Forecasting Of Short-term Power Load Based On KPCA-WOA-BiLSTM Model

The stable operation of the power system is related to the national economy and people’s livelihood.Accurate forecasting of power load is a necessary condition to realize the stable operation of power grid.The power load data have characteristics of time series and are affected by various nonlinear factors such as meteorological conditions.Neural network is a popular tool in the field of short-term load forecasting.But due to the complex historical data,traditional neural network forecasting models often encounter the problem of variable redundancy and convergence difficulties,and the forecasting accuracy is often unsatisfactory.In view of this situation,the KPCA-WOA-Bi LSTM model is improved in this study based on Bidirectional Long Short-Term Memory(Bi LSTM)neural network which has the strong temporal data processing ability for the purpose of accurately forecasting short-term power load.The improvement measures are divided into two steps: Firstly,the Whale Optimization Algorithm(WOA)is used to optimize the super parameters of the model,which can avoid the poor prediction accuracy caused by unreasonable parameter setting;Secondly,the Kernel Principal Component Analysis(KPCA)method is used to reduce the dimensions of the impact factors of high dimensions in the original data,which can avoid the difficulty of model convergence caused by the high dimension of input data.In this study,the historical load data of a county in North China were selected for prediction experiment.Firstly,BP(Back Propagation),Long Short-Term Memory(LSTM)and Bi LSTM neural network models were used to predict,respectively.The results showed that the mean absolute percentage error of Bi LSTM model was 3.72%,which was obviously superior to the other two models.This suggested that the improvements based on the Bi LSTM model in this study had certain advantages.Secondly,WOA and KPCA algorithms were used to optimize Bi LSTM model in two steps,and the improved WOA-Bi LSTM and KPCA-WOA-Bi LSTM models were used to predict.The results showed that the mean absolute percentage error of the two models decreased by 0.79% and 1.61%,respectively,compared with those before the improvement.The effectiveness of the improved methods was demonstrated.Finally,the same improvement methods were used to optimize the basic BP and LSTM models,respectively.The prediction results of nine models,including BP,WOA-BP,KPCA-WOA-BP,LSTM,WOA-LSTM,KPCA-WOA-LSTM,Bi LSTM,WOA-Bi LSTM and KPCA-WOA-Bi LSTM,were compared comprehensively.The results showed that the forecasting accuracy of the three basic neural network models was significantly improved after the improvements,which further proved the effectiveness of the optimization methods.The mean absolute error of KPCA-WOA-Bi LSTM model was4.97 MW,the mean absolute percentage error was 2.11%,and the root mean square error was 5.38 MW.These key performance indexes were significantly better than other models,which further verified the excellent forecasting performance of KPCA-WOA-Bi LSTM model.It was also proved that this model had certain application value in the actual short-term load forecasting task.