Research On Combined Forecasting Methods Of Short-term Load In Power System

Electric energy is very important to the national life and economic development of any country.Because the electric energy can not be stored in large quantities,the power system must keep the balance of supply and demand at any time during its operation,and provide reliable and standard electric energy to all kinds of users.Insufficient power supply will bring negative impact on social and economic growth,but also affect social stability and people's life.Therefore,in order to ensure the safety and economic operation of the power system,it is necessary to master the law and trend of load change,and accurately predict the change and characteristics of load demand.Accurate load forecasting plays an important role in the operation of power system,and effective load forecasting method can improve the operation efficiency of power system:This paper aims to improve the short-term load forecasting accuracy of the power system,and improves the forecasting method from the perspectives of transfer learning,deep learning,and combined forecasting model,so as to improve the load forecasting accuracy.Specifically,it includes the following aspects:1.Because of the significant correlation between the power load data in nearby areas,transfer learning can be applied to power load forecasting to improve the accuracy of load forecasting.However,due to the fact that time series prediction is not exactly the same as the traditional data regression problem,the general transfer learning method may bring negative transfer to load forecating and reduce the accuracy of load forecating.Therefore,this paper proposes a combined short-term load forecasting method based on seasonal decomposition and transfer learning of time series.First of all,using a time series seasonal decomposition approach,the power system load data is decomposed into three parts:seasonal components,trend components and irregular components.Then the traditional machine learning method is applied on the trend and seasonal components,in order to better explain the seasonal cycle of power load data.Secondly,a two-stage transfer learning approach is applied on both the source domain and target domain of irregular components to forecast the irregular component of the target domain of,in order to improve prediction accuracy.Through the prediction analysis of two real power data sets in the United States and Australia,it is verified that this method can effectively avoid negative transfer and improve the load prediction accuracy2.In order to further improve the effectiveness of transfer learning,knowledge transfer can be carried out from multiple source domains to increase the chances of finding samples closely related to the target domain.Based on this,this paper proposes a multiple similarity and multi-source instance transfer algorithm based on inter-domain similarity and inter-sample similarity,and constructs an improved Bagging based resampling transfer regression framework.This combined short-term load forecasting method uses multiple similarity to migrate more samples with higher similarity from multiple source domains.Resampling method can be used to select the combination of transfered data and target domain data,which can make more effective use of source domain data and target domain data,thus improving the accuracy of load forecasting.Experimental evaluation of two real data sets in the United States and Australia shows that trasfering data with higher similarity to the target domain from more data sources can significantly improve prediction accuracy and effectively avoid negative transfer.3.A combined short-term load forecasting approach based on a multi-head attention is also proposed.A seasonal and trend decomposing technique is used to preprocess the original electrical load data.Each decomposed datum is regressed to predict the future electric load value by utilizing the encoder–decoder network with the multi-head attention mechanism.Through multi-head attention mechanism,this method can more easily capture the characteristics of medium and long distance interdependence in the load data,and make better preparation for feature extraction of the training model,so as to effectively improve the prediction accuracy.The simulation results on a real power data sets in the United States show that the proposed method is superior to the the other five counterpart models based on traditional machine learning or nural network approachs.At the same time,the simulation experiments on Australian power dataset demonstrate that the prediction accuracy of this method is better than the other two combined load forecasting methods based on transfer learning proposed in this paper.