Transient Predictive Control (TPC)

• Transient Predictive Control (TPC) is a receding-horizon optimal control paradigm that integrates MPC with explicit transient safety constraints for critical state regulation.
• TPC employs a two-layer structure combining a bottom-layer MPC with a top-layer real-time feedback controller to enforce state invariance and attractivity between sampling intervals.
• TPC enhances system performance by reducing quadratic control effort and ensuring strict constraint adherence, validated in large-scale energy systems and adaptable to data-driven variants.

Transient Predictive Control (TPC) is a receding-horizon optimal control paradigm designed to guarantee strong transient performance and constraint enforcement in dynamical systems—especially those with safety-critical, time-sensitive requirements—by integrating model predictive control (MPC) principles with transient-invariance and attractivity guarantees. While commonly associated with real-time frequency and security envelopes in large-scale power networks, TPC has theoretical and computational generalizations across robust MPC, adaptive MPC, and data-driven predictive controllers.

1. Mathematical Structure and Core Principles

The central feature of Transient Predictive Control is the explicit embedding of transient or “safety” constraints for critical states (e.g., frequency in power systems) into the receding-horizon MPC optimization, often combined with additional architectural layers for real-time invariance enforcement.

A canonical TPC architecture comprises two layers (Zhang et al., 2018):

• Bottom-layer MPC: Periodically (every $T_s$ seconds) solves a convex quadratic program over a finite horizon $N$, with decision variables $\hat u$ (future control actions) and slack variables (for soft constraint relaxation). The predicted dynamics reconstruct the linearized (or lifted) system evolution, constraints enforce soft (or hard) bounds on target states (e.g., $|\hat \omega_i| \leq \omega^{\max}_i + \beta$), and the objective penalizes control effort and violations.
• Top-layer enforcement: Direct-feedback, typically real-time continuous controllers designed to strictly guarantee state invariance or attractivity to the prescribed transient envelope (e.g., $$\underline\omega_i \le \omega_i(t) \le \bar\omega_i$$), even between MPC sampling events. Analytical constructions ensure feedback is locally Lipschitz in the state, vanishes in the safe region, and “kicks in” only near constraint boundaries.

The structure is formalized for a linearized swing system:

$$\dot f(t) = Y_b D \omega(t) \ M \dot\omega(t) = -E \omega(t) - D^\top f(t) + p(t) + \alpha(t)$$

with receding-horizon constraints on predicted trajectories and stability filters on control action (Zhang et al., 2018, Zhang et al., 2019).

2. Transient Constraint Formulations

A defining characteristic of TPC is the use of transient constraints that enforce not only steady-state performance but also bounded state excursions during disturbances or transients. These constraints can be framed as hard invariance:

$$\underline\omega_i \le \omega_i(t) \le \bar\omega_i \quad \forall t$$

or, in the optimization, as soft bounds:

$$\underline\omega_i - \beta \le \hat\omega_i(k+1) \le \bar\omega_i + \beta$$

with quadratic penalties on $\beta$, ensuring that the violation cost is explicitly minimized when invariant satisfaction is not initially feasible due to system initialization or exogenous shocks.

The real-time layer may encode invariance via direct-feedback decoupled from the MPC prediction, using locally Lipschitz control laws and switching gains when the state approaches the constraint boundary (Zhang et al., 2018, Zhang et al., 2019):

$$\alpha_{DF,i} = \begin{cases} \min\{0, \gamma_i (\bar\omega_i - \omega_i)/( \omega_i - \bar\omega_i^{thr} + v_i ) \} & \omega_i > \bar\omega_i^{thr} \ 0 & \underline\omega_i^{thr} \le \omega_i \le \bar\omega_i^{thr} \ \max\{0, \gamma_i (\underline\omega_i - \omega_i)/( \underline\omega_i^{thr} - \omega_i + v_i ) \} & \omega_i < \underline\omega_i^{thr} \end{cases}$$

3. Stability, Invariance, and Performance Guarantees

TPC methodologies provide rigorous guarantees simultaneously on trajectory invariance, asymptotic stability, and transient suboptimality. Key analytical tools include:

• Composite Lyapunov functions: These demonstrate non-increase of a total energy-like function (system state plus filter and controller memory) under the closed-loop TPC law, implying global or regional asymptotic stability:

$$V(f, \omega, \alpha_{MPC}) = \frac{1}{2}\|f-f_{\infty}\|^2 + \frac{1}{2} \omega^T M \omega + \frac{1}{2} \|\alpha_{MPC}\|^2$$

with $$\dot V \le - \omega^T E \omega - \sum_{i} (1/T_i - \epsilon_i) \alpha_{MPC,i}^2 \le 0$$ given appropriate parameter choices (Zhang et al., 2018, Zhang et al., 2019).

• Nagumo-type conditions: Real-time feedback layers are constructed so that for each protected state, the closed-loop derivative points strictly into the safe set at the boundary, guaranteeing invariance, while attractivity is ensured via vanishing corrective control (Zhang et al., 2018).
• Distributed/Regional Feasibility: Partitioned implementations preserve the above guarantees by treating couplings at boundaries as exogenous, resulting in provable invariance and stability for each region with only neighbor-to-neighbor coordination (Zhang et al., 2018, Zhang et al., 2019).

4. Distributed and Computational Architectures

TPC is designed for large-scale and geographically partitioned networks. Regionalization is achieved by partitioning the network into subgraphs, with each subgraph (region) independently solving a local MPC using only local system states, forecast injections, and boundary flow estimates (Zhang et al., 2018, Zhang et al., 2019):

• Each region solves an MPC with restricted dynamics and modified forecast injection to encode cross-region flows as exogenous signals.
• The same two-layer (predictive + real-time feedback) structure is instantiated at each regional controller.
• All proofs of invariance and attractivity generalize to the regional case with no cross-region iteration required except for boundary data.

This enables full scalability: distributed TPC scales with region size rather than network size, and sparsity in network structure is exploited.

5. Performance Metrics and Trade-offs

TPC performance is characterized both by metric suboptimality (such as integrated quadratic control effort and maximum deviation) and by qualitative guarantees (strict invariance, attractivity rates, and convergence).

In benchmark simulations on IEEE 39-bus networks (Zhang et al., 2018, Zhang et al., 2019):

• TPC ensures all protected frequencies remain within $[59.8, 60.2]$ Hz under 20% step/sinusoidal disturbances.
• When compared to open-loop or direct-feedback-only controls, TPC reduces the total quadratic control effort by up to 4x.
• With infeasible initializations (e.g., $\omega_{30}=59.5$ Hz), TPC restores the frequency to the safe interval within short timeframes.
• Increased MPC horizon $N$ and decreased sampling period improve performance (tighter constraint satisfaction and less control effort), at the expense of increased computation time.

A summary of simulation outcomes:

Configuration	Max Excursion	Total Control Effort	Safety Violation
Open Loop	Violates	N/A	Significant
TPC (two-layer)	None	≈98	None
Direct-Feedback Only	Violates	≈364	Frequent
Regional TPC	None	Comparable/Better	None

6. Connections to Generalized and Data-Driven MPC

While the original TPC instantiations target frequency transients in power networks, the methodology generalizes broadly:

• Adaptive TPC: Extensions involving online least-squares parameter adaptation coupled with certainty-equivalent MPC achieve explicit finite-horizon suboptimality bounds, decomposing tracking error into $O(T)$ disturbance-driven and $O(\sqrt{T})$ initial-model-error-driven components (Degner et al., 2024, Degner et al., 2024). TPC thus directly quantifies and controls transient tracking-to-nominal/turnpike trajectories under model uncertainty.
• Data-driven TPC: In inverter control and power oscillation damping, TPC has been realized in model-free settings using ARX predictors or direct Hankel-based multi-step input-output models, retaining causal prediction and shifting computational cost offline (Mastroddi et al., 27 Jan 2026, Graf et al., 3 Jul 2025). Online solve times are reduced to millisecond or sub-millisecond scales, compatible with embedded hardware.
• Nonlinear/Robust TPC: In applications including grid-forming converter stabilization under post-fault current saturation and tube-based robust economic MPC, the TPC design paradigm embeds transient, tube, or invariance constraints in the prediction problem, with rigorous bounds on closed-loop cost and constraint satisfaction over finite transients (Klöppelt et al., 2021, Arjomandi-Nezhad et al., 2023).

7. Practical Implementation and Tuning Guidelines

Key operational choices in TPC involve tuning the MPC horizon $N$, sampling $T_s$, penalty parameters $(c_i,d)$, and filter time constants $(\epsilon_i,T_i)$. The product $\epsilon_i T_i < 1$ is fundamental for stability. Soft constraint and slack penalty parameters calibrate the aggressiveness of constraint enforcement.

For large-scale deployments, regionalization and operator-splitting/ADMM solvers can be adopted to ensure tractable solve times even under tight real-time constraints (Zhang et al., 2018, Mastroddi et al., 27 Jan 2026).

Trade-offs include computation vs. control performance, tightness of invariant envelope vs. actuation limits, and robustness to model mismatch (addressed in adaptive and data-driven variants).

---

Transient Predictive Control thus represents a unifying paradigm for receding-horizon control with strong transient and invariance guarantees, validated across large-scale energy systems, embedded inverters, and general constrained linear/nonlinear systems (Zhang et al., 2018, Zhang et al., 2019, Degner et al., 2024, Mastroddi et al., 27 Jan 2026).