Short-term photovoltaic power generation power prediction method based on BO-TCN-BiLSTM

Abstract

A short-term photovoltaic power generation power prediction method based on BO-TCN-BiLSTM comprises the steps that firstly, an original photovoltaic sequence is preprocessed, abnormal values are removed, processed data are divided into different similar daily sets, then a TCN-BiLSTM model is used for extracting features in the different similar daily sets, the TCN-BiLSTM model can better extract local and time sequence features of the photovoltaic sequence, and the local and time sequence features of the photovoltaic sequence can be predicted through the TCN-BiLSTM model. In the process, a BO algorithm is used for optimizing hyper-parameters of the TCN-BiLSTM model, and the prediction accuracy of the model is further improved; local features of a photovoltaic sequence are obtained through causal convolution and extended convolution of TCN, a long-term dependency relationship of the sequence is captured through a bidirectional structure of a BiLSTM network, hyper-parameters such as convolution kernel size, learning rate and LSTM unit number of a model are optimized by using a BO algorithm, and generalization ability of the model is enhanced through feature learning of different levels. And the prediction performance of the model is improved.

Description

Technical Field

[0001] The present invention relates to a short-term photovoltaic power prediction technology, and specifically refers to a short-term photovoltaic power prediction method combining a bidirectional long short-term memory network (BiLSTM), a temporal convolutional network (TCN) and a Bayesian optimization algorithm (BO), which is referred to as BO-TCN-BiLSTM. Background Art

[0002] With the rapid development of smart grid and renewable energy technologies, accurate prediction of photovoltaic power generation has become particularly critical. Traditional prediction methods often fail to fully consider the temporal characteristics and nonlinear relationships of data, resulting in insufficient prediction accuracy. To overcome these limitations, this study proposes a method that integrates bidirectional long short-term memory network (BiLSTM), temporal convolutional network (TCN) and Bayesian optimization algorithm (BO) to improve the accuracy and robustness of short-term photovoltaic power generation prediction. By combining the advantages of BiLSTM in processing sequence data and the ability of TCN in capturing long-term dependencies, this method can effectively predict photovoltaic power generation and provide strong technical support for the optimal scheduling and energy management of power grids. Summary of the invention

[0003] The purpose of the present invention is to provide a short-term photovoltaic power prediction method based on BO-TCN-BiLSTM.

[0004] The present invention is a short-term photovoltaic power generation prediction method based on BO-TCN-BiLSTM, and the steps are: first pre-process the original photovoltaic sequence, remove outliers, divide the processed data into different similar day sets, and then use the TCN-BiLSTM model to extract the features of different similar day sets respectively. The TCN-BiLSTM model can better extract the local and temporal features of the photovoltaic sequence. In this process, the BO algorithm is used to optimize the hyperparameters of the TCN-BiLSTM model to further improve the prediction accuracy of the model. The specific steps are: Step (i) Data preprocessing: Collect photovoltaic power data sets, and perform outlier processing and missing value filling. In the present invention, the mean square method is used to process and fill the outliers and missing values ​​in the data set. The specific formula of the method is expressed as follows: Where N represents the total number of samples in the data base, x i represents the power value at the i-th moment; Secondly, the processed data is divided into similar days, and the Pearson correlation coefficient is used to divide different similar day sets. The method is expressed as follows: Among them, x i and i Respectively represent the values ​​of the x variable and the y variable at the i-th moment, and represents the average of x and y, and n represents the total number of the entire sample. Step (ii) Feature extraction: The TCN-BiLSTM model is used to extract features of similar daily sets. The present invention uses the parallel convolution capability of TCN and the sequence modeling capability of BiLSTM to capture the short-term and long-term dependencies of the sequence. The TCN layer extracts local features of the sequence through causal convolution and dilated convolution. The method is specifically expressed as follows: Among them, ω(k) represents the weight of the convolution kernel. Step (iii) Feature fusion and parameter optimization: The feature information extracted from different similar days is fused, and the hyperparameters of the TCN-BiLSTM model are optimized using the BO algorithm. Step (iv) uses the training set data for training and uses the optimized model for prediction, and the root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R 2 ) is used to evaluate the performance of the model. It is expressed as:

[0005] The benefits of the present invention are as follows: the local features of the photovoltaic sequence are obtained through the causal convolution and dilated convolution of TCN, the long-term dependencies of the sequence are captured through the bidirectional structure of the BiLSTM network, and the hyperparameters of the model such as the convolution kernel size, the learning rate and the number of LSTM units are optimized by the BO algorithm. Through feature learning at different levels, the generalization ability of the model is enhanced and the prediction performance of the model is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 This is the block diagram of the BO-TCN-BiLSTM technology system. Figure 2 This is a detailed diagram of the BiLSTM layer. Figure 3 This is a detailed diagram of the TCN layer. Figure 4 For the comparison and ablation experiments between the model and other baseline models under different prediction indicators, Figure 5 A comparative experiment of power sampling between the BO-TCN-BiLSTM model and other baseline models on the same dataset is shown. DETAILED DESCRIPTION

[0007] Model training and prediction: The training method used in this invention includes using labeled training data to fine-tune the parameters of the model, and using advanced optimization algorithms to ensure the optimization of the model parameters. Once the model training is completed, it will be used for the task of predicting short-term photovoltaic power generation. By inputting new photovoltaic power generation data, the model can accurately predict the power output in the upcoming time period, thereby providing valuable information for grid operators to facilitate their effective energy management and scheduling. This prediction capability is crucial for integrating photovoltaic energy into smart grids and optimizing the operation and maintenance of photovoltaic power generation systems.

[0008] The present invention proposes a short-term photovoltaic power prediction method based on BO-TCN-BiLSTM, aiming to overcome the problems of gradient vanishing and gradient explosion in traditional methods.

[0009] The present invention is a short-term photovoltaic power generation prediction method based on BO-TCN-BiLSTM, and the steps are: first pre-process the original photovoltaic sequence, remove outliers, divide the processed data into different similar day sets, and then use the TCN-BiLSTM model to extract the features of different similar day sets respectively. The TCN-BiLSTM model can better extract the local and temporal features of the photovoltaic sequence. In this process, the BO algorithm is used to optimize the hyperparameters of the TCN-BiLSTM model to further improve the prediction accuracy of the model. The specific steps are: Step (i) Data preprocessing: Collect photovoltaic power data sets, and perform outlier processing and missing value filling. In the present invention, the mean square method is used to process and fill the outliers and missing values ​​in the data set. The specific formula of the method is expressed as follows: Where N represents the total number of samples in the data base, x i represents the power value at the i-th moment; Secondly, the processed data is divided into similar days, and the Pearson correlation coefficient is used to divide different similar day sets. The method is expressed as follows: Among them, x i and i Respectively represent the values ​​of the x variable and the y variable at the i-th moment, and represents the average of x and y, and n represents the total number of the entire sample. Step (ii) Feature extraction: The TCN-BiLSTM model is used to extract features of similar daily sets. The present invention uses the parallel convolution capability of TCN and the sequence modeling capability of BiLSTM to capture the short-term and long-term dependencies of the sequence. The TCN layer extracts local features of the sequence through causal convolution and dilated convolution. The method is specifically expressed as follows: Among them, ω(k) represents the weight of the convolution kernel. Step (iii) Feature fusion and parameter optimization: The feature information extracted from different similar days is fused, and the hyperparameters of the TCN-BiLSTM model are optimized using the BO algorithm. Step (iv) uses the training set data for training and uses the optimized model for prediction, and the root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R 2 ) is used to evaluate the performance of the model. It is expressed as:

[0010] The present invention is described in detail below with reference to the accompanying drawings and specific implementation examples. 1. Basic Assumptions Data quality and completeness: It is assumed that the input data (such as historical photovoltaic power generation, meteorological data, etc.) is accurate, complete, and consistent in time. Missing values ​​or noise in the data should be properly handled to ensure the reliability of model training and prediction. Stationarity of time series: Although short-term photovoltaic power generation may have certain volatility, the model usually assumes that the time series is relatively stable in the short term, that is, the trend and seasonal changes are not too drastic. For long-term series, stabilization may be required. Feature selection: It is assumed that the selected features (such as meteorological data, historical power data, etc.) have a strong relevance to PV power prediction and can provide useful information for model training and prediction. Influence of external factors: It is assumed that in the short term, the impact of external factors such as environmental changes (such as temperature, humidity, light intensity, etc.) on photovoltaic power generation can be captured and processed by the model. Prediction target stability: Assuming that the law of change in photovoltaic power generation in the short term is predictable, the model can accurately predict the power for a period of time in the future based on historical data. The technology of the present invention integrates the advantages of bidirectional long short-term memory network (BiLSTM) and gated temporal convolutional network (TCN), which can efficiently analyze the time series characteristics and nonlinear dynamics in photovoltaic power data, and significantly improve the accuracy of prediction. After a series of experimental tests and verifications, the BO-TCN-BiLSTM model has demonstrated performance that exceeds the current technical level in multiple prediction evaluation indicators including root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE). This achievement not only proves the superiority of the model in dealing with photovoltaic power generation prediction problems, but also provides strong technical support for the wider application of photovoltaic energy and the optimized management of smart grids.

[0011] Figure 1 A framework diagram showing the technology of the present invention, wherein the convolutional layer of TCN is used to extract local features of a sequence and can process multi-scale features, BiLSTM can consider both past and future information, and the BO algorithm is used to optimize the hyperparameters of TCN and BiLSTM.

[0012] Figure 3 The structure of the TCN layer is described in detail. TCN mainly captures long-range dependencies through dilated convolution. Dilated convolution introduces a "dilation factor" in the application of the convolution kernel, allowing the convolution kernel to skip the intermediate elements, thereby expanding the receptive field. The residual connection is used to add the input of the convolution layer directly to the output of the convolution layer, helping the model to learn the incremental features of the convolution layer, improve training efficiency, and reduce the problem of gradient disappearance. Multiple convolution layers are stacked together to extract features of different scales of time series layer by layer. Each layer of convolution operation can introduce different dilation factors to capture the multi-level dependencies of the time series.

[0013] Figure 2 The structure of the BiLSTM layer is described in detail. The input photovoltaic sequence [x 1 ,x 2 ,x 3 ,...,x T ], input into the forward LSTM network from left to right, and the forward LSTM layer outputs a hidden state sequence The input data is input into the backward LSTM layer, and a hidden state sequence is also obtained. The forward and backward LSTM layers are connected in the connection layer to form a final state

[0014] Figure 4 Comparison and ablation experiments of the BO-TCN-BiLSTM model and other baseline models at different time scales are shown.

[0015] Through the description of the above implementation modes, those skilled in the art can implement the methods described in the embodiments of the present invention through software, in part of the contribution of the present invention to the prior art.

Claims

1. A short-term photovoltaic power prediction method based on BO-TCN-BiLSTM, characterized in that: The steps are as follows: first, preprocess the original photovoltaic sequence, remove outliers, divide the processed data into different similar day sets, and then use the TCN-BiLSTM model to extract the features of different similar day sets. The TCN-BiLSTM model can better extract the local and temporal features of the photovoltaic sequence. In this process, the BO algorithm is used to optimize the hyperparameters of the TCN-BiLSTM model to further improve the prediction accuracy of the model. The specific steps are as follows: Step (i) Data preprocessing: Collect photovoltaic power data sets, and perform outlier processing and missing value filling. In the present invention, the mean square method is used to process and fill the outliers and missing values ​​in the data set. The specific formula of the method is expressed as follows: Where N represents the total number of samples in the data base, x i represents the power value at the i-th moment; Secondly, the processed data is divided into similar days, and the Pearson correlation coefficient is used to divide different similar day sets. The method is expressed as follows: Among them, x i and i Respectively represent the values ​​of the x variable and the y variable at the i-th moment, and represents the average value of x and y, and n represents the total number of the entire sample; Step (2) Feature extraction: Use the TCN-BiLSTM model to extract features from similar daily sets. Use the parallel convolution capability of TCN and the sequence modeling capability of BiLSTM to capture the short-term and long-term dependencies of the sequence. The TCN layer extracts local features of the sequence through causal convolution and dilated convolution. The specific method is as follows: Among them, ω(k) represents the weight of the convolution kernel; Step (iii) Feature fusion and parameter optimization: The feature information extracted from different similar days is fused, and the hyperparameters of the TCN-BiLSTM model are optimized using the BO algorithm; Step (iv) Use the training set data for training and use the optimized model for prediction, and use the root mean square error RMSE, mean absolute error MAE and determination coefficient R 2 The performance of the model is evaluated; it is expressed as:

Images