SEYED-MASOUD MOGHADDAS-TAFRESHI - Landis & Gyr Austria Vienna-Austria

The paper outlines a framework for mid-term prediction based on a hybrid fuzzy-neural approach: The first step to get estimates of typical daily load curves for one year in advance as needed in medium-term operation planning is the classification of characteristic load profiles for different daytypes. For this clustering a self-organized Kohonen-network with unsupervised learning is used. The result of the analysis which is performed separately onnormalized-load curves from summer and winter season, are load profile classes for the various types of days Mondays, working days, Saturdays, holidays). In a second step a weather-load-correlation model is identified on behalf of a multilayer perceptron with supervised backpropagation learning mode to enable different scenarios for various (fuzzy) assumptions about weather conditions. The input-layer neurons corresponding to explaining weathervariables are fed with temperature values. To account for the nonprecise character0f input data the temperature values are fuzzified by a fuzzy front-end processor.In the final section of the paper results and experiences obtained by tuning the etwork with A real test data from two different electric power utilities are presented to demonstrate the effectiveness of the proposed fuzzyneural forecasting methodology.