Load Forecasting Research Based On Deep Learning In Smart Grid

In recent years,with the rapid development of China’s economy and the continuous improvement of people’s living standards,both industrial production and daily life require more and more electricity.The change of energy production and consumption mode brings more uncertainly to power grid,and demands higher reliability and power quality of power grid.For power suppliers,accurate load prediction can be relied on to cope with the generation scheduling of power shortage or surplus events in some time periods,which is beneficial to improve the system load rate and reduce the system operation cost.Since the birth of power industry,load forecasting has always been a basic commercial problem.Accurate load forecasting is beneficial to the safe and economical operation of power system.Therefore,based on the above background,this thesis conducts an in-depth study on the problem of improving the accuracy of load forecasting,and designs and implements a forecasting model.The main contribution of this thesis lies in the following three points:(1)With the use and integration of renewable energy and external power stations,the study of power load forecasting is becoming more and more important.Here,this thesis divides power load forecasting into point load forecasting and probabilistic load forecasting according to the output form of power load forecasting.For point load forecasting,this thesis reviews,evaluates and analyzes the performance and research gaps of load forecasting schemes based on traditional forecasting models and neural network forecasting models.For probabilistic load forecasting,this thesis discusses and summarizes three different probabilistic forecast output forms of load forecasting.(2)Most deep learning tasks are single-task learning,and single-task learning often has the deficiency of ignoring the related information between problems.In order to solve this problem,based on MTL(Multi-Task Learning),and considering the similarity of load behavior among users in the same residential area,this thesis proposes an effective residential user load prediction scheme.The scheme uses K-means clustering technology and Pearson coefficient to select two similar users,and combines the data of the two users into the network model for training and testing.The experimental results on real datasets show that the proposed MTL load prediction scheme improves the prediction accuracy compared with the existing deep learning prediction schemes.(3)On the one hand,in view of the small amount and sudden peak load in the process of production or domestic electricity consumption,it will have a great negative impact on load forecasting.On the other hand,the most existing forecasting schemes use one-fit-all-size model to forecast the load,irrespective of spike or normal load.Based on the above observation,this thesis proposed a novel load forecasting scheme assisted by MTL-based spike occurrence prediction.In detail,at the first stage,a MTL-based spike occurrence prediction is explicitly proposed,which adopts MTL neural network to predict whether the loads to be forecasted would be spike or not.Rather than designing one-fit-all-size scheme,the second stage,based on the spike occurrence prediction in the first stage,intentionally selects the different data pre-processing technologies(i.e.,variance stabilization transformation)and deep neural network(DNN)prediction models that are suitable for respectively forecasting spike and normal load to forecast spike load and normal load.The experiment results on the Europe electricity load dataset from ENTSO-E Transparency Platform demonstrate that our proposed scheme increases the accuracy of spike occurrence prediction compared with the conventional models.