Process-guided deep learning for lake water temperature

This repository contains the code used to predict daily water temperature profiles (0-20 m) in Lake Mendota using deep learning models with physical constraints.

Overview

The project investigates four aspects of lake temperature modeling:

1. Base Models — Compare four deep learning architectures (LSTM, Transformer, CNN-LSTM, AttentionLSTM) trained on observational data with varying training set sizes (20%-100%).
2. Pretraining — Pretrain models on simulation data from the General Lake Model (GLM), then finetune on real observations.
3. Ensemble — Combine predictions from the four base models using a depth-wise weighted ensemble.
4. Process-Guided Loss — Add an energy conservation constraint to the loss function to improve physical consistency.

Repository Structure

process_guided_deep_learning/
├── Base Model/              # Train four DL models on observational data
├── Pretraining/             # Pretrain on simulation, then finetune on observations
├── Ensemble/                # Depth-wise weighted ensemble of base models
├── Loss Function/           # Energy-conservation-constrained ensemble
├── Simulation/              # GLM setup and simulation script
└── Validation/              # Monthly energy balance analysis

Environment and Data Setup

This repository was developed primarily for Google Colab + Google Drive.

The project uses daily weather and water temperature data from Lake Mendota, together with lake bathymetry, ice-cover records, and GLM simulation output. The expected input files are referenced in each module's environment_configuration.py.

Before running the scripts:

1. Mount Google Drive in Colab.
2. Update the paths in each environment_configuration.py file if your folder structure is different.
3. Check the additional hard-coded Google Drive paths in the main training and evaluation scripts, such as:
   • Base Model/basemodel_training.py
   • Base Model/figures_plot.py
   • Pretraining/basemodel_pretraining.py
   • Pretraining/basemodel_training.py
   • Pretraining/figures_plot.py
   • Ensemble/ensemble_data_processing.py
   • Ensemble/ensemble_evaluation.py
   • Ensemble/parameter_tuning_ensemble.py
   • Loss Function/ensemble_data_processing.py
   • Loss Function/energy_parameter_tuning.py
   • Loss Function/ensemble_energy_evaluation.py
   • Validation/monthly_energy_analysis.py

Requirements

• Python 3.8+
• PyTorch
• TensorFlow
• hyperopt
• scikit-learn
• numpy
• pandas
• matplotlib
• seaborn
• tqdm
• For simulation: R with packages GLM3r, glmtools, rLakeAnalyzer, tidyverse, ncdf4

Recommended Execution Order

1. Simulation

If you want to regenerate GLM simulation output:

• Run Simulation/general_lake_model.R
• This script should be launched from the Simulation folder

2. Base Model

• Run Base Model/basemodel_training.py
• Run Base Model/figures_plot.py

3. Pretraining

• Run Pretraining/basemodel_pretraining.py
• Run Pretraining/basemodel_training.py
• Run Pretraining/figures_plot.py

4. Ensemble

• Run Ensemble/parameter_tuning_ensemble.py
• Run Ensemble/ensemble_evaluation.py

5. Process-Guided Loss

• Run Loss Function/energy_parameter_tuning.py
• Run Loss Function/ensemble_energy_evaluation.py

6. Validation

• Run Validation/monthly_energy_analysis.py

Notes

• The repository is organized around comparative experiments on model architecture, pretraining, ensemble learning, and process-guided loss design.
• The current implementation uses random splitting of sliding-window samples, so the code is most appropriate for comparative evaluation under a shared protocol.