| In the power system,short-term power load forecasting plays an important role,and its forecast results directly affect the rational use of energy and the balance of power supply and demand.Therefore,it is becoming more and more important to construct an accurate,efficient and stable model to complete short-term power load forecasting.With the rapid development of social economy and the change of global ecological environment,the factors that affect power load are becoming more and more complex,the single model cannot predict the changes of modern power load.Therefore,the network model of deep learning,data decomposition and feature extraction are used to study the short-term power load forecasting.Firstly,the paper analyzes RNN,LSTM and GRU model.Based on LSTM,GRU was added as an improvement.By comparing the fitting effect,prediction accuracy and training duration of LSTM and GRU,it is proved that GRU has obvious advantages in time complexity and fitting effect.when the input load sequence is long,the GRU will lose the sequence information when processing the information of the structure between the data,thus affecting the accuracy of the model prediction.Therefore,on the basis of GRU and the Attention mechanism,a GRU-Attention prediction model with more prominent key features and independent of sequence length Was established.Meanwhile,According to the nonlinear and non-stationary characteristics of power load data,the load data is fitted into random fluctuation signals.In order to reconstitute the load sequence into a certain number of signals with different scales and relative stability,so as to indirectly improve the quality of power load data,this paper further introduces EMD and establishes a prediction model of EMD-GRU-Attention.In order to avoid the problems of too many components and mode mixing in the EMD reconstructed components,on the one hand,this paper combined as many low-frequency components as possible with Res,and on the other hand,used CNN to further extract hidden features of high-frequency components,constructed the EMD-CNN-GRU-Attention prediction model.Finally,in order to verify the feasibility and effectiveness of the model proposed in this paper,the power load data of western Denmark in 2016 are used to call and drive graphics card resources by using CUDA and cu DNN in the Spyder integrated development tool under the Anaconda software management environment.The network structure and code of the prediction model were designed and written based on Keras library function of Tensorflow-GPU framework,and the simulation experiment was carried out.By comparing and analyzing the prediction results of different models of GRU-Attention,EMD-GRU-Attention and EMD-CNN-GRU-Attention,it is shown that the model proposed in this paper has higher prediction accuracy and is an efficient shortterm load forecasting method. |
