Short-term Power Load Forecasting Based On TPE-GOSS-LightGBM Model

With the continuous development of society,the development of power industry is more and more closely related to the development of the country.At the same time,people’s living standards have also improved to a certain extent,so the demand for electricity is also growing.A high-quality power supply can provide a stable and efficient development environment for the whole country,which is an important guarantee for social development.In the real power energy supply system,short-term power load forecasting is of great significance to ensure the stable operation of the system,and targeted short-term power load forecasting of the system can also provide effective decision-making basis for power system planning and power grid investment.Traditional load forecasting methods can not meet the needs of today’s complex power network,and it is difficult to meet the needs of power grid regulation.With the continuous development of artificial intelligence technology,machine learning algorithm has attracted extensive attention in the field of power load forecasting because of its high efficiency and high performance.In order to improve the accuracy of short-term power load forecasting and solve the problems of weak pertinence and low forecasting efficiency of the existing load forecasting methods,this paper establishes a TPE-Goss-Light GBM model for short-term power load forecasting in a province in Northeast China by using the pure time series data set to construct the data set through window processing and external feature combination.Firstly,the research on power load forecasting based on machine learning strategy is analyzed.The classification and characteristics of power load forecasting are explored,and the steps of short-term power load forecasting are determined.This paper introduces MLP,KNN,SVR,Decision Tree and other machine learning models that can be used for time series prediction,as well as GBDT,Ada Boost,XGBoost,Light GBM and other integrated learning algorithms,further introduces the differences and advantages between Light GBM model and other integrated learning,as well as the direction of algorithm improvement,and applies Goss-Light GBM model to short-term power load forecasting.Secondly,data preprocessing and feature engineering are carried out.Aiming at the abnormal and missing values of load data in a province in Northeast China,this paper analyzes the periodicity of load data,defines the load forecasting data set with single window width by using the sliding window method,and introduces the external characteristics,as well as the evaluation index and experimental environment of power load forecasting.Prepare for the next model training and testing.Finally,the short-term power load forecasting model is studied.Through the construction of single,double and multiple window width load data sets and various combinations of external features,8 different data sets are formed KNN model,MLP model,SVR model,Decision Tree model,GBDT model,Ada Boost model,XGBoost model and Goss-Light GBM model.Through the comparison of prediction and evaluation indexes,the Goss-Light GBM model established by selecting the data set with single window width and external characteristics is the best.The model is optimized and improved by TPE(Tree-structured Parzen Estimator)algorithm.Compared with grid search optimization and random search optimization,it is proved that TPE algorithm has higher optimization efficiency,and further improve the accuracy of the model.Through the above research,the established TPE-Goss-Light GBM model for short-term power load forecasting of a province in Northeast China,Med AE,RMSE,Mae and MAPE are 166.537,278.503,210.713 and 1.659%respectively,which is 3.681,5.635,8.955 and 4.16%lower than the Goss-Light GBM model before optimization,and R~2 is 92.690%,an increase of 0.298 percentage points.While meeting the prediction accuracy,it reduces the redundancy of input data,prevents the occurrence of over fitting,and solves the problem of low efficiency of short-term load forecasting,which helps to reduce the operation cost of power grid and improve the overall regulation ability of power grid.