Data-driven modelling of an Aquifer Thermal Energy Storage (ATES) using machine learning

In the face of climate change and the imperative to reduce greenhouse gas emissions, harnessing geothermal energy for mitigating environmental impact and curbing global warming stands as a promising solution. The increasing reliance on intermittent renewable energy resources necessitates effective energy storage solutions for both electricity and heat. Underground reservoirs provide a substantial avenue for heat storage, predominantly achieved through Aquifer Thermal Energy Storage (ATES) systems. Interestingly, the widespread adoption of ATES remains limited, with the Netherlands hosting nearly 90% of global ATES installations. Furthermore, the majority of existing ATES systems predominantly store low-temperature heat (below 30°C). Enhancing their efficiency and energy density hinges on the storage of higher-temperature water during warm seasons for subsequent heat recovery in colder periods. This study investigates the potential of the Long Short-Term Memory (LSTM) deep learning approach to forecast the well temperatures at the ATES of a horticultural facility of Koppert-Cress, in the Netherlands. It delves into optimal data resolution, input and target variables, and loss functions. The research presents two LSTM-based model architectures, harnessing historically measured data to generate an extended forecast horizon concurrently. Robustness and stability are assessed via cross-validation, with model performances meticulously compared against original data. Furthermore, the LSTM-based model's performance is benchmarked against available data from the ATES system. The findings indicate that when employing ATES historical data as input, the LSTM model achieves consistent and robust performance across the forecast horizon, rendering it suitable for operational deployment. Notably, the system predominantly operates in heating mode (37% of the time), reflecting the region's climate demands for heat. Moreover, a strong correlation is observed between environmental conditions and warm-well temperatures. The initial LSTM model serves as a foundational part of this study for model development and data familiarization, with the primary objective of predicting two categories of wells’ temperatures. Subsequently, the second LSTM architecture is tailored to extend the forecasting horizon and also predict in a shorter time. The research outcomes reveal the reliability of both models, characterized by low evaluation metrics for the regression errors (MSE, RMSE, MAE) and high accuracy (R2) values in predicting wells’ temperatures. The singular LSTM model achieves the highest R2 of 0.97, while the parallel model achieves an R2 of 0.87. This study advances the operational understanding of temperature outcomes from ATES wells through the application of LSTM deep learning. Future research avenues may explore the integration of this ML model into control systems and assess the quantification of heat requirements for upcoming time horizons within buildings, facilitating proactive energy management strategies.