Files
Abstract
Weather factors are playing key impacts on electricity load consumption. The proper selection of weather stations will contribute to the final electric load forecasting accuracy. How to efficiently select the stations among a good number of candidates within the territory of interest remains a pressing issue. This thesis proposes several comprehensive weather station selection (WSS) frameworks along with different statistical tests to determine the stations to be used for electric load forecasting. We demonstrate comprehensive implementation and effectiveness of these methods based on the Global Energy Forecasting Competition 2012 (GEFCom2012) data and Global Energy Forecasting Competition 2014 (GEFCom2014) data are evaluated by comparing to the selection results obtained from the WSS framework introduced in (Hong et al., 2015). We introduce theoretical optimum (TO) selection to unveil what the best WSS looks like given we have access to the future load and from which, we gain further insights on why some WSS frameworks outperform the others. Additionally, we extend our discussion on several practical data fitting issues on the WSS subject and suggest several actionable rules of thumb that load forecasting practitioners can follow. Our experimental results show that the forecasting accuracy can be significantly improved by several proposed selection frameworks. Meanwhile, several heuristic methods have been applied to cut down the computational cost.