Different Models to Forecast Electricity Loads

Forecasting electricity load is an important economic problem. Sometimes suited prediction methods are evaluated by benchmark scenarios, as in case of the world-wide competition within the EUNITE network. Here maximal daily electricity load has to be forecasted over a whole month given data of two years before. Because no formal model exists of how load depends on e.g. temperature or other environmental data, non-formal methods must be used. In this paper different approaches are presented. One is based on simple statistics, whereas other approaches rely on a general model in order to be applicable even for a long-term prediction. It is determined out of the given training data by simple methods. Then this approach is extended mainly by incorporating specific general background knowledge on holidays. Therefore a hybrid crisp-fuzzy system is used which is specified by rules as well as by learning.