Technologies Of Distribution Area Sequential Load Forecasting Based On Big Data

With the optimization and upgrading of SCADA system,the power user side can not only obtain the traditional power usage information but also the data of user power quality,96-point load curve,power behavior preference which reflect user's multidimensional characteristics.As a subdivision of the traditional load forecasting area,the load forecasting in the distribution area is an emerging means for the refinement of power management,operation scheduling and grid structure optimization.Due to the randomness and diversity of distribution area load,the traditional prediction method is poor when meet this situation,and with the aid of the new big data processing platform,a large amount of user characteristic information is collected from the user side Which can effectively improve the prediction precision and prediction adaptability of the load forecast.Based on the clustering algorithm of K-means,BIRCH and WARD in big data platform,this thesis classifies the 2977 distribution areas in Shanghai;analyzes the characteristics of each cluster and the validity of partitioning combined with the composition of the actual industry area.The clustering results show that the users have different sensitivities to the meteorological and economic factors.Based on the clustering results,it is possible to improve the accuracy of the load forecasting by establishing the forecasting models for the users with different electrical characteristics.An adaptive ridge regression model based on ridge regression and adaptive thought in machine learning is proposed,and the model is divided into three patterns according to the degree of self-adaptation.By training and using the three patterns in actual area forecasting work,it shows different prediction accuracy,sensitivity when meet mutative prediction environment,each pattern has it suitable forecasting environments.Based on clustering and regression model construction,this thesis proposes a load forecasting method combined clustering and adaptive ridge regression techniques,and designs three clustering feature selection methods and five clustering algorithms as predictive and adaptive optimization modules,Which can enhance the ability of dynamic optimization and error control.In this thesis,the actual prediction is made for a 487-user distribution area in Shanghai,and the results show that this prediction method can achieve high prediction by optimizing clustering features,clustering algorithms and model parameters in different forecasting environments,so it can achieve a higher prediction accuracy and environmental adaptability.