Numerous techniques of artificial intelligence have been used for building prediction models. One of such tasks is the prediction of heat consumption in a district heating system. Not only is it required for ensuring sufficient heat production, but also it is necessary to avoid substantial heat loss due to overestimated demand for heat. The work presents the use of multilayer perceptrons for building prediction models. However, instead of building prediction models based on artificial neural networks only, hybrid approach is considered and evaluated. Evolutionary approach used to combine neural networks and a number of simple methods into hybrid prediction models is presented. Such models are developed for groups of consumers sharing similar thermal properties identified by self-organising map. It has been shown that by combining neural networks with simple predictive strategies lower prediction error rates can be achieved than in case of using neural networks only.