Machine-learning-assisted long-term G functions for bidirectional aquifer thermal energy storage system operation

Optimization of aquifer thermal energy storage (ATES) performance in a building system is an important topic for maximizing the seasonal offset between energy demand and supply and minimizing the building's primary energy consumption. To evaluate ATES performance with bidirectional operation, this study develops an analytical solution-based model to simulate the spatiotemporal thermal response in an aquifer. The model consists of three temperature response functions, similar to the G functions in borehole thermal energy storage (BTES), to estimate the transient temperature profile in the aquifer during seasonally varying injection and extraction of hot/cold water. Applying machine learning (ML) based data classification and regression techniques to the results of a series of finite element (FE) benchmark simulations of typical ATES configurations, model input parameters are linked to the subsurface thermal, hydrogeological, and ATES operational properties. Compared to the benchmark simulation results, the errors of the proposed model in estimating the annual energy storage and locating the thermally affected area are about 3 % and 1 %, respectively. The model was applied to a previous short-term case study, and the error in the transient production temperature estimation is about 1 %. The long-term heat recovery ratio estimated from the model also compares well to those calculated from the previous study and the validated numerical model. Because of its fast computation, the proposed model can be coupled with the individual building system simulation and used for preliminary ATES design, and this will allow for greater exploration of ATES operational space and, therefore, better choices of ATES operating conditions. The proposed model can also be coupled with the district heating and cooling network simulation for computationally efficient city-scale long-term ATES potential assessment.