Evaluating different artificial neural network forecasting approaches for optimizing district heating network operation

Accurate heat demand forecasting is essential for energy-efficient management of district heating networks (DHN), which face complexities such as varying weather, user behavior, and energy availability. This paper evaluates the effectiveness of Artificial Neural Networks (ANN), including recurrent Long Short-Term Memory Networks, Convolutional Neural Networks, and the Temporal Fusion Transformer, against the statistical SARIMAX model. These models are assessed based on their ability to predict diverse heat demand profiles and provide interpretable forecasts with optimization strategies, particularly emphasizing comprehensible confidence intervals. Utilizing a year's data from Stiftung Liebenau DHN, which includes multiple energy sources like CHP, biomass, and natural gas, and varied heat sinks such as residential buildings and greenhouses, we find that despite the CNN model achieving the lowest MAPE of 27 % for summer and winter, and 17 % for winter only, prediction accuracy is significantly affected by data volatility and irregularity. However, the models successfully capture the overall trend but face challenges in accurately predicting demand peaks and fluctuations. An economic analysis indicates that applying these predictive methods significantly enhances energy efficiency and offers economic benefits due to low investment costs. •Comparison of different ANNs for heating load forecasting•Experimental results for one year of data from real heating network•Applicable to different types of consumers (residential, non-residential, greenhouses)•Economic savings from operation improvements by demand forecasting in reference DHN