Short-Term Power Load Forecasting Based On Improved Multi-BiLSTM-Attention Neural Network

In the power system,electrical load forecasting is an significant foundation for power supply arrangement and machine maintenance.As an significant part of the electricity supply management system,electrical load forecasting straightly influences the analysis results of subsequent power security inspections,and plays a key role in load scheduling,so it is particularly significant to create a electrical load forecasting model with efficient and accurate forecasting performance.According to the principle and algorithm of artificial neural network,this thesis gradually creates a short-term electrical load forecasting model.Considering the impact of the characteristic changes of date and environmental factors on load data forecasting,an attention mechanism(Attention)is introduced on the basis of Bidirectional Long Short-Term Memory Neural Network(BiLSTM),analyze the advantages of multi-layer and multi-dimensional(Multi),and optimize the number and dimension of the hidden layer respectively,that is,update the number of hidden layers from a single layer to multi-layer and the size of the hidden layer from a single dimension to multi-dimension.After that,an improved particle swarm optimization algorithm(IPSO)is introduced to enhance the model parameters.The short-term electrical load of IPSO-Multi-BiLSTM-Attention neural network is structured.The main contents of this article are as follows:First,preprocess the data,introduce the principles and algorithms of various neural network models,find out the shortcomings of different models through comparison,and make improvements.After the continuous optimization and improvement,the selected target model is the multi-layer and multi-dimensional of IPSO-Multi-BiLSTM-Attention neural network model.Then,a simulation experiment environment is established to effect the computer simulation measurement and control of the artificial neural network model,also the loss curve of the training set and the test set,with a comparison chart of the predicted value for the test set and the real value curve is obtained.Finally,the data results obtained by the models are scientifically analyzed through the evaluation indicators,and the optimal neural network prediction model is selected.Compare the prediction results of the modified IPSO-Multi-BiLSTM-Attention model with the other five models.The result is that the target model has better accuracy in terms of Mean Squared Error(MSE),Mean Absolute Percentage Error(MAPE),Root Mean Squared Error(RMSE),Mean Absolute Error(MAE),with Coefficient of Determination(R~2),and the error is smaller.It is proved that the IPSO-Multi-BiLSTM-Attention neural network has the characteristics of objectivity and practicability,also can meet the requirements of short-term electrical load forecasting.