A composite k-nearest neighbor model for day-ahead load forecasting with limited temperature forecasts

Abstract

Load forecasting is an important task in power system operations. Considering the strong correlation between electricity load demand and weather condition, the temperature has always been an input for short-term load forecasting. For day-ahead load forecasting, the whole next-day's temperature forecast (say, hourly or half-hourly forecast) is however sometimes difficult to obtain or suffering from uncertain forecasting errors. This paper proposes a k-nearest neighbor (kNN)-based model for predicting the next-day's load with only limited temperature forecasts, namely minimum and maximum temperature of a day, as the forecasting input. The proposed model consists of three individual kNN models which have different neighboring rules. The three are combined together by tuned weighting factors for a final forecasting output. The proposed model is tested on the Australian National Electricity Market (NEM) data, showing reasonably high accuracy. It can be used as an alternative tool for day-ahead load forecasting when only limited temperature information is available.