AI System for Short Term Prediction of Hourly Electricity Demand

Companies supplying electrical energy rely mainly on long term agreements with electricity produsers, but on the other hand the actual demand should be precisely predicted for 48 h ahead, to take into account the actualweather conditions. Some time series models can be used for this purpose, and the best results can be achieved by combining forecasts from regression models, exponential smoothing, ARIMA models and neural networks. In practice - more popular are average demand profiles showing the average demand distribution over 24 h. We propose to build an AI system to choose the future profile. First - from the historical data - daily profiles are obtained, by cutting the time series into 24-h periods. Then, these empirical profiles are clustered with hierarchic and non-hierarchic clustering procedures to form homogeneous groups (types of profiles). Finally the classification methods are applied using weather data and observed demand from previous days (up to one week backwards). The measure for the forecasting evaluation has been proposed. Out of the tested classification methods, neural networks performed the best, followed by some voting procedures.