Introduction
Maximizing the economic return of geothermal reservoirs, typically measured by Net Present Value (NPV), is a complex task requiring precise management of well controls such as injection rates and bottom-hole pressure. Two primary challenges hinder real-time optimization: significant geologic uncertainty regarding subsurface permeability and the high computational cost of traditional numerical simulators. This paper proposes a deep learning-based closed-loop framework designed to overcome these hurdles by enabling simultaneous reservoir characterization and well-control optimization.
Methodology
The core of the proposed framework is a hybrid convolution-recurrent neural network surrogate model that replaces slow numerical simulations.
The architecture consists of several key components:
- Convolutional Neural Network (CNN): Functions as an encoder to extract spatial features from 2D permeability maps, compressing complex geologic data into low-dimensional "latent variables."
- Long Short-Term Memory (LSTM): A recurrent structure that maps sequences of well controls to sequences of production responses, such as water rate, temperature, and pressure.
- Iterative Ensemble Smoother (IES): An ensemble-based data assimilation method used to update the reservoir's permeability map as new production data becomes available.
- Differential Evolution (DE): A heuristic optimization algorithm employed to identify the optimal well-control settings based on the surrogate model's predictions.
By coupling these elements, the system performs iterative data assimilation and production optimization, effectively "learning" the reservoir characteristics while maximizing NPV.
Results and Discussion
The framework was validated using synthetic datasets generated via the MATLAB Reservoir Simulation Toolbox (MRST), focusing on 2D geothermal models with multiple injection and production wells over a 10-year span.
Computational Efficiency and Accuracy
The deep learning-based optimization achieved a massive speedup, completing in approximately 30 seconds—a stark contrast to the hours or days required by traditional numerical simulators. Despite this speed, the surrogate model maintained high accuracy, with predictions for water production rate, temperature, and bottom-hole pressure closely matching numerical references.
Economic and Characterization Gains
In scenarios with both known and unknown permeability, the optimized cases yielded significantly higher NPV compared to reference cases using constant controls. Furthermore, the framework demonstrated robust performance against geologic uncertainty by optimizing the average objective function across an ensemble of reservoir realizations. As production progressed, the Root Mean Square Error (RMSE) of permeability estimations dropped steadily, indicating that the data assimilation process effectively reduced subsurface uncertainty.
Conclusion
This research demonstrates the potential of hybrid deep learning architectures to revolutionize geothermal reservoir management. By integrating CNNs for spatial feature extraction and LSTMs for temporal dynamics, the proposed closed-loop framework enables real-time, uncertainty-aware optimization. Future work will focus on extending this approach to 3D reservoir models and incorporating more complex physical processes to bridge the gap between synthetic studies and real-world applications.
Summary updated