Short-Term Load Forecasting Based On Long Short-Term Memory Network And Bayesian Optimization

Short-term load forecasting is very important to maintain the balance between power production and consumption of the power grid.It not only affects the safety and reliability of power generation in the power system,but also affects the economic development of the power grid.With the reform of the power market and the construction of smart grids,short-term load forecasting has become one of the most important tasks in power system management.However,traditional load forecasting methods have certain limitations when dealing with non-linear load sequences.The emergence of artificial intelligence algorithms has made up for this defect.In this paper,a hybrid load forecasting model is constructed based on convolutional neural network,bidirectional long short-term memory network,Bayesian optimization algorithm and attention mechanism.Introducing the classification of power load and load forecasting according to power usage and forecast range.Focusing on the characteristics and influencing factors of short-term load.Calculating the correlation coefficient between the influencing factors and the load.And selecting the influencing factors with a high degree of correlation as the input features of the forecasting model.Subsequently,a bidirectional long short-term memory network is used for short-term load forecasting.The neural network selectively forgets or updates the state by means of memory units,so as to deal with the long-term dependence of the load sequence.In order to improve the forecasting performance and training speed of the model,Bayesian optimization is also used to automatically tune the model’s hyperparameters.In actuality,a single model cannot accurately forecast the actual load due to the interference of various complex factors.In order to improve the stability and accuracy of the model,the convolutional neural network and attention mechanism are introduced to the bidirectional long short-term memory network and Bayesian optimization.The convolutional neural network can amplify the significant features of the input data because of the unique convolution structure,so it is good at feature extraction.The attention mechanism can help the model to focus on important information by assigning weight probability.The hybrid model combines the advantages of several advanced algorithms.In order to measure the forecasting performance of the hybrid model proposed in this paper,the real load data and meteorological factor data of a city in North China are collected for a whole year.Using the hybrid model and other contrast models to perform daily load forecasting and weekly load forecasting on the basis of this data set.Comparing the forecasting performance among those models by MAE,RMSE,MAPE and R~2 score.The results of each simulation experiment show that the model proposed in this paper has the highest forecast accuracy and the best degree of fit.