District Heating networks (DHNs) are promising technologies for heat distribution in residential and commercial buildings since they enable high efficiency and low emissions. Within the recently proposed paradigm of smart grids, DHNs have acquired intelligent tools able to enhance their efficiency. Among these tools, there are demand forecasting technologies that enable improved planning of heat production and power station maintenance. In this work we propose a comparative study for heat load forecasting methods on a real case study based on a dataset provided by an Italian utility company. We trained and tested three kinds of models, namely non-autoregressive, autoregressive and hybrid models, on the available dataset of heat load and meteorological variables. The optimal model, in terms of root mean squared error of prediction, was selected. It considers the day of the week, the hour of the day, some meteorological variables, past heat loads and social components, such as holidays. Results show that the selected model is able to achieve accurate 48-hours predictions of the heat load in several conditions (e.g., different days of the week, different times, holidays and workdays). Moreover, an analysis of the parameters of the selected models enabled to identify a few informative variables.