Appraisal of soft computing methods for short term consumers' heat load prediction in district heating systems

District heating systems can play a significant role in achieving stringent targets for CO2 emissions with concurrent increase in fuel efficiency. However, there are numerous possibilities for future improvement of their operation. One of the potential domains is control, where short-term prediction of heat load can play a significant role. With reliable prediction of consumers' heat consumption, production could be altered to match the real consumers' needs. This will have an effect on lowering the distribution cost, heat losses, and especially primary and secondary return temperatures, which will consequently result in increased overall efficiency of district heating systems. This paper compares the accuracy of different predictive models of individual consumers in district heating systems. For that purpose, we designed and tested numerous models based on the SVR (support vector regression) with a polynomial (SVR–POLY) and a radial basis function (SVR–RBF) as the kernel functions, with different set of input variables and for four prediction horizons. Model building and testing was performed using experimentally obtained data from one heating substation. The results were compared using the RMSE (root-mean-square error) and the coefficient of determination (R2). The prediction results of SVR–POLY models outperformed the results of SVR–RBF models for all prediction horizons and all sampling intervals. Moreover, the SVR–POLY demonstrated high generalization ability, so we propose that it should be used as a reliable tool for the prediction of consumers' heat load in DHS (district heating systems).