Estimating temperature profiles for short-term load forecasting: neural networks compared to linear models

Short-term load profile forecasts (forecasts of the next day's 24 hourly loads, for example) have become vital tools for the efficient operation of power systems. Many new forecasting systems have been proposed in recent years, most based on models that relate the load profile to the temperature profile. Weather services, however, do not usually supply temperature profile forecasts, but only predictions of maximum and minimum values. Some methods to estimate temperature profiles by linearly interpolating between the predicted extreme values have been devised by the utilities. A model that replaces such methods by a multi-output neural network is proposed. In out-of-sample simulations over real data, this model outperformed the more traditional methods we used for comparison. Forecasting systems like this, based on neural networks, may help to deal with the lack of weather-service profile forecasts of temperature (or other weather-related variables), and may ultimately contribute to the reduction of the load forecasting error.