| Abstract |
Seasonal thermal energy storage in groundwater is a popular alternative to sustainably heat and cool buildings. Complex subsurface settings remain a largely untapped resource due to uncertain system efficiency. In fractured rock reservoirs, predicting recovery of stored thermal energy is challenging due to the unknown and heterogeneous presence of preferential flow paths. To promote large-scale use, informing decision-makers about the probability of low system efficiency is crucial. We propose a computationally efficient approach to do this. We combine the IC-FERST code, which efficiently models flow and heat transport in fractured reservoirs, with the Bayesian Evidential Learning (BEL) framework, which enables uncertainty quantification with a limited number of model realizations. We mimic the approach of a consulting company in case of a lack of data. Three conceptual model scenarios of the fractured basement rock in Brussels (Belgium) are evaluated, and a broad prior uncertainty range is included on sensitive model parameters. For each scenario, we randomly sample 250 model realizations and simulate a short-term tracer test and the long-term efficiency. Statistical modeling is used to find a direct relation between data and prediction. It combines dimension reduction techniques with statistical/machine learning. When constrained by field data, this relation can model the posterior distribution of the long-term prediction (i.e. a refined prediction constrained to observed data including uncertainty quantification). This is crucial for decision-making because it predicts efficiency while still quantifying the investment risk related to subsurface uncertainties. The efficiency prediction is validated against long-term monitoring data of the operational ATES system. |