Thermal load prediction of communal district heating systems by applying data-driven machine learning methods

Abstract

Load forecasting is an essential part of the operational management of combined heat and electrical power units, since a reliable hour- and day-ahead estimation of their thermal and electrical load can significantly improve their technical and economic performance, as well as their reliability. Among different types of prediction techniques, data-driven machine learning methods appear to be more suitable for load estimation in operational systems, compared to the classical forward approach. Research so far has been concentrated mainly on the magnitude of buildings with single load types. It has only been extended to a limited degree on the level of a district heating network where several end users with different characteristics merge into one bigger scale heat consumer (city or group of communities). In this study, artificial neural networks are utilized, to develop a load prediction model for district heating networks. A segmented analytical multi-phase approach is employed, to gradually optimize the predictor by varying the characteristics of the input variables and the structure of the neural network. The comparison against the load prediction time series generated by a local communal energy supplier using a commercial software reveals that, although the latter is enhanced by manual human corrections, the optimized fully automatic predictors developed in the present study generate a more reliable load forecast.