Improvement of electricity consumption forecasts using temperature inputs

Forecasting electricity consumption requires the projection of past consumption patterns into the future using a mathematical model. The simplest model for doing this is the univariate model, which has no inputs, and therefore requires no knowledge of future trends in variables affecting consumption. It is well known that input driven (causal) models can have significantly better performance than univariate models, particularly when the input sequences are known. The paper demonstrates the improvement obtained in the case of a temperature input. Two case studies are examined, these case studies are carefully selected, since they represent situations where the temperature variable is a dominant and non-dominant input. It is shown that increased accuracy in the forecasting results can be achieved even when a non-dominant input is used. The analysis includes the determination of the optimal configuration of the temperature input and this is justified for two case studies examined. It is also shown that when future inputs are unknown, using a separate model to estimate future values of the input (and subsequent use of these estimates as inputs to the forecasting model) results in improved electricity consumption forecasts. In addition, two methods for the determination of the electricity forecasting models are employed.