Research On Household Short-term Power Load Forecasting Based On Ensemble Learning

Problems such as resource shortage and low utilization efficiency have always been the deep-seated contradictions faced by the current global energy development,and energy transformation is imminent.According to statistics from the International Renewable Energy Agency,by the end of 2020,China’s installed photovoltaic capacity has reached 254 GW,and the installed wind power capacity has reached 282 GW,occupying a leading position in the world.With the increase in the installed capacity of new energy,the problem of energy consumption has become increasingly prominent.Power load forecasting plays an important role in the modern management of power systems.Accurate forecasting can alleviate the contradiction between power supply and demand,and is of great significance for maintaining the safe and stable operation of the power grid system.Electricity load is a typical time series data.With seasonal changes,user load will also show different periodic laws.Political environment,economic policies,human activities and other factors will also have a certain impact on user load.Therefore,it is difficult to use a single model.Accurately predict electricity load.This paper takes the load data of 114 households disclosed by the Smart* project of the University of Massachusetts as the research object.The main contents are as follows:（1）For the preprocessing of power load data,analyze the given power load data of114 households and 13-dimensional influencing factors such as load-related temperature,humidity,and electricity price,and analyze data correlation with PCC.The influencing factors show a typical nonlinear relationship,and the PCC processing is not comprehensive.Therefore,the feature＿importances method in the ensemble learning model is used to assist the calculation,and the optimal input feature set of the model is finally determined as 8-dimensional influence data and 1-dimensional load data.（2）In view of the problem of user load data processing,considering the poor effect of K-means on curve clustering,this paper proposes a new load curve clustering decomposition model based on the combination of K-shape clustering and STL decomposition.Firstly,the time series characteristics of users’ electricity consumption are analyzed,and the K-shape algorithm is used to divide the user clusters according to the load curve;secondly,the load data of different clusters are divided into seasonal items,trend items and random items by using the STL algorithm.Finally,combined with the real load data,the short-term load forecasting model is compared with various load clustering and decomposition models to verify the feasibility of the algorithm.（3）Aiming at the problem of power load forecasting,gc Forest is applied to the field of short-term load forecasting,and a two-layer gc Forest regression classification model is proposed.The first layer is the user load regression prediction problem.The model input is household power load data and impact data,and the prediction target is the user power load demand in the future.The second layer is the user load classification prediction problem,that is,the power load prediction is converted into a binary classification problem.The input features are consistent with the regression problem,but the prediction target is modified to output the power load demand information at the same time the next day according to the power load demand information at a certain time of the previous day.Load demand fluctuations.This part proves the advantages of the gc Forest model in dealing with the classification and forecasting of user power loads,and at the same time,the framework of the gc Forest model is modified to apply it to the regression forecasting problem,which expands the research scope of the model.To sum up,this paper deeply analyzes the user’s electricity load from the three dimensions of data preprocessing,load decomposition and load prediction,and proposes a new load clustering decomposition framework K-shape＿STL,and applies the new framework to household users.On the real-load dataset,the efficiency of the K-shape＿STL framework is demonstrated.At the same time,the deep ensemble learning model gc Forest is applied to short-term load forecasting,and experiments show that the model has a forecasting accuracy of 91.9%.The research results of this paper not only promote the development of short-term load forecasting,but also expand the application scope of gcForest,which has certain practicability.