Short-term Power Load Forecasting Based on Machine Learning

Sammanfattning: Short-term power load forecasting is an important part of power system management.It is the premise of network structure planing, electricity tradingand load scheduling. The accuracy of power load forecasting is directly relatedto power system security, stability and economic operation. In this thesis,machine learning based short-term load forecasting methods are studied andanalyzed, which include extreme gradient boosting (XGBoost), light gradientboosting machine (LightGBM), long short-termmemory network (LSTM)and gated recurrent unit network (GRU). Apart from historic load data, environmentalfactors are also taken into consideration and used as input vectorsto the model. Pearson similarity method is adopted to analyse the correlationbetween each environmental factor and the corresponding power load.The performance of the machine learning based load forecasting methods arecompared to a commonly used traditional load forecasting method - autoregressiveintegrated moving average model (ARIMA).The results show thatthe proposed machine learning based methods have a higher prediction accuracythan that of traditional method.