Load-prediction with neural and statistical components in power systems with instationary load patterns

A hybrid pattern algorithm is presented combining statistical and neural methods to forecast hourly load of an electrical power supplying system. Compared with ordinary neural techniques which require a large stationary data-set for the parametrization of the huge number of net-weights, the algorithm yields to a sufficient prediction even with a small reference data-set and is especially suited for power utilities with instationary load patterns. In this sense the choice of appropriate model structures and parsimonious parametrization are considered in particular. The presented modular design yields to a high transparency of the entire prediction algorithm. Furthermore a clear assessable performance measure of the prediction accuracy of the four individual steps of the forecasting algorithm is presented.