Physics-constrained deep learning for reservoir thermal structure prediction: Enhanced interpretability and extrapolation capability | JoVE Visualize

Area of Science:

Environmental Science
Water Resource Management
Machine Learning

Background:

Accurate reservoir thermal structure prediction is vital for ecological protection and optimized reservoir operations.
Existing data-driven models struggle with limited data, poor physical interpretability, and unreliable extrapolation.
Challenges include predicting water temperature dynamics and understanding stratification.

Purpose of the Study:

To propose a physically constrained deep learning framework (P-DL) to overcome limitations of current data-driven models.
To enhance prediction accuracy, physical interpretability, and extrapolation stability for reservoir thermal structure.
To provide a reliable tool for intelligent thermal management in reservoirs.

Main Methods:

Developed a physically constrained deep learning framework (P-DL).
Augmented training data using mechanism-driven process models and identified key influencing factors.
Transformed vertical temperature profiles into interpretable parameters (A, B, D) to represent stratification intensity and improve extrapolation via weak physical constraints.
Compared P-DL with Random Forest (RF), Support Vector Machine (SVM), and Long Short-Term Memory (LSTM) using Xiangjiaba (XJB) reservoir data.

Main Results:

P-DL demonstrated superior accuracy in predicting short-term local fluctuations compared to RF, SVM, and LSTM.
Interpretable parameters (A, B, D) effectively captured stratification intensity, peak timing, and temporal evolution.
Under the SSP5-8.5 scenario, P-DL achieved high accuracy (RMSE: 0.83-1.1 °C; R²: 0.88-0.9) for surface temperatures during strong stratification.
P-DL showed superior consistency (KLD: 2.85-5.71; KSS: 0.2-0.4) at both local and overall levels.

Conclusions:

The proposed P-DL framework significantly improves prediction accuracy, physical interpretability, and extrapolation stability for reservoir thermal structure.
The framework offers a valuable reference for intelligent thermal management and ecological protection in reservoirs.
The hybrid model and weak physical constraints approach can advance data-driven predictions for other environmental factors.