Application Of Residual Self-correction Deep-learning Integrated Neural Network In Short-term Load Forecasting

With the development of power systems,load forecasting is playing an increasingly important role.Accurate load forecasting effectively improves the stability and reliability of power systems.Therefore,it is of great significance to their operation and maintenance.Load data of power systems exhibits certain properties in terms of both space and time.Spatially,it is easily affected by various external factors,such as weather conditions,holidays,etc.Load data belongs to a kind of time series data,and the load data at each specific moment is closely connected to the related historical data,which is highly valuable as references for prediction.Accurate load prediction requires both the spatial and time properties to be taken into consideration.Unfortunately,traditional load predicting models are incapable to take care of them both at the same time.In comparison,time-series predicting model based on deep learning is a good choice to handle this problem.Moreover,correcting the residual of predicting results is helpful in further improving the accuracy of the model.Residuals are difficult to fit with traditional residual correcting methods,but deep learning is able to extract useful information from the disordered residual data,so that the data can be well fitted and predictedBased on the above analysis,this paper presents research about related issues,The main research contents include:A new missing data supplementing method and a new anomaly data detecting method are proposed according to analysis about the properties of load data.According to the characteristics of load data,the defects of the traditional load forecasting methods are analyzed Considering deep learning timing models' ability of memorizing and automatically selecting historical information,load forecasting prediction models based on Long Short Term Memory(LSTM)and Gated Recurrent Unit(GRLU)are established.In the experiment,LSTM and GRU are compared with traditional prediction models,and the advantages of LSTM and GRU are analyzed.Considering the disorder of the residual time series,the defects of the existing residual correction method are analyzed.Then the residual prediction models based on LSTM and GRU are established.Based on the idea of ensemble learning,the load forecasting models and the residual prediction models are combined into residual self-correcting load forecasting models.A linear regression model is established to search for the linear relationship between the real load results and the load forecasting results of multiple residual self-correcting load forecasting models.The prediction results of multiple models are assigned reasonable weights and linearly combined.Experimental results show that the results of the linear combination are more accurate than a single prediction.