Load Forecasting

Abstract

Load forecasting is critical to provide decision-making support for power generation. An accurate load forecasting can ensure sufficient power being generated to fulfil actual need of the community and reduce waste of over generation. In this project, short-term load forecasting models were built based on load and temperature history to forecast the up-coming 24-hour load usage for 20 zones in USA. Load history across 20 zones and temperature history of 11 stations from 2004 to 2008 in USA were aggregated to build 20 non-linear regression models using regression tree in Matlab. The built model could predict the next 24-hour load usage using the temperature of the predicted day and the load history of previous seven days. The mean absolute percentage errors between the forecasted load and actual load usage were calculated to evaluate the performance of the prediction model. This project concluded that with only limited information, sufficient good short-term load forecasting models can be built to forecast the up-coming 24-hour load usage which most of them have around 10% of mean error or less with known temperature while an outlier zone with 16% error. The built models can be used as preliminary information for short-term load generation plan.