Research On The Application Of Machine Learning To Short-term Power Load Prediction

Short-term power load prediction is an important basis for ensuring the safety,stability,and economic operation of power system.It is mainly used to predict the load from one day to several days in advance.Accurate load prediction results help to guide and determine the operation of generator set,coordinate the supply and demand balance between various grid companies and the planned maintenance of grid equipment,reduce resource waste and electricity costs,improve the economic benefits of power grid operations.Based on the relationship between load data and various influencing factors,as well as the underlying trends of the data itself,this paper uses improved temperature and humidity variables combined with ES-LASSO model method,DMD-NARX model method to predict short-term load,and uses actual load data to verify the two methods.The main content of this article is as follows:(1)Study the construction of input features in short-term load prediction models.This part first introduces the common preprocessing and normalization methods of load data;then,by analyzing the regular characteristics of short-term loads and external factors affecting load changes,the input feature range of the prediction model is determined;finally,feature extraction,quantization,and selection methods are used to select a subset of input features in the input feature range.(2)A short-term load prediction method based on improved temperature and humidity variables and ES-LASSO model is proposed.First,the improved temperature and humidity multi-form variables are used as model input variables;then,stratify the load according to the weekly changes,establish a prediction model based on LASSO for each layer of load,and use the ES solving algorithm to eliminate redundant input variables;finally,by calculating the coefficients corresponding to the remaining variables,the load distribution in each period is estimated.The example analysis results verify the prediction accuracy and robustness of the proposed method.(3)A short-term load prediction method based on DMD-NARX model is proposed.First,based on the results of ACF analysis of historical load data,construct the input feature set;then,use the Hankel matrix to convert the normalized univariate input feature sequence into a multidimensional data matrix,and perform feature decomposition and dynamic mode estimation on the multidimensional data matrix through the DMD method;finally,the estimated dynamic mode is used as the external input feature of the NARX neural network to estimate the load data of forecast day.The example analysis results show that the model caneffectively improve the load prediction accuracy.(4)Study the comparison between the two methods and each method.First,the improved temperature and humidity variables are combined with the ES-LASSO model method under different input variables and solution algorithm conditions;then,the DMD-NARX model method is compared with the traditional neural network model and machine learning model.finally,the two model methods mentioned in the paper are compared and analyzed.The results of comprehensive comparative analysis confirm the effectiveness and stability of the two short-term load prediction methods.