Hybrid NMPC-DNN-PI Control Architectures

• Hybrid NMPC-DNN-PI is a control strategy that integrates offline NMPC policies approximated by DNNs with PI corrections to mitigate model mismatch and disturbances.
• The architecture encompasses variants where the DNN acts as a policy surrogate, process model, or residual learner, each offering distinct feedback and computational advantages.
• Empirical studies demonstrate significant computation speedups and enhanced tracking performance in applications such as exoskeleton squat assistance and hybrid electric vehicles.

to=arxiv_search 招商总代 天天爱彩票是json {"query":"Hybrid NMPC-DNN-PI exoskeleton squat movements (Aliyari et al., 30 Sep 2025)", "max_results": 5} to=arxiv_search ӡамեծილი 彩神争霸输钱json {"query":"(Park et al., 2024) Deep Neural Network NMPC for Computationally Tractable Optimal Power Management of Hybrid Electric Vehicle", "max_results": 5} to=arxiv_search 平台直属 ฝ่ายขายข่าวjson {"query":"(Gordon et al., 2023) Introducing a Deep Neural Network-based Model Predictive Control Framework for Rapid Controller Implementation", "max_results": 5} to=arxiv_search 彩神争霸苹果 北京赛车微信json {"query":"(Sathyanarayanan et al., 2023) Deep Neural Network based Optimal Control of Greenhouses", "max_results": 5} to=arxiv_search 天天中彩票网站ിയ qq上json {"query":"(Chee et al., 2022) Learning-enhanced Nonlinear Model Predictive Control using Knowledge-based Neural Ordinary Differential Equations and Deep Ensembles", "max_results": 5} to=arxiv_search 娱乐平台招商 天天彩票怎么json {"query":"(Asadi, 2021) Constrained Deep Learning Based Nonlinear Model Predictive Control", "max_results": 5} Hybrid NMPC-DNN-PI denotes a class of control architectures that combine nonlinear model predictive control (NMPC), deep neural networks (DNNs), and proportional-integral (PI) action within a single feedback structure. In the narrowest usage, the term refers to an NMPC policy learned offline by a DNN and corrected online by a PI term, as in exoskeleton squat assistance (Aliyari et al., 30 Sep 2025). In a broader research sense, it also encompasses closely related architectures in which NMPC acts as an offline policy generator, a DNN replaces the online optimizer, and PI-like or other simple feedback layers regulate slow variables, reject mismatch, or provide fallback behavior, as seen in hybrid-electric vehicles, HCCI combustion control, greenhouse control, constrained learning-based MPC, and learning-enhanced NMPC with neural residual models (Park et al., 2024, Gordon et al., 2023, Sathyanarayanan et al., 2023, Asadi, 2021, Chee et al., 2022).

1. Conceptual scope and lineage

Hybrid NMPC-DNN-PI is not a single canonical algorithm. The literature represented here contains at least three distinct placements of the DNN. In one pattern, the DNN approximates the NMPC control law directly and is executed online as a fast explicit policy; this is the dominant interpretation in the HEV, greenhouse, and exoskeleton studies (Park et al., 2024, Sathyanarayanan et al., 2023, Aliyari et al., 30 Sep 2025). In a second pattern, the DNN is not the controller but the prediction model embedded inside the NMPC optimization, as in HCCI combustion control, where an LSTM-based engine model is solved online inside the optimal control problem (Gordon et al., 2023). In a third pattern, the DNN models only the residual dynamics or model mismatch and is integrated into an otherwise standard NMPC formulation, as in the KNODE ensemble framework (Chee et al., 2022).

A common misconception is to treat all “NMPC-DNN” architectures as interchangeable. They are not. A DNN policy surrogate replaces the optimizer; a DNN process model keeps the optimizer but changes the prediction model; a residual DNN augments a nominal model. The PI component is equally heterogeneous. Only the exoskeleton paper implements an explicit additive PI overlay on top of the DNN output (Aliyari et al., 30 Sep 2025). The HEV work instead uses an equivalent-factor feedback on battery state of charge (SOC) that plays a PI-like outer-loop role without being an actual PI controller (Park et al., 2024). The HCCI and greenhouse papers do not implement PI within the reported controller stacks (Gordon et al., 2023, Sathyanarayanan et al., 2023).

2. Principal architectural patterns

The main architectural variants can be summarized as follows.

Architecture	Role of DNN	Role of PI or simple feedback
Policy-surrogate NMPC	Learned approximation of the NMPC control law	Additive PI in exoskeleton; SOC feedback shaping in HEV; none in greenhouse
Model-in-the-loop NMPC	Learned prediction model inside the online NMPC	No explicit PI in reported formulation
Residual-learning NMPC	Learned correction to a nominal dynamics model	No explicit PI; bias-correction is model-based

In the policy-surrogate pattern, NMPC is first run offline to generate optimal state-action data, and the DNN is trained in supervised fashion to approximate the receding-horizon policy. At runtime the DNN receives the current measured state together with references, previews, or disturbance information, produces a control sequence or control vector, and only the first control move is applied. The HEV paper uses current SOC and a short-term predicted acceleration sequence to predict the optimal torque distribution sequence; the greenhouse paper uses a 16-dimensional vector comprising current state, reference state, reference input, and disturbance forecast to predict actuator commands; the exoskeleton paper uses posture, velocity, and interaction information to predict joint torques (Park et al., 2024, Sathyanarayanan et al., 2023, Aliyari et al., 30 Sep 2025).

In the model-in-the-loop pattern, the DNN does not replace the optimizer. The HCCI framework embeds an LSTM surrounded by fully connected layers inside the NMPC state-space model and solves the OCP online using acados on an ARM Cortex-A72. Here the hybridization is between NMPC and a learned nonlinear process model, not between a DNN policy and PI correction (Gordon et al., 2023).

In the residual-learning pattern, the DNN corrects the mismatch between a nominal physics model and the true plant. The KNODE ensemble paper learns a neural residual term, combines several such models in an ensemble, and inserts the resulting corrected model into NMPC. This architecture is structurally close to offset-free or disturbance-augmented MPC, although no explicit PI term is introduced (Chee et al., 2022).

3. Control formulations and feedback structure

The NMPC component is consistently formulated as a constrained finite-horizon optimal control problem. In the exoskeleton application, NMPC minimizes

$$J = (y_a - y)^T R_1 (y_a - y) + \Delta u^T R_2 \Delta u$$

subject to discrete nonlinear dynamics, incremental-input structure, and the torque-increment bound $|\Delta u_j(k)| \le 5\,\text{Nm}$ for ankle, knee, and hip, with prediction and control horizons $P=M=3$ (Aliyari et al., 30 Sep 2025). In the HEV application, the NMPC cost sums fuel consumption and an SOC-weighted electrical term over a prediction horizon, subject to SOC dynamics, SOC bounds, and the torque-split constraint $0 \le u(k|t) \le 1$ (Park et al., 2024). In the HCCI application, the online OCP tracks IMEP and CA50 while penalizing NOx, fuel and water use, and input increments under actuator and combustion constraints (Gordon et al., 2023).

The explicit Hybrid NMPC-DNN-PI law reported for exoskeleton squat assistance is

$$T_R(k) = \theta(k) + K_P\big(F_{\text{int}}(k)-F_{\text{int}}^{a}(k)\big) + K_I \sum_{i=1}^{k}\big(F_{\text{int}}(i)-F_{\text{int}}^{a}(i)\big),$$

where $\theta(k)$ is the DNN-predicted torque vector, $F_{\text{int}}(k)-F_{\text{int}}^{a}(k)$ is the interaction-force error, and the gains are chosen as $K_P=0.2\,I_3$ and $K_I=0.13\,I_3$ in simulation (Aliyari et al., 30 Sep 2025). The DNN provides the nominal NMPC-like action; the PI term is a corrective overlay. The paper characterizes the PI as a “fine stage and stabilizer,” and the reported decomposition shows that the DNN term is dominant while the PI term remains smaller in magnitude.

The HEV study provides a distinct but closely related outer-loop mechanism through the equivalent factor

$$\lambda(k) \approx \lambda_0 + K_a\big(SOC(k)-SOC_{\text{ref}}\big)^2.$$

This is a static nonlinear feedback law that shapes the relative weight of fuel versus electrical energy and indirectly regulates SOC toward charge-sustaining operation. The paper explicitly notes that this structure is conceptually compatible with replacing the static law by a PI form in the equivalent factor, but does not implement such a replacement (Park et al., 2024).

These formulations show that “PI” in Hybrid NMPC-DNN-PI can mean different things. It may be an additive actuator correction, an outer-loop multiplier adjustment, or a low-level loop assumed but not modeled. This suggests that the defining feature is not a specific signal path but the coexistence of optimal predictive structure, learned approximation, and low-order error integration within the same closed loop.

4. Data generation, neural architectures, and training

The policy-surrogate literature relies on offline NMPC data generation at scale. For HEV energy management, online NMPC is first used to collect a dataset of 7,500 samples, each formed from the current SOC and a short acceleration preview sequence as inputs, and the optimal torque-distribution sequence as output. The DNN is a feed-forward fully connected network with 3 hidden layers, initialized with $|\Delta u_j(k)| \le 5\,\text{Nm}$0 neurons per layer and trained by Levenberg–Marquardt with backpropagation using mean squared error (Park et al., 2024).

For greenhouse control, the lower-level tracking NMPC is solved over approximately 8 months of simulated operation to produce about 285,000 input-output pairs. The learned mapping is

$|\Delta u_j(k)| \le 5\,\text{Nm}$1

The resulting DNN is fully connected, uses 5 hidden layers, contains about 4,000 parameters, requires less than 14 kB of memory, and is trained with Adam on mean squared error after normalization and hyperparameter search with TensorFlow and KerasTuner (Sathyanarayanan et al., 2023).

For exoskeleton squat movements, more than $|\Delta u_j(k)| \le 5\,\text{Nm}$2 training samples are generated from offline NMPC solutions obtained across 25 squat trajectories, 9 human weight-height combinations, and 6 squatting speeds. The DNN is a feedforward fully connected network with 3 hidden layers of 50 neurons each, bipolar sigmoid activations, input scaling to $|\Delta u_j(k)| \le 5\,\text{Nm}$3, and RPROP training. The reported training MSE is approximately $|\Delta u_j(k)| \le 5\,\text{Nm}$4 after 275 epochs (Aliyari et al., 30 Sep 2025).

The model-in-the-loop HCCI architecture uses a different training target. There the DNN learns combustion dynamics rather than the control law. The network contains 7 hidden layers, including 6 fully connected layers and 1 LSTM layer, for a total of 2260 learnable parameters. It is trained on 65,000 consecutive engine cycles, and the reported validation normalized RMSEs are $|\Delta u_j(k)| \le 5\,\text{Nm}$5 for IMEP, $|\Delta u_j(k)| \le 5\,\text{Nm}$6 for NOx, $|\Delta u_j(k)| \le 5\,\text{Nm}$7 for MPRR, and $|\Delta u_j(k)| \le 5\,\text{Nm}$8 for CA50 (Gordon et al., 2023).

These examples establish two dominant supervised-learning regimes. One learns $|\Delta u_j(k)| \le 5\,\text{Nm}$9 from an NMPC oracle; the other learns $P=M=3$0 and leaves optimization online. Hybrid NMPC-DNN-PI, in the strict exoskeleton sense, belongs to the first regime.

5. Empirical performance and computational characteristics

A central motivation for Hybrid NMPC-DNN-PI is the mismatch between NMPC performance and real-time computational budgets. The exoskeleton study reports that standard NMPC requires about 60 ms per step on a Ryzen 3 3200U, whereas the control period is 2 ms, making direct online use infeasible. The Hybrid NMPC-DNN-PI controller runs in approximately 0.04 ms per step, corresponding to a 99.93% reduction in computational cost. Under unseen conditions—an 80 kg, 1.90 m human, a 1.75 s squat cycle, and the disturbance $P=M=3$1—the hybrid controller tracks the desired interaction torques well, whereas NMPC-DNN alone is reported to fail to track the desired signals. In the same studied case, the RMS values of human joint torques are reduced by 30.9% at the ankle, 41.8% at the knee, and 29.7% at the hip (Aliyari et al., 30 Sep 2025).

The HEV study shows a similar computational transition from online NMPC to learned explicit control. On a dSPACE MicroAutoBox III with a 10 ms production-standard control period, the online NMPC requires 14.80 ms minimum, 18.32 ms average, and 178.67 ms maximum per step, while DNN-MPC requires 0.063 ms minimum, 0.068 ms average, and 0.072 ms maximum. In simulation, NMPC yields a 4.76% fuel reduction versus the rule-based strategy at comparable final SOC, and DNN-MPC closely approximates NMPC within the trained region while remaining competitive on moderately untrained profiles (Park et al., 2024).

The greenhouse study reports mean NMPC computation times of roughly 155.76–162.47 ms and DNN inference times of roughly 0.0203–0.0227 ms, a difference of about four orders of magnitude. Under a 10% disturbance uncertainty, the reported net cost is 68.76 € for NMPC and 103.08 € for DNN; under 20% uncertainty, 76.44 € for NMPC and 101.63 € for DNN; under 30% uncertainty, 70.25 € for NMPC and 99.21 € for DNN. The paper notes that with disturbance uncertainty the DNN policy can achieve higher profit and lower operating cost than the online NMPC baseline, while maintaining similar climate trajectories and hard constraint satisfaction in simulation (Sathyanarayanan et al., 2023).

The HCCI paper occupies a different operating point in the design space. Because the DNN is used as the model rather than the policy, online optimization is retained. Even so, acados solves the DNN-based NMPC in 1.4 ms on a Raspberry Pi 400, well within the approximately 22 ms available for combustion control, and the reported IMEP RMS tracking error is 0.133 bar while process constraints are observed (Gordon et al., 2023).

Taken together, these results indicate that the learned-policy variant is primarily a computational tractability strategy, whereas the learned-model variant is a model-accuracy strategy. Hybrid NMPC-DNN-PI adds a robustness layer to the first strategy.

6. Robustness, guarantees, limitations, and extensions

The chief limitation motivating Hybrid NMPC-DNN-PI is that NMPC-DNN alone is only as reliable as its training distribution. The exoskeleton study makes this point explicitly: disturbances, model errors, or operating conditions outside the training data can produce large tracking errors and even instability in pure NMPC-DNN, and the PI overlay is introduced precisely to compensate residual error in the interaction force (Aliyari et al., 30 Sep 2025). The HEV study reaches a closely related conclusion from a different application: performance is good on trained or similar speed profiles, but degrades for substantially untrained profiles, with accumulating errors and SOC deviations, and no stability analysis or online fallback is provided (Park et al., 2024).

The absence of formal guarantees is common in the policy-surrogate literature represented here. The greenhouse work does not include an explicit safety filter, switching logic, or PI backup; hard constraints are inherited from NMPC only through imitation and empirical generalization (Sathyanarayanan et al., 2023). The HCCI work likewise does not add tube MPC, an explicit disturbance model, or PI/PID layers; robustness is handled through conservative constraints, short horizons, and receding-horizon feedback (Gordon et al., 2023).

By contrast, the abstract of “Constrained Deep Learning Based Nonlinear Model Predictive Control” states that constrained DNN training can enforce MPC constraints effectively, and that recursive feasibility and robust stability conditions, as well as probabilistic feasibility and optimality empirical guarantees, are derived for learning-based MPC (Asadi, 2021). The KNODE ensemble paper also provides sufficient conditions guaranteeing asymptotic stability of the closed-loop system for learning-enhanced NMPC (Chee et al., 2022). This suggests that rigorous theory exists for some hybrid NMPC-DNN configurations, but the specific additive DNN-plus-PI overlay used in Hybrid NMPC-DNN-PI remains primarily an empirical robustness device in the currently cited application set.

Several extension paths recur across the papers. The HEV study notes the value of supervisory logic or fallback to NMPC outside the training envelope and suggests incorporating stochastic traffic models such as Markov processes (Park et al., 2024). The exoskeleton study suggests other network architectures, including CNN and LSTM, and naturally motivates adaptive or gain-scheduled PI because the reported gains are fixed and no anti-windup is described (Aliyari et al., 30 Sep 2025). The greenhouse work points toward embedded implementation on low-cost hardware, where PI-type actuator loops or safety backups would be operationally natural (Sathyanarayanan et al., 2023). Across applications, the recurring design principle is stable division of labor: NMPC provides the constrained optimal template, DNN provides computational tractability or model correction, and PI provides low-order robustness where the learned component is weakest.