TL;DR
This paper introduces a novel Gaussian Process Model Predictive Control (GP-MPC) approach to stabilize the output voltage of Polymer Electrolyte Fuel Cells (PEFCs). By using Gaussian Processes to model PEFC dynamics, the GP-MPC can effectively regulate hydrogen and airflow rates while respecting safety constraints, offering a data-driven alternative to traditional physics-based MPC.
What Problem Does This Paper Address?
Polymer Electrolyte Fuel Cells (PEFCs) are a promising renewable energy technology, but their widespread adoption is hindered by challenges in ensuring reliable and stable operation. Specifically, maintaining a consistent output voltage under varying workloads is crucial for practical applications. Traditional Model Predictive Control (MPC) methods often rely on complex, physics-based models that require extensive system knowledge and can be computationally intensive. This paper addresses the need for a more data-driven, less information-intensive MPC approach that can effectively control PEFC voltage while adhering to critical safety constraints, such as hydrogen pressure limits.
Methods
The core of this work is the development of a Gaussian Process Model Predictive Control (GP-MPC) framework.
- Gaussian Process (GP) Modeling: Two GPs are employed to model the PEFC's dynamic behavior: one for the output voltage (VFC) and another for the hydrogen pressure (PH₂). These GPs learn the relationship between system inputs (hydrogen flow rate QH₂, air flow rate Qair, and current I) and the predicted states (VFC and PH₂). This approach leverages data rather than relying solely on first principles.
- Model Predictive Control (MPC): A standard MPC framework is used, but instead of a physics-based model, it utilizes the predictions from the trained Gaussian Processes.
- Linearization for MPC: The non-linear GP models are linearized to obtain a discrete-time state-space model (Ad, Bd, Cd) suitable for the Quadratic Programming (QP) solver used in MPC.
- Constraint Handling: The GP-MPC incorporates constraints on input change rates and hydrogen pressure. A novel aspect is the inclusion of a special inequality that uses the GP's prediction variance and the potential for hydrogen pressure change to compensate for model imperfections and ensure robust constraint satisfaction.
Datasets
The paper utilizes a dataset generated from a detailed physical fuel cell system model built in MATLAB Simulink, which serves as the "true system dynamics" for simulation.
- Training Data: 1000 data points were collected using Latin Hypercube Sampling (LHS) for the inputs QH₂, Qair, and current I, with a time step of 0.5 seconds.
- Testing Data: 300 test points were generated using a different set of LHS inputs to validate the trained GP models.
Key Insights and Ideas
- Data-Driven Control for PEFCs: Gaussian Processes offer a viable data-driven alternative to complex physics-based models for PEFC control, reducing the need for deep system knowledge.
- Robust Constraint Handling with Uncertainty: Incorporating GP prediction variance into constraint formulation allows the controller to act more cautiously and conservatively, leading to safer operation, especially when model uncertainties are present.
- Reduced Information Requirement: The GP-MPC requires less information about the PEFC system (only VFC and PH₂) compared to traditional MPC, making it more practical and potentially more economical in real-world implementations.
- Trade-off between Performance and Simplicity: While the GP-MPC exhibits slightly higher overshoot and slower response times compared to a physics-based MPC, its advantages in terms of reduced modeling effort and information requirements are significant.
Key Results
- The GP-MPC successfully maintained the fuel cell voltage at the target 48 V under workload disturbances ranging from 110-120 A.
- The safety constraint for hydrogen pressure (limited to 2.5 atm) was satisfied throughout all tested scenarios.
- Compared to a physics-based MPC, the GP-MPC showed a 43% higher overshoot and a 25% slower response time.
- The GP-MPC's input control actions (QH₂ and Qair) were observed to be more conservative than the physics-based MPC, particularly when increasing QH₂, due to the added constraint that accounts for prediction uncertainty.
Limitations and Future Work
- Overshoot and Response Time: The GP-MPC exhibited a higher overshoot and slower response time compared to the physics-based MPC.
- Conservative Control Actions: The conservative nature of the GP-MPC's control actions, while beneficial for safety, might lead to slightly less optimal performance in terms of speed.
- Future Work: The paper suggests that the GP-MPC framework provides a satisfactory constraint handling result under model error and linearization error. Further tuning of the parameter 'a' in the constraint formulation is mentioned as a way to balance performance and safety requirements. The authors also highlight the potential for this data-driven approach to be more economical due to reduced sensor requirements.
Summary updated