"Short Term Load Forcasting Using Artificial Neural Networks"

The prediction of the electric demand has become as one of the main investigation fields in the electric engineering. The electric industry needs to predict the load consumption with lead time in the range from the short term (hours or days ahead) to the long term (with several years ahead). The short-term prediction, in particular, has become increasingly important for various operations in power systems, such as economic scheduling of generating capacity, fuel purchase scheduling, security analysis, and planning activities. In addition, since many countries have recently privatized and deregulated their power systems, load forecasting play a crucial role in the final price of the energy. Small errors in the load forecasting have a significant economic impact. However, load forecasting is a difficult task because the load series is complex. First, the series exhibits several level of seasonality, and second, there are many exogenous variables that must be considered, specially weather-related variables. Thus, the relationships between hourly load and these factors are non-linear, so the forecasting problem requires a non-linear specification with a wide number of variables. Conventional load forecasting techniques, categorized into statistical methods, such as multiple regression and Box-Jenkins time series methods, present several limitations : complexity of modeling, lack of flexibility, low accuracy of results, mainly in special days, weekends and holidays, etc. In recent times, much research has been carried out on the application of artificial intelligence techniques to the load forecasting problem. Among these techniques, the models with the highest attention have been the Artificial Neural Networks (ANNs), mathematical tools originally inspired by the way the human brain processes information. ANNs are being applied to forecasting problems since they have a distributed architecture and their weights store interrelationships between variables without specifying them explicitly in advance. ANNs applications to the forecasting problem usually employ supervised learning in order to implement the non-linear mapping between historical data and future values of load. However, although the ANNs are being used by many utilities, there is certain skepticism among the researchers and the industries. And this, among another reasons, because the issues derived from the design of ANN-based forecasting system. In this conference we shall address the problem of designing a proper ANN attending to the main task to be performed, such as data pre-processing, the ANN design itself and its implementation, and the ANN validation