Fast prediction of aquifer thermal energy storage: a multicyclic metamodelling procedure

The metamodel-based approach (also referred to as the surrogate approach) is commonly applied to overcome the computational burden of numerical models that are used to simulate the evolution of reservoir fluids and pressures in response to any production scheme. In this study, we propose an adaptation of this approach for aquifer thermal energy storage (ATES) systems. ATES systems are characterized by cyclic loading/unloading production schemes, which result in a strong similarity in the dynamics of the intercyclic evolution of variables such as the temperature at the producer well. Instead of training several metamodels, i.e., one per cycle (independent metamodelling approach), we take advantage of the intercyclic similarity to train a single metamodel within the setting of multifidelity cokriging (multicyclic metamodelling approach). To explore the predictive performance of this approach, we applied a random subsampling validation approach multiple times to 300 simulation results of a realistic ATES system in the Paris basin by considering three characteristics, i.e., the minimum and maximum temperature, and the rate of temperature decrease at each cycle. Numerical experiments with varying training dataset sizes (from 33 to 66% of the total number of results) and using 100 test samples show that (1) the predictive error of the multicyclic metamodelling reaches lower levels (by 20-50%) than that of the independent approach; (2) this higher predictive performance is achieved while saving computational time cost because the training phase only needs a few tens of complete simulations (run over all cycles) together with a few hundreds of partial simulations (stopped at the first cycle); the latter simulations are less expensive to evaluate because of shorter simulated time.

Automated summary

This study proposes an adaptation of the metamodel-based approach for aquifer thermal energy storage (ATES) systems. By training a single metamodel within the setting of multifidelity cokriging, the predictive performance is improved while saving computational time cost.