Research On Mid-long-term And Short-term Load Forecasting Based On Machine Learning

Load forecasting is an important foundation of power system planning,construction,and operation.It is instrumental in heightening the operation efficiency and dropping the operation cost of the power system.Power load shares a close uncertainty correlation with multiple factors such as economic policy,human activities,and meteorological conditions,by which precise forecasting is rendered quite a challenge and the traditional statistical methods have been difficult to meet the accuracy requirements of load forecasting.Given the fact,the study exerted the powerful data digging and fitting capabilities of various machine learning algorithms to investigate the mid-long-term and short-term load forecasting.Below are listed the prime tasks achieved in the study.(1)The study presented an overview of the basic theories of load forecasting and machine learning.It first introduced the primary principles,characteristics,classification and forecasting steps of load forecasting,and then determined the data pre-treatment methods and the indexes used to evaluate the forecasting models,following which was a description about the fundamental principles,development history,and categories of machine learning.(2)A mid-long-term load forcasting combined model was set up under the new normal of China’s economy based on GWO-SVM.In this section,the study first analyzed the challenges of trend changes of load development,insufficient available samples and increased uncertainties faced by mid-long-term load forecasting under the new normal of China’s economy.To probe into these challenges,improved gray Verhulst model,partial least square regression,and generalized regression neural network were selected as single forecasting models,while a support vector machine(SVM)was operated to reach combined forecasting,thereby it could utilize data information comprehensively and thoroughly.Moreover,gray wolf optimization(GWO)which harbors an efficient whole-domain search competence was adopted to select superior SVM parameters.This process not only helped avoid the blindness and subjectivity of parameter selection but also enhanced the generalization and popularization capabilities of SVM.At last,the load forecasting calculation example of Tianjin in 2016 and 2017 verified the precision and reliability of the combined model proposed in this thesis which can available references to the mid-long-term load forecasting in the new normal of China’s economic.(3)A short-term load forecasting model based on MIC-FA and improved LSTM neural network was raised to address the multiple input factors and high randomicity of short-term load.First of all,the periodicity of short-term load and the influence of meteorological factors were analyzed.Then,maximal information coefficient(MIC)method was applied to screen the input features of load forecasting,which avoided a shortage of artificially or traditional methods selected input features.Furthermore,these input features were processed with dimension reduction by factor analysis(FA).This simplified the dimensions of these factors in the premise that precision was ensured.Last,long short-term memory(LSTM)neural network was used to regressively analyze the screened and dimension reduced input variables.Meanwhile,the training of the LSTM neural network was optimized by a means of adaptive moment estimation so that a high-precision short-term load forecasting was achieved.The practical calculation example suggested that the short-term load forecasting model proposed in this thesis had a higher forecast precision and anti-interference ability.