Weather sensitive short-term load forecasting using artificial neural networks and time series

In this paper, a novel weather sensitive (WS) load model using a back-propagation artificial neural network (ANN) is proposed to reflect the load's dependence on temperature, hours of a day, days of a week, and seasonal effects. The model inputs are temperature offsets from the winter and summer changeover temperatures, average weekly and daily temperatures to cater for seasonal effects, and radial basis functions (RBF) centered at peak sensitivity points of load versus temperature to indicate the hours of a day. The other load components, which are the base load (BL) and the residual load (RL) components, are modelled using time-series ANN models. A special load history set is put into focus and used for the base load forecast. It is derived using a third order difference scheme that reduces the base load sequence into a stationary white-noise-like process, thus allowing linear estimation to be used. This same load history has also proven to be very successful when applied to one step ANN load forecast models. It allows the forecast to be carried out based on load and temperature inputs only, with no need to indicate the hour or day-type of the forecast. This leads to a big reduction in the number of inputs required and allows for the use of linear activation functions, which results in a dramatic reduction in the learning time required to calculate the model's coefficients. An investigation of radial basis function (RBF) network models also shows the appropriateness of this load history set for such models, but reveals their weak responsiveness to temperature inputs. Results from the three models are then compared when they are used to forecast the hourly loads of a winter and spring week of a typical north-European utility.