The aim of this study is to investigate the potential of soft computing methods for selecting the most relevant variables for predictive models of consumers’ heat load in district heating systems (DHS). Data gathered from one of the heat substations were used for the simulation process. The ANFIS (adaptive neuro-fuzzy inference system) method was applied to the data obtained from these measurements. The ANFIS process for variable selection was implemented in order to detect the predominant variables affecting the short-term multistep prediction of consumers’ heat load in district heating systems. It was also used to select the minimal input subset of variables from the initial set of input variables – current and lagged variables (up to 10 steps) of heat load, outdoor temperature, and primary return temperature. The obtained results could be used for simplification of predictive methods so as to avoid multiple input variables. While the obtained results are promising, further work is required in order to get results that could be directly applied in practice.