Photovoltaic power generation output prediction model construction method based on three-channel characteristics
By constructing a three-channel CNN-BLSTM model that integrates spatial, temporal, and seasonal features, the shortcomings of existing photovoltaic power output prediction models in capturing spatial features and multi-scale temporal patterns are addressed, resulting in more accurate photovoltaic power output prediction and improved grid dispatch stability and renewable energy grid connection capabilities.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUIZHOU ELECTRIC POWER DESIGN INST
- Filing Date
- 2025-12-05
- Publication Date
- 2026-04-28
AI Technical Summary
Existing photovoltaic power output prediction models are insufficient in capturing spatial characteristics and multi-scale time-series patterns, especially in power plant-level applications with multi-source data, making it difficult to achieve accurate photovoltaic power output prediction.
A photovoltaic power generation output prediction model with three-channel features is adopted. The original features, seasonal features and time-frequency features are processed in parallel by CNN-BLSTM branches, and feature fusion is performed by combining attention mechanism to construct a three-channel CNN-BLSTM model, which solves the problem of unstable prediction accuracy of existing models in seasonal and multi-scale time series patterns.
It has enabled more accurate photovoltaic power output forecasting, reduced the differences in forecast accuracy across different seasons, and improved the stability and reliability of grid dispatching and renewable energy grid connection.
Smart Images
Figure CN121936652A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of photovoltaic energy and deep learning interdisciplinary technology, and more specifically, to a method for constructing a photovoltaic power generation output prediction model based on three-channel features. Background Technology
[0002] The construction scale of distributed photovoltaic (PV) power stations in China is accelerating year by year. However, with the continuous expansion of installed capacity, the instability caused by fluctuations in PV output is becoming increasingly prominent. PV power generation is inherently affected by a series of complex meteorological factors, the effects of which change dynamically with time and weather conditions. This volatility leads to intermittent and unstable PV output, posing a challenge to the stable operation of the power grid. To alleviate the stability problem of distributed PV systems, accurate prediction of PV output is needed to optimize grid dispatch, improve operational safety and reliability, and promote the grid connection of renewable energy and optimize load management.
[0003] Currently, photovoltaic (PV) power output prediction methods are mainly divided into three categories: physical models, statistical models, and machine learning / deep learning models. Physical models rely on the dynamic relationship between solar radiation and physical laws; statistical models mainly rely on historical data, statistical analysis, and probability principles to predict future time series trends; machine learning models establish a direct mapping between input and output, and can extract complex nonlinear features. Currently, machine learning and deep learning methods have been widely applied in the field of PV power output prediction. Long Short-Term Memory (LSTM) models and Bidirectional Long Short-Term Memory (BLSTM) models are adept at capturing complex temporal features in PV power output data, and many studies have adopted these methods. However, BLSTM models have shortcomings in capturing spatial features and modeling multi-scale temporal patterns.
[0004] Existing technologies offer several solutions to address these shortcomings: combining time-frequency transformation with convolutional neural networks to capture local frequency features, and using CNN-LSTM hybrid models to perform multi-step predictions by fusing spatial and temporal information. However, general-purpose CNN-LSTM models are not optimized for specific seasonal or micro-meteorological patterns. This deficiency is particularly pronounced in power plant-level applications that require refined analysis of multi-source data.
[0005] Therefore, a technical solution is needed to integrate spatial and temporal characteristics to achieve more accurate photovoltaic power output prediction. Summary of the Invention
[0006] To achieve the above objectives, this application provides a method for constructing a photovoltaic power generation output prediction model based on three-channel characteristics, comprising the following steps: Raw data is collected, preprocessed, and sampled to generate sample datasets. The raw data includes meteorological variables and target variables, with the target variable being active power. During sample partitioning, the preprocessed raw data is categorized into four seasonal sample datasets: spring, summer, autumn, and winter. Each seasonal sample dataset is further divided into training, testing, and prediction sets. The sample dataset is filtered, decomposed, and extracted to obtain three-channel input features; Three parallel CNN-BLSTM branches are constructed, and features are fused through an attention mechanism to achieve multi-dimensional modeling of "spatial-temporal-seasonal-time-frequency" and generate a three-channel CNN-BLSTM model; Based on the normalized predicted value output by the three-channel CNN-BLSTM model, the normalized predicted value is restored to the actual power prediction value; Evaluate the regression model, obtain the optimal regression model, and construct a photovoltaic power generation output prediction model.
[0007] The preprocessing includes time alignment and normalization. Time alignment is used to unify the timestamps of all variables and eliminate outlier samples with mismatched times. Normalization maps all variables to the [0,1] interval to eliminate the impact of dimensional differences on model training.
[0008] Furthermore, obtaining the three-channel input features includes: obtaining the original features as the first channel input features; extracting seasonal features as the second channel input features; and extracting time-frequency features as the third channel input features.
[0009] Obtaining the original features includes: Obtain principal component information, and then obtain collinearity feature clusters by measuring the magnitude of variable collinearity; Multiple collinear feature pairs were obtained by using the feature correlation coefficients between collinear feature clusters. After confirming the prevalence of collinearity in feature clusters, dimensionality reduction was performed by principal component analysis to obtain the core information in the collinear feature clusters after eliminating collinearity. The principal component information after dimensionality reduction and the non-collinear feature clusters are merged to generate non-collinear features as the original features.
[0010] Seasonal feature extraction is achieved through STL decomposition; after STL decomposition, trend component, seasonal component and residual component are generated; among them, trend component is used to reflect the long-term power change trend, seasonal component is used to reflect the intraday power fluctuation pattern, and residual component is used to reflect random fluctuation.
[0011] Time-frequency feature extraction refers to the process of decomposing the residual components and extracting low-frequency and high-frequency coefficients. The low-frequency coefficients reflect the slow variation trend of the residuals, while the high-frequency coefficients reflect the rapid fluctuations of the residuals. The low-frequency and high-frequency coefficients are then reconstructed into a time-frequency feature vector.
[0012] Furthermore, time-frequency feature extraction supports Fourier transform and wavelet transform; wavelet transform is used for scenarios that generate local sudden fluctuations.
[0013] Furthermore, each CNN-BLSTM branch includes a CNN feature extraction layer + a BLSTM temporal modeling layer, and the parameters are set uniformly. The parameter settings include: The CNN layer consists of one convolutional layer and one max pooling layer, with ReLU activation function, used to extract spatial correlations of input features; the spatial correlations include the spatial mapping relationship between meteorological parameters and power. The BLSTM layer is a single-layer BLSTM used to capture the bidirectional temporal dependencies of features, including the impact of radiation changes in the previous hour on the current power.
[0014] The three parallel CNN-BLSTM branches each have three channels. After assigning input features to the first channel, the second channel, and the third channel respectively, the branch attention is fused through weight calculation. The first channel input feature is used to capture the basic spatiotemporal features of the raw data; the second channel input feature is used to capture seasonal and long-term trend features; and the third channel input feature is used to capture the local frequency fluctuation features of power. Among them, the weight calculation supports adaptive allocation and fixed weight fusion with weights dynamically set according to the season.
[0015] Furthermore, the output layer of the three-channel CNN-BLSTM model is a single Dense layer; the output layer will fuse the... The mapping is to normalized predicted values.
[0016] This invention can construct a photovoltaic power generation output prediction model that balances accuracy and seasonal stability. It extracts spatial features (such as the spatial distribution of the correlation between meteorological parameters and power) through CNN, and combines this with the bidirectional temporal modeling capability of BLSTM to solve the problem that LSTM / BLSTM models cannot effectively capture spatial features and multi-scale temporal patterns. A three-channel design processes the "original features after dimensionality reduction," the "seasonal / trend / residual features from STL decomposition," and the "time-frequency features from wavelet transform" separately, and then weights and fuses them through a branch attention mechanism to weaken the impact of seasonal differences on the prediction results, thus solving the problem of unstable seasonal prediction accuracy in single-channel CNN-BLSTM models. By using VIF (variance inflation factor) and correlation coefficients to screen for collinear features, and combining this with a dimensionality reduction strategy that retains 95% of the cumulative variance using PCA, it eliminates multicollinearity while maximizing the preservation of original feature information, thus resolving the contradiction between "multicollinearity" and "feature information loss" in existing feature processing technologies. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating the steps of constructing a photovoltaic power generation output prediction model according to an embodiment of the present invention; Figure 2 This is a time series curve of each variable of the original data provided according to an embodiment of the present invention; Figure 3 This is a comparison chart of multi-model accuracy evaluation provided according to an embodiment of the present invention; Figure 4 This is a quarterly accuracy evaluation chart provided according to an embodiment of the present invention. Detailed Implementation
[0018] This invention integrates data collection and preprocessing, feature selection, a three-channel CNN-BLSTM model, and accuracy testing to achieve rapid and accurate short-term photovoltaic output prediction, and effectively reduces the difference in prediction accuracy across different seasons, providing technical support for grid dispatch and renewable energy grid connection.
[0019] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.
[0020] Figure 1 A method and steps for constructing a photovoltaic power generation output prediction model based on three-channel characteristics are provided, including the following steps: Step S100: Collect raw data, preprocess the raw data, and generate a sample dataset; In this step, the sampling interval and time span are specified to collect relevant raw data on photovoltaic power generation. The raw data consists of meteorological variables and target variables; among them, meteorological variables include: wind direction, ambient temperature (°C), ambient relative humidity, global horizontal radiation, diffuse horizontal radiation, global tilted radiation, and diffuse tilted radiation; the target variable is active power (i.e., photovoltaic power generation output).
[0021] The present invention provides an embodiment that collects publicly available data from a local solar energy center, with the data spanning from July 1, 2016 to July 1, 2017, and a sampling interval of 5 minutes (288 sampling points per day).
[0022] After the raw data is collected, data preprocessing is performed, which includes the following operations: 1) Time alignment: unify the timestamps of all variables and remove outlier samples with mismatched times (such as meteorological data but no power data at a certain moment).
[0023] 2) Normalization process: Min-Max normalization is used to map all variables to the [0,1] interval, eliminating the impact of dimensional differences on model training. The normalization formula is: , in, The original variable value, , These are the minimum and maximum values of the variable, respectively. This is the value after normalization.
[0024] The time distribution characteristics of each variable after normalization are as follows: Figure 2 As shown in the figure, the vertical axis represents the normalized variable values; the curves correspond to: (a) - ambient temperature (°C), (b) - ambient relative humidity, (c) - global horizontal radiation, (d) - scattered horizontal radiation, (e) - wind direction, (f) - global tilted radiation, (g) - scattered tilted radiation, and (h) - active power.
[0025] After data preprocessing, sample splitting is performed: first, the original data is split according to the season, and the preprocessed original data is classified into four seasonal sample datasets: spring, summer, autumn and winter. Then, each seasonal sample dataset is divided into training set, test set and prediction set in an 8:1:1 ratio for subsequent model training and seasonal accuracy verification.
[0026] Step S110: Filter, decompose, and extract from the sample dataset to obtain three-channel input features; This step involves feature extraction based on four seasonal sample datasets to address the issues of multicollinearity and missing seasonal features. The specific feature extraction process includes the following steps: 1) Obtain the original features as the input features for the first channel; In practice, the principal component information is first obtained, and then collinear feature clusters are obtained by measuring the magnitude of variable collinearity. Further, multiple collinear feature pairs are obtained by measuring the feature correlation coefficients between collinear feature clusters. After confirming the prevalence of collinearity in the feature clusters, dimensionality reduction is performed by principal component analysis (PCA) to obtain the core information in the collinear feature clusters after eliminating collinearity. The principal components after PCA and the non-collinear feature clusters are then merged to obtain the non-collinear features as the original features. The original features are used as the first channel input features to capture the basic spatiotemporal features of the original data.
[0027] Specifically, for principal component information, the magnitude of variable collinearity can be obtained through the `variance_inflation_factor()` function in Python's `statsmodels.stats.outliers_influence`, the correlation coefficient can be calculated using NumPy's `.corrcoef()` function, and the PCA() function in `sklearn.decomposition` can be used to obtain the information.
[0028] 2) Extract seasonal features as the second channel input features: This step performs STL decomposition (Seasonal and Trend decomposition using Loess) on the target variable (active power). For the embodiments of the present invention, the parameter settings for STL decomposition include: core parameter period=288 (corresponding to the number of 5-minute sampling points per day, i.e., 1 day is 1 seasonal cycle). After decomposition, three types of components are obtained: trend component, seasonal component, and residual component. These three types of components are used as the second channel input features to capture seasonal and trend features.
[0029] Specifically, trend components can reflect long-term power change trends, such as power rises and falls caused by seasonal changes; seasonal components can reflect intraday power fluctuation patterns, such as midday power peaks; and residual components can reflect random fluctuations, such as sudden power drops caused by sudden cloud formations.
[0030] The residual components generated in this step reflect random fluctuations and have large random fluctuations, therefore, further time-frequency feature processing is required.
[0031] 3) Extract time-frequency features and use them as input features for the third channel; In this step, the residual components after STL decomposition are further decomposed to extract low-frequency and high-frequency coefficients. The low-frequency coefficients reflect the slow variation trend of the residuals, while the high-frequency coefficients reflect the rapid fluctuations. Finally, the low-frequency and high-frequency coefficients are reconstructed into a time-frequency feature vector, which serves as the input feature for the third channel to capture local frequency fluctuations in power.
[0032] The extraction method supports both Fourier transform and wavelet transform. For power plants with extremely large data volumes (such as GW-level photovoltaic power plants), Fourier transform is used to ensure higher computational efficiency. For scenarios with sudden local fluctuations caused by short-term cloud cover, wavelet transform is used to ensure capture capability. When using wavelet transform, a 3-level wavelet decomposition is performed using the db4 wavelet basis function to extract low-frequency and high-frequency coefficients.
[0033] Step S120: Construct three parallel CNN-BLSTM branches, fuse features through an attention mechanism to achieve multi-dimensional modeling of "spatial-temporal-seasonal-time-frequency", and generate a three-channel CNN-BLSTM model; The specific implementation is as follows: 1) In terms of model structure, each CNN-BLSTM branch includes a CNN feature extraction layer + a BLSTM temporal modeling layer, and the parameters are uniformly set; the parameter settings include: CNN layer: 1 convolutional layer (3×3 convolutional kernels, 32 in number) + 1 max pooling layer (2×2 pooling kernels), with ReLU activation function, used to extract spatial correlations of input features (such as the spatial mapping relationship between meteorological parameters and power). BLSTM layer: 1 BLSTM layer with 64 neurons and a dropout rate of 0.2, used to capture bidirectional temporal dependencies of features (such as the effect of radiation changes in the previous hour on the current power).
[0034] 2) Feature allocation strategies include: The raw features after PCA processing are input into the first channel to capture the basic spatiotemporal features of the raw data; The trend component, seasonal component, and residual component of STL decomposition are input into the second channel to capture seasonal and long-term trend characteristics. The third channel inputs a time-frequency feature vector, which is responsible for capturing the local frequency fluctuation characteristics of the power.
[0035] 3) Achieve branch attention fusion through weight calculation: The hidden states of the three branches of the BLSTM output are used to calculate attention weights through a fully connected layer. (satisfy This achieves an adaptive allocation where "branches with higher contributions have higher weights"; among which, These are the weights of the three branches.
[0036] The hidden states of the three branches are weighted and fused, and the formula is: ,in, These are the hidden states of the BLSTM for the three branches.
[0037] 4) Output layer: The fused product is processed through a single Dense layer (with Linear activation function). The mapping is to normalized predicted values.
[0038] In summary, the three-channel CNN-BLSTM model extracts high-dimensional features of photovoltaic (PV) power generation from three branch networks and constructs a branch attention mechanism to automatically assign weights, resulting in the final PV power generation features. The attention mechanism includes a fully connected (Dense) layer with softmax activation, which takes the high-dimensional feature vectors from the connections of the three CNN-BLSTM branches as input. This layer computes a set of normalized weights (summing to 1) to quantify the contribution of each branch to the final prediction. These weights are dynamically learned through end-to-end training with the entire network, following a rule of "fixed within seasons, adaptive between seasons." This means the mechanism learns weight configurations specific to each season (e.g., assigning higher weights to the wavelet branch in summer when cloud interference is frequent, and assigning higher weights to the STL branch in spring when seasonal patterns are stable) to address the seasonal impact of the feature fusion stage. Finally, after obtaining weighted fused features through the branch attention mechanism (where the features of each branch are scaled according to their learned weights to emphasize more informational features of the current season), the integrated final features are fed into the Dense layer.
[0039] Furthermore, the normalized predicted values output by the three-channel CNN-BLSTM model are restored to the actual power values through Min-Max inverse normalization. The method is as follows: ,in, The normalized value predicted by the model. This is the predicted value of actual power.
[0040] Step S130: Evaluate the regression model, obtain the optimal regression model, and construct a photovoltaic power generation output prediction model.
[0041] In this step, three metrics are selected to evaluate the regression model. These three metrics are: Mean Squared Error (MSE), Coefficient of Determination (R²), and Mean Absolute Error (MAE), defined as follows: , , , in: This represents the measured value of photovoltaic power output. This represents the predicted photovoltaic power output, where n represents the total number of samples. denoted as the average value of the measured photovoltaic output, where i represents the i-th sample.
[0042] Specifically, the predict method is used to predict the test set, and the mean squared error, coefficient of determination, and mean absolute error are calculated by combining the r2_score, mean_squared_error, and mean_absolute_error methods in sklearn, thereby obtaining the optimal regression model.
[0043] This invention also provides an overall evaluation of multiple models, including a three-channel CNN-BLSTM model, an LSTM model, a single-channel CNN-BLSTM model, and a BLSTM model, based on a test set according to embodiments. The evaluation results are as follows: Figure 3 As shown.
[0044] It is evident that the three-channel CNN-BLSTM model of this invention exhibits the best performance: achieving a minimum MSE of 0.1077 and MAE of 0.1559, representing reductions of 42.1% and 29.0% respectively compared to the LSTM model (MSE=0.186, MAE=0.221); and achieving a maximum R² of 0.9705, approaching a perfect fit between the actual and predicted values (R²=1), representing a 1.3% improvement over the single-channel CNN-BLSTM (R²=0.958). This demonstrates that by fusing features from three channels, the model overcomes the blind spots in capturing spatial, temporal, and seasonal features, such as reducing the prediction error at the midday power peak by 35% compared to BLSTM.
[0045] Furthermore, tests were conducted on multiple models, including the three-channel CNN-BLSTM model, the LSTM model, and the single-channel CNN-BLSTM and BLSTM models, in conjunction with tests performed in spring, summer, autumn, and winter. The test results are as follows: Figure 4 As shown.
[0046] Among them, the MSE fluctuation of the three-channel CNN-BLSTM model was only 46.0%, from the lowest MSE of 0.1058 in autumn to the highest MSE of 0.1547 in winter. The fluctuation range was much lower than that of LSTM (452.4%), BLSTM (392.0%) and single-channel CNN-BLSTM (418.8%). Especially in the complex weather conditions of winter (weak radiation and large fluctuations), the error of traditional models increased sharply, while the three-channel CNN-BLSTM model of this invention still maintained MSE=0.1547 and R²=0.9501, which is 23.9% lower than that of LSTM and 1.57% higher in R². In a stable summer environment (strong radiation, small fluctuations), the three-channel CNN-BLSTM model of this invention achieves R²=0.99, which is comparable to the single-channel model, and its MSE=0.0448 is only slightly higher than that of the LSTM (0.0368), with no significant loss of accuracy. A comparison of the MSE / MAE / R² fluctuation ranges for each model is shown in Table 1. Table 1. Comparison of the seasonal fluctuation ranges of MSE / MAE / R² for each model This invention proposes a CNN-BLSTM structure based on the fusion of three channels: STL decomposition, wavelet time-frequency features, and PCA dimensionality reduction, for the construction of a photovoltaic power generation output prediction model. Regarding the determination of model features, it simultaneously covers four dimensions: spatial, temporal, seasonal, and time-frequency features, filling the gap in existing models for collaborative modeling of multi-source features. Due to the introduction of a branch attention mechanism to achieve adaptive feature weighting, there is no need to add an adaptive seasonal weight unit, enabling the model to maintain high stability in spring, summer, autumn, and winter. In terms of sample preparation, it is designed based on a 5-minute high-sampling-rate dataset containing 8 core variables, adapting to short-term prediction scenarios at the power plant level, and through normalization processing, it can be compatible with photovoltaic data from different regions and under different climatic conditions.
[0047] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.
Claims
1. A method for constructing a photovoltaic power generation output prediction model based on three-channel characteristics, characterized in that, Includes the following steps: Raw data is collected, preprocessed, and sampled to generate a sample dataset. The raw data includes meteorological variables and a target variable, with the target variable being active power. During sample segmentation, the preprocessed raw data is categorized into four seasonal sample datasets: spring, summer, autumn, and winter. The sample dataset is filtered, decomposed, and extracted to obtain three-channel input features; Three parallel CNN-BLSTM branches are constructed, and features are fused through an attention mechanism to achieve multi-dimensional modeling of "spatial-temporal-seasonal-time-frequency" and generate a three-channel CNN-BLSTM model; Based on the normalized predicted value output by the three-channel CNN-BLSTM model, the normalized predicted value is restored to the actual power predicted value; Evaluate the regression model, obtain the optimal regression model, and construct a photovoltaic power generation output prediction model.
2. The method for constructing a photovoltaic power generation output prediction model based on three-channel characteristics according to claim 1, characterized in that, The preprocessing includes time alignment and normalization. The time alignment is used to unify the timestamps of all variables and eliminate abnormal samples with mismatched times; the normalization process maps all variables to the [0,1] interval to eliminate the impact of dimensional differences on model training.
3. The method for constructing a photovoltaic power generation output prediction model based on three-channel characteristics according to claim 1, characterized in that, The acquisition of the three-channel input features includes: Obtain the original features as the input features for the first channel; Seasonal features are extracted and used as input features for the second channel. Time-frequency features are extracted and used as the input features for the third channel.
4. The method for constructing a photovoltaic power generation output prediction model based on three-channel characteristics according to claim 3, characterized in that, The acquisition of original features includes: Obtain principal component information, and then obtain collinearity feature clusters by measuring the magnitude of variable collinearity; Multiple collinear feature pairs were obtained by using the feature correlation coefficients between collinear feature clusters. After confirming the prevalence of collinearity in feature clusters, dimensionality reduction was performed by principal component analysis to obtain the core information in the collinear feature clusters after eliminating collinearity. The principal components and non-collinear feature clusters after dimensionality reduction are merged to generate non-collinear features as the original features.
5. The method for constructing a photovoltaic power generation output prediction model based on three-channel characteristics according to claim 3, characterized in that, The seasonal feature extraction is achieved through STL decomposition; after STL decomposition, trend components, seasonal components, and residual components are generated. The trend component is used to reflect the long-term power change trend, the seasonal component is used to reflect the intraday power fluctuation pattern, and the residual component is used to reflect random fluctuations.
6. The method for constructing a photovoltaic power generation output prediction model based on three-channel characteristics according to claim 3, characterized in that, The time-frequency feature extraction refers to: decomposing the residual components and extracting low-frequency coefficients and high-frequency coefficients; wherein, the low-frequency coefficients reflect the slow variation trend of the residuals, and the high-frequency coefficients reflect the rapid fluctuations of the residuals; the low-frequency coefficients and high-frequency coefficients are reconstructed into a time-frequency feature vector.
7. The method for constructing a photovoltaic power generation output prediction model based on three-channel characteristics according to claim 6, characterized in that, The time-frequency feature extraction supports Fourier transform and wavelet transform; the wavelet transform is used to generate scenarios with local sudden fluctuations.
8. The method for constructing a photovoltaic power generation output prediction model based on three-channel characteristics according to claim 3, characterized in that, Each of the CNN-BLSTM branches contains a CNN feature extraction layer and a BLSTM temporal modeling layer, and the parameters are set uniformly. The parameter settings include: The CNN layer consists of one convolutional layer and one max pooling layer, with ReLU activation function, used to extract spatial correlations of input features; the spatial correlations include the spatial mapping relationship between meteorological parameters and power. The BLSTM layer is a single-layer BLSTM used to capture the bidirectional temporal dependencies of features, including the impact of radiation changes in the previous hour on the current power.
9. The method for constructing a photovoltaic power generation output prediction model based on three-channel characteristics according to claim 8, characterized in that, The three parallel CNN-BLSTM branches are three channels, and after being assigned input features for the first channel, the second channel, and the third channel respectively, branch attention fusion is achieved through weight calculation. The first channel input feature is used to capture the basic spatiotemporal features of the raw data; the second channel input feature is used to capture seasonal and long-term trend features; and the third channel input feature is used to capture the local frequency fluctuation features of power. Among them, the weight calculation supports adaptive allocation and fixed weight fusion with weights dynamically set according to the season.
10. The method for constructing a photovoltaic power generation output prediction model based on three-channel characteristics according to claim 1, characterized in that, The output layer of the three-channel CNN-BLSTM model is a single Dense layer. The output layer will merge the results. The mapping is to normalized predicted values.