Training Fuel Cell Energy Managers on the Sensors They Will Actually Have

A fuel cell hybrid vehicle carries two power sources: a hydrogen fuel cell and a battery. At every control step, software decides how much power to draw from each. Get that split right and the car uses less hydrogen, recovers more braking energy, and keeps the fuel cell out of the load swings that wear it out. Most controllers that make this choice, from rule-based schemes to model predictive control and deep reinforcement learning, are designed and scored in simulation. In that setting the controller knows the battery's true charge level, the fuel cell's true efficiency curve, and a clean speed signal. On a real vehicle none of these holds. Battery charge is estimated by integrating a current sensor that has offset and drift, the fuel cell's efficiency degrades as it ages, and speed and demand signals arrive noisy and late.

This note proposes treating that measurement chain as the main source of the simulation-to-reality gap for fuel cell energy management. The method is to distill a controller that sees everything into one that sees only what a real controller would see. The teacher is dynamic programming (an offline solver that finds the best power split over a whole known drive cycle) run with privileged access to true battery charge and true fuel cell health. The student is a small neural network. It receives only a short history of realistic, corrupted measurements: terminal voltage and current, estimated charge, and speed. During training, the corruptions are temporally correlated and biased in the way real automotive sensors fail, and fuel cell degradation parameters are randomized alongside them. The central question is whether such a student can match a carefully built "estimate first, then optimize" pipeline on hydrogen use while violating battery limits less and cycling the fuel cell less. If it cannot, the answer is still worth having.

Core Claim

Under realistic sensor bias and fuel cell aging, a student policy distilled from a privileged dynamic-programming teacher will do three things. It will keep equivalent hydrogen consumption within a small margin of the modular pipeline (an extended Kalman filter for charge, online identification of the fuel cell's voltage–current curve, and adaptive equivalent consumption minimization). It will reduce charge-limit violations and high-frequency fuel cell load cycling relative to that pipeline. And it will run at the sub-millisecond step times needed on an automotive control unit. The part that is new is where the robustness comes from. Prior fuel cell work filters one clean kinematic signal or randomizes drive cycles. This proposal instead randomizes the electrochemical sensing chain and the plant's aging, and uses asymmetric information, where the teacher sees true states and the student sees measurements. The student is expected to infer charge bias and stack health from the voltage–current history rather than trusting a single estimated number.

What Prior Work Leaves Open

The energy management literature supplies the teachers and baselines, and it motivates the gap. The equivalent consumption minimization strategy converts battery power into equivalent fuel through a weighting factor. Jung et al. extend it with a receding-horizon search over that factor and validate their powertrain model on a dynamometer. They assume the controller knows its states. Li et al. (2024) come closest on the noise side. They denoise the speed trace with a wavelet transform, feed it to a recurrent network that recognizes driving conditions, and switch between precomputed and online model predictive control. Their noise handling covers only the kinematic input, and their plant is ideal and non-degrading. Li et al. (2022) show that injecting noise into network parameters, rather than into actions, speeds transfer of a reinforcement learning controller from UDDS and FTP75 to NEDC. That "transfer" is across clean synthetic schedules, not from simulation to hardware. Dong et al. pair an upper-level reinforcement learner for speed planning with dynamic programming for the power split, assuming perfect perception.

Two further papers constrain the design because they tie efficiency to health. De Lorenzo et al. and Shi et al. show that load transients and start–stop events degrade the stack. A controller that recovers hydrogen efficiency by chattering the fuel cell has therefore not really won, so load cycling must be a primary metric.

The robotics literature supplies the method. Kaufmann et al. distill a privileged model predictive controller into a student that sees abstracted visual-inertial features and transfers zero-shot to a real quadrotor. Their raw-pixel students failed completely on new backgrounds, which shows that the choice of student input is decisive. APEX shows how to distill slow trajectory optimizers with dataset aggregation, where the student's own visited states are relabeled by the teacher. We reuse that aggregation loop. Two other works are relevant but not adopted. Karnan et al. correct simulator dynamics from real transitions through a learned action transformer. We do not assume enough real driving data for that kind of grounding, so it stays a later extension. Michelmore et al. gate driving commands on Bayesian uncertainty but report that Monte Carlo dropout collapsed in rain, and their sampling cost is high. This is why we do not rely on heavy uncertainty machinery inside the control loop.

Research Program

We will build a forward-looking fuel cell hybrid model, called the training plant, with three parts. The fuel cell has a parameterized polarization curve, the curve relating stack voltage to current. The battery is an equivalent-circuit model with open-circuit-voltage hysteresis. The sensing layer produces what the control unit actually receives. That sensing layer is the core artifact. It includes current-sensor offset and gain error, which integrate into charge-estimate drift. It also includes voltage quantization, controller-network latency and dropped samples on speed and demand, and a slow random walk in fuel cell degradation that lowers and flattens the polarization curve. These are the correlated, non-Gaussian corruptions that the robustness literature identifies as breaking policies trained on independent Gaussian noise. We sample them per episode and let them evolve within each episode.

For each training episode, dynamic programming runs on the true states and the true degraded plant, which gives optimal power-split labels. The student is a small recurrent or temporal-convolution network over roughly the last tens of seconds of measurements. It is trained by imitation with dataset aggregation, so the teacher relabels the states the student actually reaches, including the drifted-charge states that clean imitation never visits. The training loss adds a penalty on the rate of change of fuel cell power, following the degradation evidence above. A thin hard safety layer clips commands that would push measured charge outside its band. That layer is also ablated to check that it is not doing all the work.

Evaluation happens on a held-out reality-proxy plant. It uses a different fuel cell model family and a battery model with thermal effects, and its sensor corruptions are drawn outside the training ranges. We also run a hardware-in-the-loop check with a real battery pack and current sensor, where a bench is available. We are explicit that the file contains no physical fuel cell data. For that reason the headline claims are stated against the reality-proxy plant, and the hardware result is treated as confirmatory.

Baselines

The naive reference point is a thermostat-style rule-based controller and a plain equivalent consumption minimization controller with a fixed weighting factor. Both consume raw measured charge. They show how much the measurement chain costs a controller that ignores it. A privileged dynamic-programming solution on the reality-proxy plant gives the lower bound on hydrogen use. Every method is reported as a percentage gap to it.

The strongest existing method from the literature is the wavelet-plus-recurrent dual model predictive controller of Li et al. (2024). It is the only fuel cell controller in the file that explicitly handles noisy inputs. On WLTC they report total equivalent hydrogen consumption of 1151.163 g, a 20.614% saving over a rule-based strategy, 43.30% of operation in the fuel cell's high-efficiency region, and an average step time of about 2.320 × 10⁻⁴ s. On UDDS they report 1162.744 g and a 5.972% saving. Those absolute numbers belong to their vehicle model. We will therefore reimplement their controller on our training plant, tuned on the same cycles with the same tuning budget as the student, and compare percentage gaps to dynamic programming.

The strongest principled competitor is not in the file, and we construct it ourselves. It is an extended Kalman filter for battery charge, recursive least-squares identification of the fuel cell polarization curve, and adaptive equivalent consumption minimization that uses both estimates. This modular pipeline is what an automotive engineer would ship. It receives the same sensor streams, is tuned with the same number of simulation hours as the student, and its filter noise settings are fitted on the training plant only.

A reinforcement learning baseline follows Li et al. (2022). It is DDPG with parameter-space noise, trained on the same randomized plant, with the same environment-step budget as the teacher-labeling budget. Their reported figure for this approach is convergence in 22 iterations versus 36 from scratch when transferring to NEDC. No reported result on Lacuna covers this method under sensor bias.

Three ablations isolate the proposal's ingredients. The first trains the same student on independent Gaussian noise instead of the structured sensing chain. The second gives the student only the current measurement instead of a history. The third removes the privileged teacher by training the student with reinforcement learning under the same randomization.

The proposal counts as beating the baselines if it meets three conditions on the reality-proxy plant, averaged over held-out cycles and corruption draws. Its hydrogen gap to dynamic programming must be within one percentage point of the modular pipeline. It must reduce charge-band violations and fuel cell power-rate integrals relative to that pipeline. And it must run within the step-time budget reported by Li et al. (2024).

Why not just…?

Why not just put a Kalman filter and online curve fitting in front of an equivalent-consumption controller? This is the real threat, and the proposal is built around it. When charge bias is observable from terminal voltage and the aging is slow, the modular pipeline may match the student. Our prediction is that it degrades when the estimator's model is wrong, for example under hysteresis, thermal effects, or out-of-range sensor gain on the reality-proxy plant. In those cases a confidently wrong estimate is passed straight to an optimizer that trusts it. The bias-sweep experiment below settles this question directly. If the pipeline wins everywhere, we will report that distillation is unnecessary for this problem.

Why not just denoise the inputs with wavelets, as Li et al. do? Wavelet filtering removes high-frequency noise from the speed trace. Charge-estimate drift and aging are low-frequency and biased, and a filter cannot see them because they look like signal. The Gaussian-noise ablation and the reimplemented wavelet controller test whether removing high-frequency noise is enough. We expect it is not once the bias is nonzero.

Why not just train reinforcement learning with domain randomization, as in quadrotor work? Randomization alone worked for Sa et al. on stabilization. The file notes, however, that broad randomization makes policies conservative, and that reinforcement learning controllers for energy management violate charge limits out of distribution. The teacher-free ablation measures exactly this. It keeps the randomization and removes only the privileged labels, so any gap between it and the student is attributable to distillation.

Proposed Experiments

Training uses UDDS and FTP75, plus synthetic segments. Testing uses WLTC and NEDC as held-out cycles, with a real-route replay in the style of the Columbus corridor from Jung et al. where elevation data can be reconstructed.

The decisive experiment is a sweep. On the reality-proxy plant we increase current-sensor offset and fuel cell degradation from zero to beyond the training range. At each level we plot three curves for the student, the modular pipeline, and the reimplemented wavelet controller: hydrogen gap to privileged dynamic programming, charge-band violation rate, and integrated fuel cell power rate of change. The hypothesis is supported if the curves coincide at zero corruption and the student's curves separate favorably as corruption grows, especially past the training range. It is refuted if the modular pipeline tracks or beats the student throughout.

A second check examines what the student learned. We regress true charge bias and stack health from the student's hidden state. If these can be decoded, the student has learned to infer them from voltage–current history. If they cannot, and the student still performs well, it has found a policy that is insensitive to them. Each of those findings would call for a different next step.

Finally, the hardware-in-the-loop run with a physical battery and current sensor checks whether the ranking of methods on the reality-proxy plant carries over when at least the battery-sensing half of the chain is real.