eDriveMORL: Offline Reinforcement Learning Dataset and Benchmark for FCEVs

eDriveMORL is a benchmark suite for offline reinforcement learning on fuel cell electric vehicle (FCEV) systems. It includes:

• High-fidelity FCEV dynamic simulation
• Minari-compatible offline datasets
• Multiple D3RLpy-compatible algorithm configs
• Custom reward function and thermal modeling

📦 Project Structure

.
├── run.py                      # Run benchmark for all algorithms via CLI
├── train.py                    # Generate offline dataset from Minari
├── register_minari_dataset.py  # Register Minari-compatible FCEV dataset
├── datasets/                   # Stores generated D3RLpy HDF5 datasets
├── requirements.txt            # Python dependency list
├── fcev/                       # Core model implementation
└── README.md                   # You are here

⚙️ Environment Setup

0. Git Download the code and dataset

git lfs install
git clone [email protected]:datasets/TJIET/eDriveMORL

1. Create a Conda environment (Python 3.9 recommended):

conda create -n fcev-benchmark python=3.9
conda activate fcev-benchmark

2. Install required dependencies:

cd eDriveMORL
pip install -r requirements.txt

🗂️ Step-by-Step Usage

1️⃣ Register the Minari Dataset

Before any training or dataset generation, register the Minari dataset:

python register_minari_dataset.py

This ensures that your local offline dataset (e.g., collected via MPC) is discoverable by minari.load_dataset().

2️⃣ (Optional) Regenerate Offline Dataset

If you want to regenerate a D3RLpy-compatible dataset (HDF5 format), modify and run:

python train.py

This will create a .h5 dataset under the datasets/ folder, such as:

datasets/fcev-mpc-v1.h5

You can switch to different reward shaping or normalization settings inside train.py.

3️⃣ Run Offline RL Benchmarks

Run the benchmark suite using different algorithms (TD3+BC, CQL, AWAC, etc.):

python run.py \
  --algo CQL \
  --dataset-path datasets/fcev-mpc-v1.h5 \
  --drive-cycle CLTC-P-PartI.csv \
  --n-steps 10000 \
  --wandb-project fcev-offline-benchmark

🔧 Available algorithms:

• TD3PlusBC
• IQL
• CQL
• BCQ
• CalQL
• AWAC
• ReBRAC
• TACR
• PLAS
• PRDC
• BEAR

Use --wandb to enable logging to Weights & Biases.

📊 Dataset: eDriveMORL

All offline training is based on the eDriveMORL dataset, registered through Minari. It captures state-action-reward sequences collected via expert controllers (e.g., MPC) from a simulated FCEV model.

Dataset fields include:

• State: [SOC, T_fc, T_core, T_surf, speed, acc]
• Action: [fc_ratio, cooling_level, coolant_split_ratio]
• Reward: Custom function reflecting energy and thermal efficiency
• Termination: Episode end or infeasibility

📈 Logging & Evaluation

• 📝 TensorBoard logs saved under tensorboard_logs/{algo}
• 📁 File logs (e.g., model snapshots) under d3rlpy_logs/{algo}
• 📊 WandB metrics (optional): view your experiment dashboard online.