Analysis of different testing parameters in Self-Organizing Maps for short-term load demand forecasting in Spain

Short-term forecasting is required by utility planners and electric system operators for tactical operational planning and day-to-day decision making. The forecasting is intended to obtain the system load demand over a period of hours or days, and it plays an important role in determining unit commitment, spinning reserve, economic power interchange, load management etc... Electrical load has a time-varying nature, and it is affected by various factors such as social, meteorological and financial conditions. Since the time horizon is relatively small in short-term load forecasting, the social and financial conditions have almost no influence on the forecasting process. In this research, the authors analyze the capability of a neural network, such as Self-Organizing Maps (SOM) for short-term load forecasting. The input data used concern the global load demand of Spain over several years and were obtained thanks to the Electrical Spanish System Operator: Red Eléctrica Española (REE). The study was focused on testing the first hours of the day to be forecasted in order to identify its common patterns with the historical database previously trained by the neural network. The input data has to be analysed beforehand to normalize them and filter anomalous days and holidays. Weekends were also excluded as their patterns are completely different to the rest of the week. After several simulations with different training parameters, three distinct tests were accomplished with the first 8, 10 and 12 hours of the day to be forecasted, and the errors between them were compared. Even in the case of 8 hours, the results show how the Self-Organizing Map is able to associate the evolution of the day with the most similar patterns in the database. The error in the case of the 10 hour test is lower and it reaches a minor value of around 1.6% when the test is carried out with 12 hours. Further tests will allow to select the best range of hours for creating the input data and will take into account the advantage of previously classifying the data depending on the season (summer, winter ...).