Distributed photovoltaic short-term power prediction method and system based on BiLSTM

Abstract

The invention belongs to the technical field of photovoltaic control, and discloses a distributed photovoltaic short-term power prediction method and system based on BiLSTM, and the method comprises the steps: screening meteorological feature vectors according to a unified information coefficient, wherein the unified information coefficient of the meteorological feature vectors and distributed photovoltaic output is greater than a set threshold value; performing dimension reduction on the screened meteorological feature vectors by adopting a factor analysis method to obtain a factor score matrix; performing nonlinear enhancement on the factor score matrix by using an SCN network as a feature intensifier to obtain an enhanced feature sequence and using the enhanced feature sequence as an input feature, and extracting a time sequence feature by using a BiLSTM network; and mining the internal relevance of each time sequence feature by using a self-attention mechanism to obtain a final output feature, namely a distributed photovoltaic short-term power prediction result. According to the method, meteorological features are screened through unified information coefficients, and prediction precision is improved in combination with factor analysis method dimension reduction, SCN network feature enhancement, BiLSTM time sequence modeling and a self-attention mechanism.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic control, and particularly relates to a short-term power prediction method and system for distributed photovoltaic based on BiLSTM. Background Art

[0002] The output of photovoltaic power generation is affected by meteorological factors such as actual irradiance, temperature, wind speed, and humidity, with large intermittency, randomness, and volatility, which have an impact on the power quality and power supply reliability of the power grid during grid connection. By accurately predicting photovoltaic power, power system managers can adjust the power generation plan of photovoltaic power stations or formulate a local photovoltaic consumption plan in advance, thereby achieving balanced dispatching of the power grid. However, the output of distributed photovoltaic has the characteristics of intermittency and randomness, which brings challenges to the economic operation dispatching and safe and stable operation of the distribution network connected to it. In order to ensure the safe and economic operation of the distribution network and promote the large-scale development of photovoltaic, it is very necessary to accurately predict the output of distributed photovoltaic.

[0003] Regarding the problem of short-term prediction of distributed photovoltaic power, the currently common photovoltaic power predictions mainly include: time series prediction, artificial neural network prediction, and machine learning prediction. Among them, the more common models for time series prediction include autoregressive moving average model (ARIMA), seasonal decomposition model, and exponential smoothing model, etc.; common artificial neural network predictions include recurrent neural network (RNN) and multi-layer perceptron (MLP), etc.; common machine learning predictions include decision tree, random forest, and support vector machine, etc. Different prediction methods are applicable to different scenarios and data sets.

[0004] Long short-term memory network (LSTM) has also been used for ultra-short-term prediction of photovoltaic power. LSTM can well handle the non-linear relationships in data and capture the long-term dependencies between data. Based on the LSTM model, other neural networks can also be incorporated to form a hybrid prediction model, and through the complementary advantages between different neural networks, more accurate multi-dimensional data prediction can be achieved. However, there are the following deficiencies in the current distributed photovoltaic output prediction method based on the LSTM model: (1) It is relatively sensitive to noise and abnormal data, and is prone to overfitting to time series noise in a long time series, and its robustness needs to be improved; (2) It has insufficient generalization ability for a large amount of high-dimensional historical photovoltaic data, and a large amount of feature engineering and parameter adjustment are required to achieve stable generalization ability. Summary of the Invention

[0005] The present invention aims to solve the problems of sensitivity to noise and abnormal data and insufficient generalization ability of high-dimensional data in existing photovoltaic power prediction methods, and provide a distributed photovoltaic short-term power prediction method and system based on BiLSTM. By screening meteorological features through unified information coefficient (UIC), combining factor analysis dimensionality reduction, SCN network feature enhancement, BiLSTM time series modeling and self-attention mechanism, the prediction accuracy and model robustness are improved, providing a reliable basis for power grid dispatching.

[0006] The present invention is implemented by the following technical solutions: A distributed photovoltaic short-term power prediction method based on BiLSTM, the steps are as follows: Step S1: Collect distributed photovoltaic output data on an hourly scale, and select meteorological feature vectors whose unified information coefficient with the distributed photovoltaic output is greater than a set threshold according to the unified information coefficient (UIC); Step S2: Using factor analysis to reduce the dimension of the selected meteorological feature vectors to obtain a factor score matrix; Step S3: Use the SCN network as a feature enhancer to perform nonlinear enhancement on the factor score matrix to obtain an enhanced feature sequence; Step S4: Using the enhanced feature sequence as input feature, the BiLSTM network is used to extract time series features; Step S5: Use the self-attention mechanism to mine the internal correlation of each time series feature to obtain the final output feature, which is the distributed photovoltaic short-term power prediction result.

[0007] Further preferably, in step S1, the mutual information coefficient between the meteorological characteristic vector and the distributed photovoltaic output vector is first calculated; then the meteorological characteristic vector and the distributed photovoltaic output vector are evenly divided into several segments according to a unified division method, and the unified information coefficient between the meteorological characteristic vector and the distributed photovoltaic output vector is calculated; the meteorological characteristic vectors whose unified information coefficient with the distributed photovoltaic output vector is greater than a set threshold are screened out and retained, and the remaining meteorological characteristic vectors are filtered and eliminated to form a meteorological characteristic vector screened by the unified information coefficient.

[0008] The process of reducing the dimension of the selected meteorological feature vector using factor analysis includes: First, the meteorological feature vectors screened by S1 are standardized to obtain a standardized meteorological feature matrix, and the Pearson correlation coefficient matrix is calculated based on the standardized meteorological feature matrix; Then, the Pearson correlation coefficient matrix is subjected to eigendecomposition to obtain eigenvalues and corresponding eigenvectors; the cumulative variance contribution rate of the eigenvalues is calculated, and the number of common factors is selected according to a preset cumulative variance contribution rate threshold; Finally, according to the number of common factors, the principal component analysis method is used to extract factors, convert the standardized meteorological characteristics into factor scores, and output the factor score matrix after dimensionality reduction. The factor score matrix is a linear combination of the original features. Through its incremental random learning mechanism and universal approximation property, the SCN network can efficiently learn the more complex non-linear mapping relationship between factor scores and target output power.

[0009] Further preferably, the calculation process of the factor score matrix is as follows: ; In the formula, L is the initial factor loading matrix, V is the rotation matrix determined by maximizing the variance of factor loadings; F is the factor score matrix, is a diagonal matrix containing input features, and each row of F is the factor score of a sample, is the rotated factor loading matrix, is the standardized meteorological characteristic matrix, is the transpose of.

[0010] Further preferably, collect distributed photovoltaic output data and corresponding meteorological data to form a training set and a test set; process the data of the training set according to steps S1 - S5, thereby training the distributed photovoltaic short-term power prediction model composed of the SCN network, BiLSTM network and self-attention mechanism, and then test it through the test set. Obtain meteorological data for a certain future period according to meteorological forecasts, perform dimensionality reduction according to step S2, and then input the qualified distributed photovoltaic short-term power prediction model for distributed photovoltaic short-term power prediction.

[0011] The present invention also provides a distributed photovoltaic short-term power prediction system based on BiLSTM, including a meteorological data acquisition module for collecting meteorological data; a distributed photovoltaic output acquisition module for collecting historical distributed photovoltaic output to be used for constructing a training set and a test set; a meteorological feature screening module for screening meteorological feature vectors whose unified information coefficient with distributed photovoltaic output is greater than a set threshold according to the unified information coefficient (UIC); a dimensionality reduction module that uses factor analysis method to reduce the dimensionality of the screened meteorological feature vectors to obtain a factor score matrix; a distributed photovoltaic short-term power prediction module that integrates a distributed photovoltaic short-term power prediction model composed of an SCN network, a BiLSTM network and a self-attention mechanism.

[0012] The present invention also provides an electronic device, including a memory and a processor. Computer-readable instructions are stored in the memory. When the instructions are executed by the processor, the processor is caused to implement the above-mentioned distributed photovoltaic short-term power prediction method based on BiLSTM.

[0013] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned distributed photovoltaic short-term power prediction method based on BiLSTM is implemented.

[0014] The present invention has the following technical effects: The unified information coefficient (UIC) is used to screen meteorological features strongly related to photovoltaic output, effectively capturing linear and non-linear relationships and reducing redundancy; factor analysis (FA) is combined to reduce the dimensionality of high-dimensional features, reducing the computational complexity and improving the model efficiency.

[0015] The SCN network is introduced as a feature enhancer, and the network structure is dynamically adjusted through a random configuration mechanism to enhance the non-linear representation ability of the input data and avoid the problem of gradient disappearance / explosion. Through its incremental random learning mechanism and universal approximation characteristics, the SCN network can efficiently learn more complex non-linear mapping relationships between factor scores and target output.

[0016] The BiLSTM network captures temporal dependencies bidirectionally, and the self-attention mechanism (SA) is combined to mine the associations between features, improving the modeling ability for complex meteorological-distributed photovoltaic output mapping relationships.

[0017] The collaborative work of multiple modules (UIC-FA-SCN-BiLSTM-SA) effectively suppresses noise interference and adapts to high-dimensional data. The model shows higher prediction accuracy and stability on both the training set and the test set. Description of the Drawings

[0018] Figure 1 is the flowchart of the method of the present invention.

[0019] Figure 2 are the prediction effects of three feature selection schemes. Detailed Embodiments

[0020] The following further describes the present invention in conjunction with embodiments. It is necessary to point out here that the following embodiments are only used to further illustrate the present invention and should not be construed as limiting the protection scope of the present invention. Some non-essential improvements and adjustments made by those skilled in the art according to the above-mentioned inventive content still fall within the protection scope of the present invention.

[0021] As Figure 1 shown, a distributed photovoltaic short-term power prediction method based on BiLSTM is as follows: Step S1: Collect the distributed photovoltaic output data on an hourly scale, and screen the meteorological feature vectors whose unified information coefficient with the distributed photovoltaic output is greater than the set threshold according to the unified information coefficient (UIC). Step S2: Use factor analysis to reduce the dimension of the selected meteorological feature vectors to obtain a factor score matrix. Step S3: Use the SCN network as a feature enhancer to perform non-linear enhancement on the factor score matrix to obtain an enhanced feature sequence. Step S4: Use the enhanced feature sequence as input features and use the BiLSTM network to extract temporal features. Step S5: Use the self-attention mechanism to mine the internal correlation of each temporal feature to obtain the final output feature, which is the short-term power prediction result of the distributed photovoltaic.

[0022] Compared with the widely used Pearson correlation coefficient method and Spearman correlation coefficient method, the unified information coefficient (UIC) can not only extract the linear correlation between variables, but also extract the non-linear correlation. Compared with the maximum information coefficient method, the unified information coefficient (UIC) can effectively avoid the influence of noise on the correlation analysis and has a lower calculation cost, which is suitable for feature dimension reduction with a large amount of data. The specific process of screening meteorological features whose unified information coefficient with the distributed photovoltaic output is greater than the set threshold according to the unified information coefficient (UIC) is as follows.

[0023] For the meteorological feature vector and the distributed photovoltaic output vector , the calculation method of their mutual information coefficient is as follows:[[]] (1); In formula (1), is the mutual information coefficient between the meteorological feature vector and the distributed photovoltaic output vector ; x is the characteristic element of the meteorological feature vector , y is the characteristic element of the distributed photovoltaic output vector , is the joint probability density function of the meteorological feature vector and the distributed photovoltaic output vector ; , are the marginal probability density functions of the meteorological feature vector and the distributed photovoltaic output vector respectively.

[0024] Then, divide the meteorological feature vector and the distributed photovoltaic output vector Evenly divided into several segments, which can be expressed as: (2); (3); Wherein, and are respectively the partition unit lengths of the meteorological feature vector and the distributed photovoltaic output vector ; and are respectively the maximum value and the minimum value of the meteorological feature vector ; and are respectively the maximum value and the minimum value of the distributed photovoltaic output vector ; and are respectively the number of segments of the meteorological feature vector and the distributed photovoltaic output vector ; represents the partition grid size, n is the data volume, is the grid coefficient, usually taken as the 0.6th power of the data volume.

[0025] According to this unified division method, the calculation formula of the unified information coefficient can be written as: (4); Wherein, is the unified information coefficient between the meteorological feature vector and the distributed photovoltaic output vector ; is and the minimum value in;

[0026] Based on formulas (1)-(4), calculate the unified information coefficient between each meteorological feature vector and the distributed photovoltaic output vector respectively, screen out the meteorological feature vectors whose unified information coefficient with the distributed photovoltaic output vector is greater than the set threshold of 0.4 for retention, and filter out the remaining meteorological feature vectors to form the meteorological feature vector after being screened by the unified information coefficient (UIC).

[0027] Factor analysis method (FA) is used to reduce the dimension of the meteorological feature vector to improve the efficiency and accuracy of the prediction model. The following are the detailed steps: First, standardize the meteorological feature vector screened in step S1 to eliminate the dimension difference, and obtain the standardized meteorological feature matrix , based on the standardized meteorological feature matrix Calculate the Pearson correlation coefficient matrix ; Then, perform eigenvalue decomposition on the Pearson correlation coefficient matrix to obtain the eigenvalues and corresponding eigenvectors; calculate the cumulative variance contribution rate of the eigenvalues, and select the number of common factors according to the preset cumulative variance contribution rate threshold; (5); In the formula, R is the Pearson correlation coefficient matrix; is the meteorological feature vector after being screened by the unified information coefficient (UIC), is the transpose of, is the diagonal matrix containing the eigenvalue , is the first eigenvalue, is the second eigenvalue, is the p-th eigenvalue; Calculate the cumulative variance contribution rate of each eigenvalue and select the number of factors: (6); Among them, k is the number of selected factors, is the cumulative variance contribution rate when the number of factors is k, and select the number of factors with a cumulative contribution rate > 85% as the number of common factors, is the i-th eigenvalue, and p is the total number of eigenvalues.

[0028] Finally, according to the number of common factors, use the principal component analysis method to extract factors, convert the standardized meteorological feature matrix into factor scores, and obtain the factor score matrix; (7); In the formula, L is the initial factor loading matrix, V is the rotation matrix determined by maximizing the variance of the factor loadings; F is the factor score matrix, is the diagonal matrix containing the input features, and each row of F is the factor score of a sample, is the rotated factor loading matrix, is the standardized meteorological feature matrix, is the transpose of.

[0029] The present invention uses the factor scores as the input of the distributed photovoltaic short-term power prediction model to replace the original meteorological features for distributed photovoltaic output prediction. The distributed photovoltaic short-term power prediction model consists of an SCN network, a BiLSTM network, and a self-attention mechanism.

[0030] The SCN network is a lightweight neural network based on the theory of random vector functional approximation. By randomly generating hidden layer nodes and dynamically adjusting the network structure, it can achieve efficient non-linear modeling and avoid the gradient vanishing / explosion problems of traditional neural networks through a unique random configuration mechanism. Through its incremental random learning mechanism and universal approximation characteristics, the SCN network can efficiently learn more complex non-linear mapping relationships between factor scores and target outputs.

[0031] (8); In the formula, j is the index of the hidden layer node, x is the data input to the SCN network, L is the number of hidden layer nodes, is the input weight vector of the j-th node, is the transpose of, is the bias of the j-th node, is the output weight of the j-th node (solved by the least squares method), is the activation function of the j-th node. In the time series prediction of distributed photovoltaic, the SCN network can be used as a feature enhancer to quickly extract non-linear features, thereby achieving non-linear enhancement.

[0032] Input the non-linear features into the forward LSTM layer to obtain the output of the forward LSTM layer. Then input the non-linear features into the backward LSTM layer. After obtaining the output of the backward LSTM layer, reverse the output again to get the output of the backward LSTM layer. Finally, linearly superimpose the output of the forward LSTM layer and the output of the backward LSTM layer according to a certain weight to obtain the final output result of the BiLSTM network. The calculation process of the BiLSTM network is shown in formula (9): (9); Among them, is the input feature at time t, is the hidden state of the forward LSTM layer at time t, is the hidden state of the forward LSTM layer at time t - 1; is the hidden state of the backward LSTM at time t, is the hidden state of the backward LSTM at time t - 1, and are the output weight matrices of the hidden states of the forward LSTM layer and the backward LSTM layer respectively; is the output of the BiLSTM network at time t, is the bias term, and LSTM(·) represents the calculation process of the LSTM cell.

[0033] The self-attention mechanism is a deep learning technique that mimics the human attention mechanism. Its role is to focus more on important parts when processing large-scale input data. In a neural network, the self-attention mechanism weights different parts of the input data through learned weights so that the network can better understand and utilize the information in the input data. This mechanism can improve the performance and generalization ability of the neural network, making it more suitable for handling various complex tasks. The calculation process of the self-attention mechanism is as follows: (10); In Equation (10): is the attention score of the i-th hidden unit state; is the weight matrix of the self-attention mechanism, is the bias term of the self-attention mechanism; is the value of the i-th hidden unit state; is the normalized weight coefficient of the i-th hidden unit state; is the output of the self-attention mechanism at time t, and T is the length of the input sequence.

[0034] In the present invention, it is necessary to collect distributed photovoltaic output data and corresponding meteorological data to form a training set and a test set; the data in the training set is processed according to steps S1 - S5, so as to train a distributed photovoltaic short-term power prediction model composed of an SCN network, a BiLSTM network, and a self-attention mechanism, and then test it through the test set. Meteorological data for a future period is obtained according to meteorological forecasts, and dimensionality reduction is performed according to step S2, and then it is input into the qualified distributed photovoltaic short-term power prediction model for distributed photovoltaic short-term power prediction.

[0035] In order to verify the effectiveness and superiority of the distributed photovoltaic short-term power prediction model proposed in the present invention, historical monitoring data of a certain distributed photovoltaic power station is selected for experimental verification. The original data of the example is selected as the annual data of a distributed photovoltaic power station in the north in 2024, and the time resolution is 1h. By selecting different prediction models for comparative analysis with the model proposed in the present invention, the effectiveness and superiority of the method proposed in the present invention are well verified.

[0036] In the experiment, 80% of the original data is used as the training set, and 20% is used as the test set. To evaluate the effect of the model, two indicators, root mean square error (RMSE) and mean absolute error (MAE), are used. RMSE is used to measure the deviation between the predicted value and the true value. The smaller the RMSE, the better the prediction effect. MAE represents the average distance between the predicted value and the true value. The smaller the MAE, the better the prediction effect.

[0037] (1) To explore the impact of input feature selection on the prediction accuracy of the model, three input feature selection schemes for the prediction model are set. Scheme 1 uses only historical distributed photovoltaic output data as input features; Scheme 2 uses all meteorological features as input for model training; Scheme 3 uses the feature screening method combining UIC and FA method proposed in the present invention to select input features. The prediction effects of the three feature selection schemes are as Figure 2 shown. For the features selected by Scheme 1, that is, only using the historical photovoltaic output data as the input of the prediction model, the error is the largest among the three schemes, and the RMSE and MAE are 4.02 and 2.02 respectively. For the features selected by Scheme 2, using all meteorological features as the input of the prediction model, its RMSE and MAE are 3.47 and 1.89 respectively. Compared with Scheme 1, they are reduced by 13.68% and 6.43% respectively, verifying the effectiveness of using multiple meteorological features. For Scheme 3, based on the feature screening method combining UIC and FA method proposed in the present invention to select the input of the prediction model, the average prediction error on the validation set is the smallest, and the RMSE and MAE are 3.12 and 1.70 respectively. Compared with Scheme 2 that selects all input features, they are reduced by 11.22% and 10.05% respectively, verifying the effectiveness of the input feature screening method proposed in the present invention.

[0038] (2) To evaluate the performance of the distributed photovoltaic short-term power prediction model (SCN-BiLSTM-SA) proposed in the present invention, the SCN model, BiLSTM model, and SCN-BiLSTM model are selected as comparison models. Among them, the other parameter settings of the comparison models, such as the number of iterations, training batch size, optimization algorithm, loss function, etc., are the same as those of the model proposed in the present invention.

[0039] Table 1 Comparison of prediction errors of different prediction models

[0040] As can be seen from Table 1, the SCN-BiLSTM-SA model proposed by the present invention has the best prediction performance. The RMSE of the single model BiLSTM is 4.23 kW, and the MAE is 1.98 kW. For SCN-BiLSTM, the RMSE is 3.42 kW and the MAE is 1.91 kW, with the RMSE increasing by 19.15% and the MAE increasing by 3.54%. It can be seen that the addition of the SCN layer improves the prediction ability of the BiLSTM model. At the same time, for SCN-BiLSTM-SA, the RMSE is 3.12 kW and the MAE is 1.70 kW, with the RMSE increasing by 8.77% and the MAE increasing by 10.99%. This further confirms that the prediction ability of the combined model is improved after introducing the attention mechanism. The experimental results show that the SCN-BiLSTM-SA prediction model proposed by the present invention can better mine the dependence relationship between distributed photovoltaic power output and historical meteorological feature sequences, effectively improving the prediction accuracy of distributed photovoltaic power output.

[0041] The second embodiment of the present invention provides a distributed photovoltaic short-term power prediction system based on BiLSTM, including a meteorological data acquisition module for collecting meteorological data; a distributed photovoltaic power output acquisition module for collecting historical distributed photovoltaic power output for constructing a training set and a test set; a meteorological feature screening module for screening meteorological feature vectors with a unified information coefficient greater than a set threshold with respect to the distributed photovoltaic power output according to the unified information coefficient (UIC); a dimensionality reduction module that uses factor analysis to reduce the dimensionality of the screened meteorological feature vectors to obtain a factor score matrix; a distributed photovoltaic short-term power prediction module that integrates a distributed photovoltaic short-term power prediction model composed of an SCN network, a BiLSTM network, and a self-attention mechanism.

[0042] The third embodiment of the present invention provides an electronic device, including a memory and a processor. The memory stores computer-readable instructions. When the instructions are executed by the processor, the processor is caused to implement the above-mentioned distributed photovoltaic short-term power prediction method based on BiLSTM.

[0043] The fourth embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned distributed photovoltaic short-term power prediction method based on BiLSTM is implemented.

[0044] The above only expresses the preferred embodiments of the present invention and does not limit the present invention in other forms. Any person skilled in the relevant art may use the disclosed content above to make changes or modifications into equivalent embodiments with equivalent changes. However, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A short-term power prediction method for distributed photovoltaic based on BiLSTM, characterized in that, The steps are as follows: Step S1: Collect distributed photovoltaic output data on an hourly scale, and screen meteorological feature vectors whose unified information coefficient with the distributed photovoltaic output is greater than a set threshold according to the unified information coefficient; Step S2: Use factor analysis to reduce the dimension of the screened meteorological feature vectors to obtain a factor score matrix; Step S3: Use the SCN network as a feature enhancer to non-linearly enhance the factor score matrix to obtain an enhanced feature sequence; Step S4: Use the enhanced feature sequence as input features and use the BiLSTM network to extract time series features; Step S5: Use the self-attention mechanism to mine the internal correlation of each time series feature to obtain the final output feature, which is the short-term power prediction result of the distributed photovoltaic; 2. The distributed photovoltaic short-term power prediction method based on BiLSTM according to claim 1, wherein In step S1, first calculate the mutual information coefficient between the meteorological feature vector and the distributed photovoltaic output vector; then evenly divide the meteorological feature vector and the distributed photovoltaic output vector into several segments according to the unified division method, and calculate the unified information coefficient between the meteorological feature vector and the distributed photovoltaic output vector; screen and retain the meteorological feature vectors whose unified information coefficient with the distributed photovoltaic output vector is greater than the set threshold, and filter and eliminate the remaining meteorological feature vectors to form the meteorological feature vectors after screening by the unified information coefficient.

3. The short-term power prediction method for distributed photovoltaic based on BiLSTM according to claim 1, characterized in that, The process of using factor analysis to reduce the dimension of the screened meteorological feature vectors includes: First, standardize the meteorological feature vectors screened in S1 to obtain a standardized meteorological feature matrix, and calculate the Pearson correlation coefficient matrix based on the standardized meteorological feature matrix; Then, perform eigenvalue decomposition on the Pearson correlation coefficient matrix to obtain eigenvalues and corresponding eigenvectors; calculate the cumulative variance contribution rate of the eigenvalues, and select the number of common factors according to the preset cumulative variance contribution rate threshold; Finally, according to the number of common factors, use the principal component analysis method to extract factors, convert the standardized meteorological features into factor scores, and output the reduced-dimensional factor score matrix.

4. The short-term power prediction method for distributed photovoltaic based on BiLSTM according to claim 3, wherein, The calculation process of the factor score matrix is as follows: ； In the formula, L is the initial factor loading matrix, V is the rotation matrix determined by maximizing the variance of the factor loadings, F is the factor score matrix, is a diagonal matrix containing the input features, and each row of F is the factor score of a sample, is the rotated factor loading matrix, is the standardized meteorological feature matrix, is the transpose of.

5. The distributed photovoltaic short-term power prediction method based on BiLSTM according to claim 3, characterized in that Collect distributed photovoltaic output data and corresponding meteorological data to form a training set and a test set; process the data of the training set according to steps S1 - S5 to train the distributed photovoltaic short-term power prediction model composed of the SCN network, the BiLSTM network and the self-attention mechanism, and then test it through the test set. Obtain meteorological data for a certain future period according to the weather forecast, perform dimensionality reduction according to step S2, and then input it into the qualified distributed photovoltaic short-term power prediction model for short-term power prediction of the distributed photovoltaic.

6. A system for implementing the distributed photovoltaic short-term power prediction method based on BiLSTM according to claim 1, characterized in that, It includes: A meteorological data acquisition module for collecting meteorological data; A distributed photovoltaic output acquisition module for collecting historical distributed photovoltaic output; A meteorological feature screening module for screening meteorological feature vectors whose unified information coefficient with the distributed photovoltaic output is greater than a set threshold according to the unified information coefficient; A dimensionality reduction module that uses factor analysis to reduce the dimension of the screened meteorological feature vectors to obtain a factor score matrix; Distributed photovoltaic short-term power prediction module, integrating a distributed photovoltaic short-term power prediction model composed of an SCN network, a BiLSTM network, and a self-attention mechanism.

7. An electronic device, comprising a memory and a processor, wherein computer-readable instructions are stored in the memory, characterized in that, When the instruction is executed by the processor, it causes the processor to implement the BiLSTM-based distributed photovoltaic short-term power prediction method according to any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the BiLSTM-based distributed photovoltaic short-term power prediction method according to any one of claims 1-5.

Images