Temporal Fusion Transformer for thermal load prediction in district heating and cooling networks

Abstract

Accurate forecasting of thermal loads is a critical factor for operating district heating and cooling networks economically, efficiently and with minimized emissions. If thermal loads are known with high accuracy in advance, use of renewable energies can be maximized, and fossil generation, in particular in peaking units, can be avoided. Machine learning has already proven to be an efficient tool for time series forecasting in this context. One recent advancement in machine learning is the "Temporal Fusion Transformer" (TFT), which shows especially good results in the area of time series forecasting. This paper examines the performance of TFT in the concrete context of thermal load forecasting for district heating and cooling networks. First, a brief summary of differences between TFT and other machine learning methods is given. Secondly, it is described how the method can be adopted to train a machine learning model for thermal load forecasting. The data to train and evaluate the neural network is based on 8 years of hourly operating data made available from the district heating network of the city of Ulm in Germany. The presented technique is used to produce 72 hours of heating load forecasts for three different district heating grids in the city of Ulm. The results are compared to forecasts of other machine learning methods that have been previously made as part of the publicly funded research project "deepDHC", in order to evaluate if TFT is an improvement to further reduce forecasting uncertainties.