Short term electrical load forecasting is a topic of major interest for the planning of energy production and distribution. The use of artificial neural networks has been demonstrated as a valid alternative to classical statistical methods in term of accuracy of results. However a common architecture able to forecast the load in different geographical regions, showing different load shape and climate characteristics, is still missing. In this paper we discuss a heterogeneous neural network architecture composed of an unsupervised part, namely a neural gas, which is used to analyze the process in sub models finding local features in the data and suggesting regression variables, and a supervised one, a multilayer perceptron, which performs the approximation of the underlying function. The resulting outputs are then summed by a weighted fuzzy average, allowing a smooth transition between sub models. The effectiveness of the proposed architecture is demonstrated by two days ahead load forecasting of EOS power system sub areas, corresponding to five different geographical regions, and of its total electrical load.