Sustainable Geothermal Energy for the Future: AI in ATES

ATES-Heat Pump systems enable simultaneous supply of heating and cooling and provide free heating and cooling making them economical and thermally efficient solutions. However, suboptimal operation significantly hinders performance. Current modeling practices require extraordinary effort, domain knowledge and long computation time making it infeasible for control and operation of real systems. In this project, numerical and physics-informed machine learning(PIML) models are developed for ATES leveraging PIML ability to incorporate the physical laws governing a system in the learning process, utilizing their effectiveness in solving realistic problems and fast computation time. The models are tested on a comprehensively monitored (since 2016) ATES site and facilitate its integration into the control and operation system. The project aims to facilitate geothermal storage technology development and increase Sweden's research and industry competitiveness regarding design, modeling, operation, and control.

Background

The need to find alternative energy sources other than fossil fuels has become prioritized in order to secure a reliable and sustainable energy supply, to reduce the carbon footprints to mitigate climate change. According to the Swedish Energy Agency, the residential and services (including commercial buildings) sector is responsible for 147 TWh of the total energy consumption in Sweden which represents nearly 35% of the total energy use in the country for 2018. More than half of the energy consumed in this sector was used for space heating and domestic hot water. A major challenge for renewables is the mismatch between demand and supply which led to investment in geothermal energy storage technologies such as borehole or aquifer thermal energy storage (BTES and ATES). Furthermore, the need for sophisticated efficient controls that capture the underlying physics of the building energy system components is becoming increasingly critical in the energy transition process to have optimal operation. The geothermal energy storage component of the system contributes significantly to the increased overall efficiency given its utility for short and long-term thermal energy storage, economic feasibility, short payback time (1-3 years for ATES), ability to be coupled with Heat Pumps, providing free cooling and/or heating. This highlights the need for detailed geothermal storage models to be incorporated into the energy system controls. In the case of ATES, the heterogeneity of the subsurface often requires extensive modeling work, conducting long-term monitoring and performance evaluation of ATES systems to develop a better understanding of the ATES behavior and how to optimize ATES-Heat Pump energy systems operation. Modeling such a complex system, using physical principles, requires extraordinary effort, time, and domain knowledge, and can be computationally expensive. This makes this approach of modeling not feasible to be used for the control and operation of an actual geothermal energy system. Modeling methods based on machine learning (ML) have been developed to overcome these challenges for buildings and other complex systems. However, while having found great success in certain applications, purely data-driven modeling methods often require large amounts of data, have poor generalization performance and physical consistency, and have difficulty handling low-quality data, due to their lack of physical insights. To address these limitations, physics-informed machine learning (PIML) methods have recently emerged to incorporate known physical laws governing a system into the learning process. This is a significant development to have interpretable, robust, accurate, and physically consistent ML models of physical systems, by merging ML with physics-based modeling. PIML models are useful for prediction, simulation, control, and optimization, with often higher confidence in their performance and safety compared to non-physics-informed ML models. Therefore, the physics-informed data-driven modeling approach for geothermal systems has a high potential to reduce the computational cost, improve the robustness, accuracy, and feasibility of the model to be integrated into the energy system control and operation to have more efficient use of the geothermal resource (ATES) and achieve sustainable operation and optimal interaction between the energy system components

Aim and objectives

Quantified key performance indicators for coupled ATES-Heat Pump system performance and optimized operation.
Increased knowledge about physics-informed machine learning in geothermal energy storage with a focus on ATES systems.
Implementation of physics-informed machine learning for ATES modeling.
Testing modeling methods on an operation project to provide more concrete and applicable knowledge.
Collaboration between industry and academia for knowledge exchange and strengthening Swedish research and industry competence and competitiveness nationally and internationally.