Research map / Multi-agent systems and games / Cooperative multi-agent RL
WFCRL: A Multi-Agent Reinforcement Learning Benchmark for Wind Farm Control
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces WFCRL, an open multi-agent RL benchmark suite for wind farm control interfacing the static FLORIS and dynamic FAST.Farm simulators, and benchmarks IPPO, MAPPO, QMIX and IDQN/IDRQN on yaw-control power-with-load tasks.
Problem
Wind farm control is hard because model-based strategies need tractable wake models and scale poorly with turbine count, and there was no open RL environment suite for general wind farm control, especially with dynamic simulators.
Approach
An open, customizable cooperative MARL environment suite (Gymnasium/PettingZoo) where each turbine is an agent controlling yaw, pitch or torque, offered on both a static simulator (FLORIS) and a dynamic simulator (FAST.Farm) to enable sim-to-sim transfer studies, with a fatigue-load penalty in the reward.
Key results
- On Turb3Row1 (FLORIS, Sc. 2 training), IPPO reached 406.0 ± 10.5 episode reward, the best among the compared algorithms.
- On Ablaincourt (FLORIS, Sc. 1 training), MAPPO (351.7 ± 0.1) and IPPO (351.0 ± 0.3) outperformed the value-based IDQN, IDRQN and QMIX baselines (≤263.3).
- Naive transfer fine-tuning of IPPO policies on FAST.Farm degraded performance to about 0% gain over the greedy baseline, while FLORIS-trained policies gained about 15% power on FAST.Farm before fine-tuning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train IPPO instead of MAPPO on Turb3Row1 under Wind Scenario II | IPPO and MAPPO · Turb3Row1, FLORIS Sc. 2 | score: 372.4 $\pm$ 15.5 -> 406.0 $\pm$ 10.5 | Helped |
| Train MAPPO instead of IPPO on Ablaincourt under Wind Scenario I | IPPO and MAPPO · Ablaincourt, FLORIS Sc. 1 | score: 351.0 $\pm$ 0.3 -> 351.7 $\pm$ 0.1 | No clear effect |
| Train IPPO instead of MAPPO on Ablaincourt under Wind Scenario II | IPPO and MAPPO · Ablaincourt, FLORIS Sc. 2 | score: 314.6 $\pm$ 3.4 -> 354.6 $\pm$ 4.1 | Helped |
| Expose policies to varied wind conditions during training rather than Scenario I | IPPO · Turb3Row1, FAST.Farm Sc. 1 | score: 217.4 $\pm$ 1.8 -> 218.7 $\pm$ 0.5 | Helped |
| Expose policies to varied wind conditions during training rather than Scenario I | IPPO · Ablaincourt, FAST.Farm Sc. 1 | score: 327.4 $\pm$ 1.6 -> 317.3 $\pm$ 4.1 | Hurt |
| Fine-tune the FLORIS-trained IPPO policy online on FAST.Farm | IPPO · Turb3Row1, FAST.Farm | power increase over greedy baseline: 15% -> 0% | Hurt |
Surprising findings
- Naive fine-tuning of FLORIS-learned policies on FAST.Farm degraded performance back to the greedy baseline rather than improving it.
- Value-based baselines (QMIX, IDQN, IDRQN) underperformed PPO-based methods substantially on both layouts.
Limitations
- FAST.Farm simulation remains slow, making online learning on dynamic simulators costly.
- No investigation of transferring yaw optimization results from FAST.Farm to a real wind farm.
- Fatigue load on FLORIS is only a proxy based on turbulence and rotor-plane velocity variation.
Experimental setup
Datasets: SmartEole wind data (Ablaincourt site)
Benchmarks: WFCRL Turb3Row1 (FLORIS and FAST.Farm) WFCRL Ablaincourt (FLORIS and FAST.Farm)
Metrics: episode reward evaluation score under wind scenarios I and II power output load penalty
Baselines: IPPO MAPPO QMIX IDQN IDRQN
Builds on: PPO MAPPO IPPO QMIX CleanRL
Closest papers by meaning
- PWM: Policy Learning with Multi-Task World Models (ICLR 2025 · similarity 0.83)
- Efficient Multi-agent Reinforcement Learning by Planning (ICLR 2024 · similarity 0.83)
- NeuralPlane: An Efficiently Parallelizable Platform for Fixed-wing Aircraft Control with Reinforcement Learning (NeurIPS 2024 · similarity 0.83)
- JaxMARL: Multi-Agent RL Environments and Algorithms in JAX (NeurIPS 2024 · similarity 0.82)
- Reinforcement Learning with Imperfect Transition Predictions: A Bellman-Jensen Approach (NeurIPS 2025 · similarity 0.82)
- TorchRL: A data-driven decision-making library for PyTorch (ICLR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.