Hydrogen-Diesel Dual-Fuel Engine Operation Using Real-Time GRU-Based Nonlinear Model Predictive Control
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: We're looking at the paper: "Hydrogen-Diesel Dual-Fuel Engine Operation Using Real-Time GRU-Based Nonlinear Model Predictive Control".
Dev: Here's the gist of the paper. Hydrogen-diesel dual-fuel (H2DF) combustion reduces combustion-out CO2 emissions but exhibits nonlinear cycle-tocycle dynamics at high hydrogen energy shares (HES) This work evaluates nonlinear model predictive control (NMPC) with a gated recurrent-unit deep neural network dynamics model for transient H2DF control Compared with production diesel-only control, NMPC improves load-tracking mean absolute error by 27.8% and reduces mean PM by 61.1%, while mean engine-out NOx increases by 105.1% without exhaust-gas recirculation The learned-dynamics NMPC enables practical, real-time, constraint-aware transient H2DF control with substantial diesel substitution Experimental Setup and Data Generation The experimental setup involves a modified four-cylinder Cummins QSB 4.5L diesel engine equipped with a hydrogen port fuel injection (HANA H2200A) in cylinder 1 The NMPC optimization runs on the RPi 400 and communicates with the main engine controller via User Datagram Protocol over Ethernet In-cylinder pressure is measured using a Kistler 6124A piezoelectric pressure transducer, and engine-out particle-mass concentration cPM and NOx concentration cNOx are measured directly after the exhaust valve of cylinder 1 The controller formulation optimizes over the control changes ∆u(k) = u(k) − u(k − 1), allowing for direct penalization of actuator change rates in the cost function How the training signal is built The controller uses a data-driven GRU-DNN dynamics model to capture the strongly nonlinear and engine-cycledependent H2DF combustion behavior Model training is performed on measured engine data generated by pseudo-random binary-sequence excitation of the four control inputs Each measurement run is split chronologically into training, validation, and test subsets using an 80%/15%/5% split without shuffling to preserve the temporal structure required by the GRU The prediction performance on the unseen test set shows high accuracy for load and NOx prediction (pmi: RMSE 0.23 bar, NRMSEclip 3.0%; cNOx: RMSE 40 ppm, NRMSEclip 4.0%) Where the gains come from The proposed NMPC outperformed the production diesel-only baseline controller while tracking an unseen, rapidly changing load command over 4,900 engine cycles and enabling partial substitution of diesel fuel with hydrogen The mean absolute error for pmi tracking falls to 0.117 bar (−27.8% vs.
Rosa: diesel) and an NRMSE of 3.07%, corresponding to a 13.6% improvement Soot emissions were significantly reduced, with the mean cPM dropping by 61.1% to 0.21 mg/m3 vs.
Dev: diesel The HES reaches as high as 55.1% and averages 39.7%, demonstrating that the framework can be steered towards aggressive hydrogen substitution through cost-function adjustment alone Robustness and Flexibility The NMPC maintains constraint satisfaction while executing the necessary control actions, even when subjected to perturbations like white Gaussian noise injected into the pmi feedback channel The controller executes in 3 to 7 ms per engine cycle without deadline violations on the embedded target Empirical robustness analyses demonstrate that the proposed NMPC maintains stable and constraint-compliant operation when subjected to disturbances affecting both modeled dynamics and unmodeled dynamics, such as model–plant mismatch induced by varying engine speed The framework demonstrates flexibility by allowing the controller to be steered between low and high hydrogen substitution modes through cost-function adjustment alone without model reidentification The work presented a GRU-NMPC framework for H2DF engine control, targeting simultaneous loadtracking improvement and multi-objective emission management under safety constraints The proposed NMPC outperformed the production diesel-only baseline controller while tracking an unseen, rapidly changing load command over 4,900 engine cycles and enabling partial substitution of diesel fuel with hydrogen The full controller executed in 3 to 7 ms per engine cycle without deadline violations, demonstrating real-time capability on the embedded target The quantitative results are limited to a naturally aspirated single-cylinder test configuration at 1500 min−1, without EGR Active follow-on work at the UofA targets a 6.7 L Cummins multi-cylinder H2DF platform and ultimately full-hydrogen operation Page 1 --- Hydrogen–Diesel Dual-Fuel Engine Operation Using Real-Time GRU-Based Nonlinear Model Predictive Control Alexander Winklera, Vasu Sharmaa, Julian Bedeia, Edward Sperlingb, Charles Robert Kochb, David Gordonb, Jakob Anderta, Abstract Hydrogen-diesel dual-fuel (H2DF) combustion reduces combustion-out CO2 emissions but exhibits nonlinear cycle-tocycle dynamics at high hydrogen energy shares (HES) This work evaluates nonlinear model predictive control (NMPC) with a gated recurrent-unit deep neural network dynamics model for transient H2DF control Trained on 99,800 engine cycles, the model predicts indicated mean effective pressure, nitrogen oxides (NOx), particulate matter (PM), and maximum pressure-rise rate Single-cylinder Cummins 4.5 L experiments follow an unseen 4,900-engine-cycle trajectory Compared with production diesel-only control, NMPC improves load-tracking mean absolute error by 27.8% and reduces mean PM by 61.1%, while mean engine-out NOx increases by 105.1% without exhaust-gas recirculation Mean and peak HES reach 39.7% and 55.1%; a high-hydrogen setting achieves 77.8% peak HES without constraint violations Robust to feedback noise and model-plant mismatch and executing in 3 to 7 ms per engine cycle on low-cost embedded hardware, learned-dynamics NMPC enables practical, real-time, constraint-aware transient H2DF control with substantial diesel substitution Keywords: Hydrogen-diesel dual-fuel engine, Nonlinear model predictive control, Gated recurrent unit, Recurrent neural network, Engine control, Real-time embedded control Page 2 --- Road transport remains a major contributor to global CO2 emissions and decarbonizing medium- and heavyduty applications is particularly challenging due to high energy demand and long asset lifetimes
1, 2, 3: A practical transitional option is hydrogen-diesel dual-fuel (H2DF) combustion by retrofitting existing compression-ignition engines The literature on H2DF operation consistently reports strong particulate matter (PM) and CO2 reductions with increasing hydrogen substitution, while NOx remains strongly load- and calibration-dependent Most published studies remain focused on steady-state operation and calibrationoriented control (maps, feedforward structures, or lightly tuned stock-engine-control-unit strategies), with limited evidence on closed-loop transient tracking under practical constraints This work addresses this gap using a gated recurrent unit (GRU)-based deep-neural-network (DNN) dynamics model embedded in NMPC, hereafter referred to as GRUNMPC, to investigate H2DF operation under practical constraints The controller is validated on an unseen 4,900-engine-cycle transient trajectory and demonstrates robust transient load tracking, a peak hydrogen energy share (HES) of 77.8%, constraint satisfaction, and real-time execution on low-cost embedded hardware Relative to prior H2DF DNN-NMPC, the distinguishing contribution is the integration of recurrent cycle-to-cycle dynamics with embedded experiments that test robustness to measurement noise and model–plant mismatch Page 3 --- Figure 1: Workflow connecting H2DF operation, GRU-DNN dynamics modelling, NMPC design, and embedded deployment.
Rosa: A schematic of the experimental setup is shown in Fig.
Title and authors: Rosa: So we're looking at this paper called "Hydrogen-Diesel Dual-Fuel Engine Operation Using Real-Time GRU-Based Nonlinear Model Predictive Control". It sounds like a lot of complex control systems are being put together here.
Dev: Yeah, it’s about using a gated recurrent unit deep neural network dynamics model inside a nonlinear model predictive control framework to handle hydrogen diesel dual fuel engines. That’s the core idea.
Rosa: It tackles the problem that hydrogen diesel combustion cuts out carbon dioxide emissions but it has these really nonlinear dynamics when you have high hydrogen energy shares, or HES.
Dev: Exactly. The paper is looking at how to manage those kinds of nonlinear cycle-to-cycle dynamics in real time using a GRUNMPC system.
Taro: So, what does this mean for the actual operation? Is it just a theoretical model prediction?
Rosa: No, it’s designed for transient control. It’s about managing the engine when you’re changing loads quickly and trying to keep things stable under those high hydrogen energy shares.
Dev: Right. The authors trained their dynamics model on a massive dataset of ninety-nine thousand eight hundred engine cycles to capture this behavior accurately enough for real-time prediction <ref:2609.15819#pg1>.
Taro: That training data sounds pretty extensive for something focused on specific engine conditions like the Cummins QSB four point five L test rig they used <ref:2609.15819#pg1>.
Rosa: They used a single-cylinder Cummins four point five L engine, but they validated the system on an unseen trajectory of four thousand nine hundred engine cycles to see how it performs in practice when things get tricky <ref:2609.15819#pg1,single-cylinder Cummins 4.5 L>.
The paper's summary: Dev: Okay, so what’s the actual performance claim? What does this GRUNMPC do compared to just running the production diesel-only control?
Rosa: Well, they found that this nonlinear model predictive control improves load-tracking mean absolute error by twenty seven point eight percent over the diesel baseline.
Dev: That’s a significant improvement for transient response, suggesting better stability when you’re changing what the engine is doing on the fly.
Rosa: They also saw a major reduction in particulate matter, with the mean particle mass concentration dropping by sixty one point one percent compared to diesel operation.
Dev: And they managed to handle nitrogen oxides in a way that actually increased them by over one hundred five point one percent when they weren't using exhaust-gas recirculation. That’s a bit of a trade-off, though.
Taro: So, the paper is showing that you can get better load tracking and lower PM without necessarily having to rely on things like exhaust-gas recirculation to manage NOx.
Rosa: Right. The authors also looked at the hydrogen energy share, or HES. They found that when they steered the cost function correctly, they could reach as high as seventy seven point eight percent peak HES without hitting any constraints.
Dev: That's a big deal because it shows they can actively steer the engine toward aggressive hydrogen substitution using just adjustments to the cost function weights, not by changing their underlying model.
The paper's improvements: Rosa: The main improvement here is that they are integrating these recurrent cycle-to-cycle dynamics directly into the model predictive control loop. This is what makes it different from standard NMPC setups using just a static or simple model.
Dev: They use this gated recurrent unit deep neural network to predict several things at once: indicated mean effective pressure, nitrogen oxides, particulate matter, and maximum pressure rise rate.
Rosa: That means the controller isn't just predicting one variable; it’s predicting the whole combustion picture based on what happened in the previous cycles.
Taro: And that recurrence is what they tested for robustness. They specifically put this GRUNMPC through tests with measurement noise and model–plant mismatch, like varying engine speed between twelve hundred min per minute and eighteen hundred min per minute.
Dev: So, the key contribution isn't just the prediction accuracy; it’s proving that this recurrent structure is robust to those kinds of real-world disturbances.
Rosa: They also showed that they can remove direct indicated mean effective pressure feedback from the model, and even with degraded performance there, they kept hydrogen substitution working well at forty-two point one five percent HES.
Conclusion: Dev: So to wrap up the "Hydrogen-Diesel Dual-Fuel Engine Operation Using Real-Time GRU-Based Nonlinear Model Predictive Control," what’s the big picture for us?
Rosa: Essentially, this work shows that you can get real time, constraint aware transient control for H2DF engines by using this specific model predictive control setup. It handles load tracking and emission management simultaneously.
Taro: I think the robustness part is what really matters for real-world deployment. Being able to handle those noise and mismatch issues without needing a complete re-identification of the entire dynamics model is crucial for safety cases.
Dev: Yeah, and they proved it can run in three to seven milliseconds per engine cycle on low cost embedded hardware, which means it’s practically deployable right now for real time control.
Rosa: It shows that with this GRUNMPC framework, you can achieve substantial diesel substitution by just adjusting the cost function weights. It's a way to manage competing objectives without needing complex re-tuning of the underlying deep learning model.
Taro: I think it’s promising because they didn't just show it works in a perfect lab setting; they tested it against real engine speed variations and noise, which is where most control systems fail in practice.
Dev: So that’s the gist of this paper, "Hydrogen-Diesel Dual-Fuel Engine Operation Using Real-Time GRU-Based Nonlinear Model Predictive Control". It’s a look at how data driven NMPC can provide a path toward practical hydrogen diesel dual fuel operation.
RWTH Aachen University · University of Alberta
eess.SY, cs.SY
Submitted: 2026-09-14
Updated: 2026-11-15
Comments: Submitted to Energy Conversion and Management: X
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 88/100
The gist: Hydrogen-diesel dual-fuel (H2DF) combustion reduces combustion-out CO2 emissions but exhibits nonlinear cycle-tocycle dynamics at high hydrogen energy shares (HES) This work evaluates nonlinear model…
Key concepts
- GRU-DNN Dynamics Model
- This is a deep learning model, specifically a Gated Recurrent Unit (GRU) combined with a Deep Neural Network (DNN), trained on nearly 100,000 engine cycles. It learns the complex, nonlinear way hydrogen and diesel fuel interact during combustion to accurately predict outputs like pressure and NOx in real time.
- Nonlinear Model Predictive Control (NMPC)
- NMPC is an advanced control method that uses a predictive model to look into the future. At every moment, it solves an optimization problem to decide the best sequence of fuel injections (main diesel duration, hydrogen timing) to meet goals like tracking a desired load while respecting physical limits.
- Cycle-to-Cycle Dynamics
- This refers to how the engine's behavior changes from one combustion cycle to the next. The GRU model is specifically designed to capture these nonlinear dynamics, which are crucial because hydrogen ignition sensitivity causes significant variability in performance that a simple static model cannot handle.
- Constraint-Aware Control
- This means the controller actively manages safety limits. It ensures that control actions never violate physical boundaries, such as maximum pressure rise rates or actuator change limits. This allows the system to safely operate at high hydrogen energy shares without causing engine damage.
Terminology
Summary
Hydrogen-diesel dual-fuel (H2DF) combustion reduces combustion-out CO2 emissions but exhibits nonlinear cycle-tocycle dynamics at high hydrogen energy shares (HES) This work evaluates nonlinear model predictive control (NMPC) with a gated recurrent-unit deep neural network dynamics model for transient H2DF control Compared with production diesel-only control, NMPC improves load-tracking mean absolute error by 27. p. 0 4, Aachen, 52074, NRW, Germany
bDepartment of Mechanical Engineering, University of Alberta, 116 St & 85 Ave, Edmonton, T6G 2R3, AB, Canada
Abstract
Hydrogen-diesel dual-fuel (H2DF) combustion reduces combustion-out CO2 emissions but exhibits nonlinear cycle-tocycle dynamics at high hydrogen energy shares (HES). This work evaluates nonlinear model predictive control (NMPC)
with a gated recurrent-unit deep neural network dynamics model for transient H2DF control. Compared with production diesel-only control, NMPC improves load-tracking mean absolute error by 27. 8% and reduces
mean PM by 61. Compared with production diesel-only control, NMPC improves load-tracking mean absolute error by 27. 8% and reduces
mean PM by 61. 8% and reduces
mean PM by 61. 1%, while mean engine-out NOx increases by 105. 8% and reduces
mean PM by 61. 1%, while mean engine-out NOx increases by 105.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Experimental Setup and Data Generation
The experimental setup involves a modified four-cylinder Cummins QSB 4.5L diesel engine equipped with a hydrogen port fuel injection (HANA H2200A) in cylinder 1 The NMPC optimization runs on the RPi 400 and communicates with the main engine controller via User Datagram Protocol over Ethernet In-cylinder pressure is measured using a Kistler 6124A piezoelectric pressure transducer, and engine-out particle-mass concentration cPM and NOx concentration cNOx are measured directly after the exhaust valve of cylinder 1 The controller formulation optimizes over the control changes ∆u(k) = u(k) − u(k − 1), allowing for direct penalization of actuator change rates in the cost function
How the training signal is built
The controller uses a data-driven GRU-DNN dynamics model to capture the strongly nonlinear and engine-cycledependent H2DF combustion behavior Model training is performed on measured engine data generated by pseudo-random binary-sequence excitation of the four control inputs Each measurement run is split chronologically into training, validation, and test subsets using an 80%/15%/5% split without shuffling to preserve the temporal structure required by the GRU The prediction performance on the unseen test set shows high accuracy for load and NOx prediction (pmi: RMSE 0.23 bar, NRMSEclip 3. p. 4 The GRU-DNN dynamics
model provides high accuracy for load and NOx prediction (pmi: RMSE 0. 23 bar, NRMSEclip 3. The GRU-DNN dynamics
model provides high accuracy for load and NOx prediction (pmi: RMSE 0. 23 bar, NRMSEclip 3. 23 bar, NRMSEclip 3. 0%; cNOx: RMSE
40 ppm, NRMSEclip 4. 23 bar, NRMSEclip 3. 0%; cNOx: RMSE
40 ppm, NRMSEclip 4.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the gains come from
The proposed NMPC outperformed the production diesel-only baseline controller while tracking an unseen, rapidly changing load command over 4,900 engine cycles and enabling partial substitution of diesel fuel with hydrogen The mean absolute error for pmi tracking falls to 0.117 bar (−27. p. 8 Using the H2DF
controller, the mean absolute error (MAE) for pmi tracking falls to 0. 117 bar (−27. 6% improvement in NRMSE. This load-tracking improvement results
from the NMPC’s ability to directly incorporate combustion dynamics into the optimization, in contrast to the
diesel-only feedforward map used as the baseline. By substituting a portion of
diesel with port-injected hydrogen, H2DF can reduce soot
and direct CO2 emissions while preserving existing engine
platforms and operational flexibility [10, 11, 12, 13, 14]. The literature on H2DF operation consistently reports strong particulate matter (PM) and CO2 reductions with increasing hydrogen substitution, while NOx
∗Corresponding author. 6% improvement. Soot emissions were significantly reduced,
with the mean cPM dropping by 61. 21 mg/m3 vs. The HES reaches as high as 55. The HES reaches as high as 55. 1% and averages
39.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Robustness and Flexibility
The NMPC maintains constraint satisfaction while executing the necessary control actions, even when subjected to perturbations like white Gaussian noise injected into the pmi feedback channel The controller executes in 3 to 7 ms per engine cycle without deadline violations on the embedded target Empirical robustness analyses demonstrate that the proposed NMPC maintains stable and constraint-compliant operation when subjected to disturbances affecting both modeled dynamics and unmodeled dynamics, such as model–plant mismatch induced by varying engine speed The framework demonstrates flexibility by allowing the controller to be steered between low and high hydrogen substitution modes through cost-function adjustment alone without model reidentification
Conclusion
The work presented a GRU-NMPC framework for H2DF engine control, targeting simultaneous loadtracking improvement and multi-objective emission management under safety constraints The proposed NMPC outperformed the production diesel-only baseline controller while tracking an unseen, rapidly changing load command over 4,900 engine cycles and enabling partial substitution of diesel fuel with hydrogen The full controller executed in 3 to 7 ms per engine cycle without deadline violations, demonstrating real-time capability on the embedded target The quantitative results are limited to a naturally aspirated single-cylinder test configuration at 1500 min−1, without EGR Active follow-on work at the UofA targets a 6.7 L Cummins multi-cylinder H2DF platform and ultimately full-hydrogen operation
--- Page 1 ---
Hydrogen–Diesel Dual-Fuel Engine Operation Using Real-Time GRU-Based Nonlinear Model Predictive Control Alexander Winklera, Vasu Sharmaa, Julian Bedeia, Edward Sperlingb, Charles Robert Kochb, David Gordonb, Jakob Anderta, Abstract Hydrogen-diesel dual-fuel (H2DF) combustion reduces combustion-out CO2 emissions but exhibits nonlinear cycle-tocycle dynamics at high hydrogen energy shares (HES) This work evaluates nonlinear model predictive control (NMPC) with a gated recurrent-unit deep neural network dynamics model for transient H2DF control Trained on 99,800 engine cycles, the model predicts indicated mean effective pressure, nitrogen oxides (NOx), particulate matter (PM), and maximum pressure-rise rate Single-cylinder Cummins 4.5 L experiments follow an unseen 4,900-engine-cycle trajectory Compared with production diesel-only control, NMPC improves load-tracking mean absolute error by 27. p. 0 5 L experiments follow an unseen 4,900-engine-cycle trajectory. Compared with production diesel-only control, NMPC improves load-tracking mean absolute error by 27. 1% without exhaust-gas recirculation. Mean and peak
HES reach 39. Mean and peak
HES reach 39. 1%; a high-hydrogen setting achieves 77.Where the paper says it
Where the paper says it
Where the paper says it
--- Page 2 ---
Road transport remains a major contributor to global CO2 emissions and decarbonizing medium- and heavyduty applications is particularly challenging due to high energy demand and long asset lifetimes [1, 2, 3] A practical transitional option is hydrogen-diesel dual-fuel (H2DF) combustion by retrofitting existing compression-ignition engines The literature on H2DF operation consistently reports strong particulate matter (PM) and CO2 reductions with increasing hydrogen substitution, while NOx remains strongly load- and calibration-dependent Most published studies remain focused on steady-state operation and calibrationoriented control (maps, feedforward structures, or lightly tuned stock-engine-control-unit strategies), with limited evidence on closed-loop transient tracking under practical constraints This work addresses this gap using a gated recurrent unit (GRU)-based deep-neural-network (DNN) dynamics model embedded in NMPC, hereafter referred to as GRUNMPC, to investigate H2DF operation under practical constraints The controller is validated on an unseen 4,900-engine-cycle transient trajectory and demonstrates robust transient load tracking, a peak hydrogen energy share (HES) of 77.8%, constraint satisfaction, and real-time execution on low-cost embedded hardware Relative to prior H2DF DNN-NMPC, the distinguishing contribution is the integration of recurrent cycle-to-cycle dynamics with embedded experiments that test robustness to measurement noise and model–plant mismatch
--- Page 3 ---
Figure 1: Workflow connecting H2DF operation, GRU-DNN dynamics modelling, NMPC design, and embedded deployment. A schematic of the experimental setup is shown in Fig. p. 1 Step 3: Online Nonlinear Model Predictive Controller Design
NMPC States
Augmentation
GRU DNN with
States Augmentation
Input
Output
Optimizer
Objective
Function Constraints
Reference
GRU-DNN
Experimental Setup
Nonlinear Model Predictive Control
NMPC
Pseudo Binary
Random Input Output
Step 1: Data Generation using System Identification
Experimental Setup
Validation of
Engine Model
Training Data Log
Step 2: Training of
Deep Neural Network
Engine Model
GRU-DNN
Figure 1: Workflow connecting H2DF operation, GRU-DNN dynamics modelling, NMPC design, and embedded deployment. A schematic of the experimental setup is shown in
Fig.Where the paper says it
--- Page 4 ---
H2 setpoint Diesel setpoint tcalc ≈ 18 ms @ 1500 min−1 (165 CAD) At 1500 min−1, one four-stroke engine cycle corresponds to 80 ms As shown in Fig. 3, the combustion metrics are available by 165 CAD aTDC and the next injection setpoint is required at 330 CAD aTDC The available online calculation time is therefore tcalc ≈ 18 ms, corresponding to 165 CAD Subsequently, there are four engine controls available to the NMPC, namely the main-diesel duration of injection (DOI), the pre-to-main (P2M) interval, the maindiesel start of injection (SOI), and the hydrogen DOI
--- Page 5 ---
H2DF combustion differs markedly from conventional diesel operation Rising HES levels amplify these effects, leading to steep pressure rises and elevated peak cylinder pressures, reflecting faster combustion and stronger thermal loading H2DF operation also exhibits increased cycle-to-cycle variability due to hydrogen’s ignition sensitivity and mixture dynamics, contributing to combustion instability and noise The injection timings and control setpoints over one combustion cycle are visualized in Fig. 3 The controls for the NMPC are represented in compact form as u(k) = h T [tMain(k) tP2M(k) αMain(k) tH2(k)]iT
--- Page 6 ---
While u(k) represents the physical actuator commands sent to the engine, the NMPC optimizes over the control changes ∆u(k) = u(k) − u(k − 1), namely the input increments rather than their absolute values The corresponding output vector is written as y(k) = h T [pmi(k) cNOx(k) cPM(k) dpmax(k)]iT The GRU-DNN dynamics model provides high accuracy for load and NOx prediction (pmi: RMSE 0.23 bar, NRMSEclip 3.0%; cNOx: RMSE 40 ppm, NRMSEclip 4.0%) The GRU-DNN dynamics model provides moderate accuracy for dpmax (NRMSEclip 12. p. 4 Prediction errors are reported as root
mean square error (RMSE) and normalized root mean
square error (NRMSE); the clipped variant NRMSEclip is
used for the model metrics. The GRU-DNN dynamics
model provides high accuracy for load and NOx prediction (pmi: RMSE 0. 0%), with moderate accuracy for
dpmax (NRMSEclip 12. 1%) and lower but acceptable accuracy for cPM (NRMSEclip 9.Where the paper says it
Where the paper says it
--- Page 7 ---
The NMPC determines control actions by repeatedly solving a finite-horizon optimal control problem (OCP) online The OCP requires a dynamics model that maps the applied inputs to the constrained outputs over the horizon at each sampling instant The GRU-DNN dynamics model takes the place of a physics-based model and supplies the state predictions propagated across the horizon at each sampling instant The NMPC state vector x˜(k) and NMPC output vector y˜(k) are thus defined as
--- Page 8 ---
At each engine cycle k, the NMPC solves a finite-horizon, discrete-time OCP over a prediction horizon of N engine cycles The optimization is written directly in the H2DF-specific OCP-structured nonlinearprogramming form using the augmented dynamics from Eq. 9 The resulting finitehorizon cost is J = X N i=0 h kpmi,i − pmi ref,ik squared qp,mi z Load tracking + kdpmax,ik squared qdp max z Pressure-rise moderation + kcNOx,ik squared qcNOx + kcPM,ik squared qcPM z Emissions minimization + ktMain,ik squared qt Main + ktH2,ik squared qt H2 Fuel consumption and H2 share + ksik1 qs z Fuel consumption and H2 share
--- Page 9 ---
The NMPC demonstrates robust performance in reference tracking and constraint handling while managing the fuelblend ratio The high-utilization run reaches a peak of 77.8% HES, confirming flexible hydrogen/diesel blending The control change-rate constraints set the trade-off between tracking response and actuator-rate limitation
--- Page 10 ---
Production engines generally do not provide direct incylinder pressure feedback such as pmi; the controller must therefore remain effective without this feedback signal A no-feedback GRU-DNN dynamics model was developed by removing the feedback input feature pmi(k−1) from the original model Despite degraded pmi tracking performance, with the NRMSE rising to 5.72%, an 86.3% increase over the feedback-based NMPC, the results remain acceptable The mean HES remains virtually unchanged at 42. p. 9 3% increase over the
feedback-based NMPC, the results remain acceptable. NOx and soot emissions cNOx, cPM are slightly lower
with the no-feedback variant, which can be associated
with shifted operating conditions relative to the feedbackbased NMPC, with mean pmi about 4% lower. The mean
HES remains virtually unchanged at 42. 15%, whereas the
NRMSE rises to 5.Where the paper says it
Where the paper says it
--- Page 11 ---
The NMPC maintains constraint satisfaction while executing the necessary control actions The augmentedstate OCP handles actuator change-rate constraints directly, exposing the trade-off between tracking response and actuator-rate limitation By adjusting the relative penalties on tH2 and tMain, the controller can be steered between low and high hydrogen substitution modes without re-identification or re-tuning of the underlying DNN The HES reflects the primary difference between the three scenarios, ranging from a mean of 23.1% (low) to 44. p. 0 8% peak HES without constraint violations. Robust to
feedback noise and model-plant mismatch and executing in 3 to 7 ms per engine cycle on low-cost embedded hardware,
learned-dynamics NMPC enables practical, real-time, constraint-aware transient H2DF control with substantial diesel
substitution. 1% (low) to
44. 9% (high), with the high-utilization run reaching a peak
of 77.Where the paper says it
Where the paper says it
--- Page 12 ---
The NMPC maintains constraint satisfaction while executing the necessary control actions The augmentedstate OCP handles actuator change-rate constraints directly, exposing the trade-off between tracking response and actuator-rate limitation A key advantage of the NMPC formulation is the direct exposure of its objectives through the cost function weights By adjusting the relative penalties on tH2 and tMain, the controller can be steered between low and high hydrogen substitution modes without re-identification or re-tuning of the underlying DNN The quantitative results are limited to a naturally aspirated single-cylinder test configuration at 1500 min−1, without EGR Thus, active follow-on work at the UofA targets a 6.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
- Real-time transient control for Hydrogen-Diesel Dual-Fuel Engines
The improved system can execute a constraint-aware GRU-NMPC formulation that integrates recurrent cycle-to-cycle dynamics,
which enables robust transient load tracking and constraint satisfaction during H2DF operation. This allows the AI system to manage competing objectives like load tracking, NOx moderation, and maximum pressure rise rate simultaneously in real time.
- Data-Driven Dynamics Modeling for Nonlinear Combustion
The system can utilize a gated recurrent unit (GRU)-based deep neural network (DNN) dynamics model
trained on 99,800 engine cycles to predict key combustion metrics like indicated mean effective pressure, nitrogen oxides, particulate matter, and maximum pressure rise rate. This model provides a data-driven NMPC formulation that integrates recurrent cycle-to-cycle dynamics
into the optimization loop.
- Robustness to Model Mismatch and Noise
The system is enhanced with an embedded experiments that test robustness to measurement noise and model–plant mismatch,
allowing it to maintain performance under perturbations. This includes maintaining load tracking even with injected white Gaussian noise in the pmi feedback channel or varying engine speed (nEng) between 1200 min/min and 1800 min/min.
- Modular Rapid Prototyping Workflow
The AI system can operate within a modular rapid-prototyping workflow with real-time Raspberry Pi (RPi) execution in 3 to 7 ms per engine cycle.
This allows for controller updates without modifying the main engine control software, facilitating rapid iteration from model retraining to closed-loop validation.
Abstract
Hydrogen-diesel dual-fuel (H2DF) combustion can reduce combustion-out CO2 emissions but exhibits nonlinear cycle-to-cycle dynamics at high hydrogen energy shares (HES), while the NOx-particulate matter (PM) trade-off must still be managed. This work evaluates nonlinear model predictive control (NMPC) with a gated recurrent-unit deep neural network dynamics model for transient H2DF control. Trained on 99,800 engine cycles, the model predicts indicated mean effective pressure, nitrogen oxides (NOx), particulate matter (PM), and maximum pressure-rise rate. Experiments on a four-cylinder 4.46 L Cummins engine with hydrogen port injection in cylinder 1 only, naturally aspirated, at 1500 rpm, follow an unseen 4,900-engine-cycle trajectory. Compared with diesel-only operation under the production calibration, the NMPC reduces the load-tracking mean absolute error by 27.8% while reaching a mean HES of 39.7% (peak 55.1%); a high-hydrogen setting achieves 77.8% peak HES without measured constraint violations. The mean engine-out PM concentration is about 60% lower, while the mean engine-out NOx concentration increases by 105.1% without exhaust-gas recirculation. The controller tolerates feedback noise and model-plant mismatch from engine-speed variation, and it executes in 3 to 7 ms per engine cycle on low-cost embedded hardware within a modular rapid-prototyping toolchain. Learned-dynamics NMPC thus enables practical, real-time, constraint-aware transient H2DF control with substantial diesel substitution.
Related papers
- Industrial Load Modeling and Optimization for Market-Based Interaction with Power Systems
- Exact and Approximate Convex Reformulation of Linear Stochastic Optimal Control with Chance Constraints
- Certified Set Convergence for Piecewise Affine Systems
- Symbolic Control for Stochastic Systems via Finite Parity Games
- Multiobjective optimization-based design and dispatch of islanded, hybrid microgrids for remote, off-grid communities in sub-Saharan Africa
- Physiology-informed layered sensing for intelligent human-exoskeleton interaction