Classification Tree Based AI System for Short Term Prediction for Heat and Power Plants
Abstract
Companies supplying electrical energy rely mainly on long term agreements with electricity producers, but on the other hand the actual demand should be precisely predicted for 48 h ahead, to take into account the actual weather conditions. The same type of analysis is important also for heat and power plants, but this time the temperature of returning water is the most interested. Some time series models can be used for forecasting. However in daily practice – more popular are average profiles showing the 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 two tested classification methods, CART classification tree performed better.