Research On Short-term Power Load Forecasting Based On FFT-CNN-LSTM

With the acceleration of the smart grid process,the grid scale and structure are becoming increasingly complex,which undoubtedly increases the difficulty of shortterm power load forecasting.At present,a single traditional classic forecasting method or modern forecasting method has more or less the characteristics of low forecast accuracy and weak generalization ability in short-term load forecasting.From the perspective of deep learning,this paper builds a multi-step output prediction method of combined neural network FFT-CNN-LSTM based on the real power load data of New South Wales,Australia,using the advantages of different algorithms and neural networks to improve short-term load forecasting performance.The main contents of this paper are as follows:First of all,from the perspective of the study of the periodic characteristics of the electric load and its influencing factors,the periodic characteristics of the load sequence itself are analyzed,including daily,weekly,seasonal and annual periodic characteristics,and the influence of meteorological factors,electricity prices,and date types on the electric load is studied.And based on the analysis of Peason’s correlation coefficient,a multi-influencing factor input feature set incorporating temperature and humidity,electricity price,date type and historical load is constructed.Secondly,based on the XGBoost and LSTM neural network,the XGBoost load forecasting model and the improved LSTM-2 forecasting model with double hidden layers are built,and the comparison of examples verifies that LSTM-2 is compared with XGBoost,RNN and standard LSTM in short-term load forecasting.There are certain advantages in performance.In order to further improve the prediction performance of a single LSTM neural network,this paper proposes a CNN-LSTM composite neural network model,that is,a convolutional neural network is introduced before the LSTM neural network to obtain more useful advanced features in the input data,and then the LSTM is fully considered The time series characteristics of load characteristics,experiments prove that the CNNLSTM composite neural network has the advantages of both CNN and LSTM and has a larger prediction performance improvement compared with the LSTM-2 model.Finally,combined with Fourier(FFT)decomposition method,a combination model prediction method of FFT-CNN-LSTM is proposed.The load sequence is decomposed by FFT into periodic components,low-frequency components and highfrequency components with different changing laws,and then a single component is predicted by CNN-LSTM to reconstruct the prediction results.The experimental results of the calculation examples show that the FFT-CNN-LSTM model is good at grasping the differences The changing law of the components further improves the short-term load prediction accuracy of CNN-LSTM,and proves the effectiveness of the prediction algorithm for short-term load prediction.