Model Predictive Control with NARX Metamodels: Smarter Torque Requests for Fuel Cell Vehicles

As electrification expands across commercial fleets, engineers are rethinking power management to deliver efficiency, drivability, and robustness under real‑world conditions. By anticipating what the vehicle will need moments ahead, Model Predictive Control (MPC) reduces energy waste, smooths transients, and keeps the power source operating in its most efficient window. MPC paired with a dynamic, data-driven metamodel that learns from system behavior and real-time inputs, offers a practical, high‑fidelity path to predictive torque control.

Why MPC for Vehicle Power Management?

Traditional rule‑based strategies react to the present state; MPC looks ahead. At each control interval, MPC forecasts how the vehicle will evolve over a short horizon and then selects control actions that best meet targets such as speed tracking, efficiency, and constraint compliance.

For fuel cell vehicles, this approach reduces abrupt power swings, improves stack utilization, and maintains consistent operation on grades and transients.

The persistent challenge is building a predictive model that is fast enough for real‑time use and accurate across operating conditions. Data‑driven metamodels such as NARX neatly fill this gap: they learn complex system dynamics from well‑designed experiments and then execute quickly in deployment.

Building the NARX Metamodel in GT‑SUITE

Development begins with a glider model in GT-SUITE that represents the vehicle’s body, tires, axles, and differential capturing the vehicle’s longitudinal dynamics without the added complexity of the power source.

A Design of Experiments (DOE) is constructed to cover all the operating conditions, combining transient drive‑torque profiles with variations in vehicle mass and road grade while measuring the resulting speed response.

Using GT‑SUITE’s Machine Learning Assistant, multiple metamodel candidates are trained and evaluated, and the NARX configuration with the best fit and lowest error is selected. Figure 1 displays an example of the torque estimate from the metamodel and actual torque to drive the cycle.

For control, the mapping is inverted so the metamodel becomes a fast torque estimator. The model receives the current speed, mass, and grade and returns the torque required to achieve the desired speed trajectory. Once validated, the metamodel is exported as a Functional Mock‑up Unit (FMU), providing a stable and portable interface for integration into broader simulations and control workflows.

Figure 1: MPC predicted torque vs. required torque

Embedding MPC: Python + FMU Inside GT‑SUITE

To translate this predictive capability into real-time control, the MPC framework built on the NARX metamodel is implemented within a continuous control loop. The predictive loop is implemented by embedding the FMU in a Python Function template within GT‑SUITE. At each look‑ahead step, the script collects the current state—torque, speed, elevation, and position—reads the desired speed and elevation at the look‑ahead horizon, and calls the FMU to compute the torque required to meet the target under the present mass and grade. That torque request is then converted to power, forming the input power request to the fuel cell system.

Operationally, the GT‑SUITE plant model pauses at the prescribed interval, calls the FMU, retrieves the predicted torque, and resumes with the updated command. The result is a closed loop that continuously anticipates road‑load changes—especially elevation—and requests just enough torque to stay on target while avoiding inefficient transients.

Case Study: Model Predictive Control for Fuel Cell Trucks

To quantify benefits, the MPC strategy was tested against a conventional rule‑based controller on a commercial fuel cell delivery truck operating over the Davis Dam cycle, a demanding route known for sustained grades and thermal stress.

The outcome was clear: MPC achieved approximately 5.6% reduction in hydrogen (H₂) consumption versus the rule‑based baseline. The improvement stems from predictive torque smoothing, which minimizes inefficient power excursions, and from anticipating grade changes, which prevents over‑ or under‑commanding torque as the terrain shifts.

Figure 2: Commercial fuel cell delivery truck model

Figure 3: Hydrogen consumption comparison—MPC vs. rule‑based

Why MPC with NARX Improves System Performance

Speed and fidelity are the central advantages. A NARX metamodel trained on a well‑constructed DOE captures nonlinear vehicle dynamics with high accuracy while remaining lightweight enough for control systems. Exporting the model as an FMU creates a consistent interface that is easy to integrate and reuse across programs. Embedding the metamodel inside GT‑SUITE aligns plant and controller assumptions, preserving system‑level realism throughout development. Finally, translating predicted torque into fuel‑cell power requests keeps the stack operating in a more efficient region, which converts directly into reduced hydrogen use, lower operating costs, and improved durability.

How to Implement MPC with NARX: A Step-by-Step Workflow

Teams aiming to replicate or extend this workflow can follow a straightforward path. First, scope the DOE to reflect duty cycles, covering mass, grade, and torque transients representative of your routes. Next, train and validate the NARX metamodel with the Machine Learning Assistant, selecting the fit that balances accuracy and robustness on held‑out scenarios. Then, export the model as an FMU and embed it via the Python Function template to implement MPC. Finally, measure what matters by comparing energy use, tracking error, and component operating windows across representative cycles, including the kinds of grades and ambient conditions that vehicles encounter.

Closing Thoughts

Predictive torque control with a NARX metamodel brings practical, real‑time intelligence to commercial fuel cell vehicles. The Davis Dam study underscores tangible efficiency gains—about 5.6% H₂ savings—while enhancing drive quality and operational consistency. As fleets scale and routes diversify, MPC’s ability to anticipate demand and right‑size power makes it a compelling addition to modern vehicle controls.

To explore how this approach can be applied to your systems, connect with our team to see how MPC with NARX can be integrated into your workflow. Visit our Digital Twin and Machine Learning solution pages to learn more about advanced modeling capabilities, and follow us on LinkedIn to stay updated on the latest developments in system simulation and control.