The Research Of Electricity Load Forecasting Under The Condition Of Missing Data

The electricity load forecasting is a very critical technical approach of electric power and energy efficiency services. In recent years, intelligent data analysis technology has brought new market opportunities and growing space for power industry and electricity market. As electric services continue to deepened, they triggered a series of application scenarios such as electricity marketing services, electricity trading services, and also accelerated the lift of personalized customer services in smart grid. All of these can not be separated from support of load forecasting. At present, the available data may be incomplete and inconsistent. The problem of missing data has a negative effect on the prediction of short-term load, increasing the uncertainty of system analysis result. It is improving data quality and the accuracy and robustness of the prediction that become a problem needed to be solved urgently.In this paper, considering the adverse effects caused by the missing data on load forecasting, we combined the load forecasting with filling up missing data. Filling up missing data is used in the stage of data preprocessing, and then employ the hybrid of models to predict the future load. This dissertation studied the characteristics of load change and related factors, focusing on the methods of extraction of features and feature importance. In order to isolate the interferences in adjacent times, a multi-step prediction strategy was used to build models according to the factors in different times. Combining the advantages of parametric models and non-parametric models respectively, a more powerful model was used by integration of models. The method of optimizing parameters and controlling model complexity was also offered based on the characteristics of the training data.By a variety of experimental parameters selection and analysis, comparing with other algorithms, this thesis illustrated the excellent performance of gradient boosting machine in filling missing data and hybrid of models in load forecasting. Simulation test based on 10 regions of different load characteristics shows, the mean absolute percentage error of daily load is about 3.3%, and the week load forecast is at around 4.8%. When the load changes in large fluctuations and temperature is less relevant, the performance of the algorithm is stable, which has a certain application space.