reconstructed data of the medium-frequency items; superimposing the intrinsic mode functions corresponding to the sample entropy values ​​belonging to the low-frequency items to obtain the reconstructed data of the low-frequency items; using the high-frequency item reconstructed data, the medium-frequency item reconstructed data and the low-frequency item reconstructed data as the reconstructed data respectively; making the high-frequency item reconstructed data correspond to the first photovoltaic power primary prediction model, the medium-frequency item reconstructed data correspond to the second photovoltaic power primary prediction model, and the low-frequency item reconstructed data correspond to the third photovoltaic power primary prediction model; 1-2) Obtain a set of correlation coefficients E1 between N irradiance wavelet components and each set of reconstructed data; 1-3) Take the largest correlation coefficient in the correlation coefficient set E1 as R max Calculate the correlation coefficient between the irradiance curve corresponding to the training set and each set of reconstructed data and use it as R i , if the R of the reconstructed data corresponding to the same photovoltaic power primary prediction model is i Greater than R max , the photovoltaic power primary prediction model selects irradiance as the first-category feature; otherwise, the photovoltaic power primary prediction model selects the frequency value of the irradiance wavelet component corresponding to the correlation coefficient with an absolute value greater than 0.2 in the corresponding correlation coefficient set E1 as the first-category feature; The selection rules for the second type of features include: obtaining a set E2 of correlation coefficients between eight types of meteorological data and each set of reconstructed data; the photovoltaic power primary prediction model selects the meteorological data corresponding to the correlation coefficient with an absolute value greater than 0.2 in the set E2 of correlation coefficients corresponding to the reconstructed data as the second type of features.

2. The photovoltaic power prediction model according to claim 1, characterized in that: The first photovoltaic power primary prediction model includes: a normalization module, a CNN-BiLSTM-Attention neural network and an anti-normalization module connected in sequence.

3. The photovoltaic power prediction model according to claim 1, characterized in that: The second photovoltaic power primary prediction model and the third photovoltaic power primary prediction model both include: a normalization module, an LSTM neural network and an anti-normalization module connected in sequence.

4. The photovoltaic power prediction model according to claim 1, characterized in that: The photovoltaic power prediction model also includes: a weighted module, the three photovoltaic power primary prediction models are the first photovoltaic power primary prediction model, the second photovoltaic power primary prediction model and the third photovoltaic power primary prediction model, the output of the first photovoltaic power primary prediction model is the first photovoltaic power preliminary prediction value, the output of the second photovoltaic power primary prediction model is the second photovoltaic power preliminary prediction value, the output of the third photovoltaic power primary prediction model is the third photovoltaic power preliminary prediction value, the weights corresponding to the first photovoltaic power preliminary prediction value, the second photovoltaic power preliminary prediction value and the third photovoltaic power preliminary prediction value are weights A and B respectively. ’ , weight B ’ and weight C ’ The input of the weighted module is the first photovoltaic power preliminary prediction value, the second photovoltaic power preliminary prediction value, the third photovoltaic power preliminary prediction value, and the weight A. ’ , weight B ’ , weight C ’ and the constant term d ’ , the output of the weighted module is the final predicted value of PV power.

5. The photovoltaic power prediction model according to claim 4, characterized in that: Weight A ’ , weight B ’ and weight C ’ The dung beetle algorithm is used for weighted optimization.

6. The photovoltaic power prediction model according to claim 1, characterized in that: The method for obtaining the correlation coefficient set E1 between N irradiance wavelet components and each group of reconstructed data is as follows: the irradiances in all training samples in the training set are arranged in sequence according to the acquisition time to obtain the irradiance curve corresponding to the training set; the irradiance curve is decomposed to obtain N irradiance wavelet components, and the correlation coefficient between each irradiance wavelet component and the group of reconstructed data is calculated. N correlation coefficients are obtained between the group of reconstructed data and the N irradiance wavelet components, and the N correlation coefficients constitute the correlation coefficient set E1.

7. The photovoltaic power prediction model according to claim 1, characterized in that: The method for obtaining the correlation coefficient set E2 between the eight types of meteorological data and each group of reconstructed data is as follows: each type of meteorological data of all training samples in the training set is arranged in sequence according to the collection time to obtain a meteorological data curve; a total of eight meteorological data curves are obtained for the eight types of meteorological data; the correlation coefficients between the group of reconstructed data and the eight meteorological data curves are calculated and form the correlation coefficient set E2.

8. The photovoltaic power prediction model according to claim 5, characterized in that: The fitness function of the dung beetle algorithm is the mean absolute percentage error between the final predicted value of PV power and the true value of the samples.

9. The photovoltaic power prediction model training method according to any one of claims 1 to 8, characterized in that: include: According to the selection rules and the training set, the first and second features of each photovoltaic power primary prediction model are determined and the first and second features are input into the photovoltaic power primary prediction model to train the photovoltaic power prediction model to obtain a trained photovoltaic power prediction model.

10. A short-term photovoltaic power combination forecasting method based on feature screening and weight optimization, comprising: Input the first and second characteristics corresponding to the test sample into each photovoltaic power primary prediction model in the trained photovoltaic power prediction model described in any one of claims 1 to 8, so that the photovoltaic power prediction model performs prediction to obtain the final predicted value of photovoltaic power.

Images