A neural network model to forecast urban electricity consumption from weather data

The short-term load forecasting (STLF), with lead times ranging from a few hours to several days ahead, helps grid operators to make a cost effective scheduling of resources, purchase of energy, maintenance and security analysis studies. The use of reliable load forecasting models is necessary for a rational use of electricity, taking into account that it is not storable. Climatic conditions certainly have a remarkable influence on electric energy demand, and their incidence became quite overwhelming also for supply of energy, when it is obtained by renewable energy sources such as wind power and hydroelectric power. This paper investigates the correlation of weather variability with the electricity demand of a suburban area, focusing attention on load forecasting with a prediction time of 24 hours. A preventive classification of the historical load data is carried out by means of a hierarchical clustering algorithm. The actual forecast is obtained using a two layered feedforward neural network, trained with the back-propagation with momentum and variable learning rate algorithm. The neural network is trained using weather data along with historical load data related to a part of the electric grid of the town of Palermo (Italy).