Adaptive Current-Limiting Strategy

• The paper presents an adaptive current-limiting strategy that dynamically adjusts controller setpoints via online QP optimization to enforce operational current constraints.
• It integrates adaptive cost weights, conditional QP engagement, and time-varying limit updates to maintain performance while preventing hardware saturation.
• Experimental validation on systems like TCV demonstrates rapid response (<10 ms) and minimal shape errors, ensuring reliable operation under dynamic conditions.

An adaptive current-limiting strategy is a control methodology that dynamically enforces operational current constraints on driven physical systems—such as power electronic converters, superconducting tokamak coils, or resistive memory measurement circuits—by online adjustment of controller setpoints or reference targets in response to bounded, real-time estimations of constraint violation risk. This form of constraint-aware control preserves primary performance objectives (e.g., shape control, voltage regulation, waveform tracking) while ensuring protection of hardware from saturation, loss of synchronism, or destructive overcurrent. Adaptive mechanisms are essential in environments where system states, operational margins, or disturbance magnitudes are time-varying and difficult to forecast with static or open-loop thresholds. Notable implementations include quadratic-program (QP) allocators for plasma coil arrays, estimator-driven reference rescaling, and layered feedback structures that inject constraint information into primary control loops.

1. Mathematical Formulation of Adaptive Current Limiting

An adaptive current-limiting architecture is mathematically defined by casting the constraint-enforcement problem as a real-time optimization task. In the context of plasma shape control in TCV (Tokamak à Configuration Variable), the relevant variables are:

• $I_{EF} \in ℝ^m$: Vector of shape-coil currents (e.g., E1–E8, F1–F8, $m=16$).
• $u_{CLA} \in ℝ^m$: Corrective current allocator outputs added to the hybrid controller references.
• $$I^0_{EF} = I_{EF,meas} + P^0_{EF}(u_{sh} + r_h - y_h)$$: Predicted steady-state currents post-actuation.
• Static gain matrices $P^0_{EF}$, $P^0_{sh}$: Map between reference errors and coil currents/outputs, respectively.

At every control interval, the current allocator solves a quadratic program (QP):

$$\min_{u_{CLA}} \tfrac{1}{2} u_{CLA}^T H\, u_{CLA} + f^T u_{CLA}$$

subject to: $G u_{CLA} \leq h$

Here, $H$ encodes penalties on shape deviation and coil effort (weighted by matrices $W$, $m=16$0), and constraints $m=16$1 encode current bounds and any additional operational requirements. The bounds $m=16$2 can be made adaptive and time-varying, e.g.,

$m=16$3

where $m=16$4 is a dynamically adjusted margin that enforces a proximity “warning” band as a given coil nears its hard limit (Frattolillo et al., 31 Jan 2025).

Such formulations generalize across domains where steady-state (or predicted) system outputs can be mapped linearly from control actions, and constraint compliance must be maintained at all times.

2. Mechanisms for Real-Time Adaptivity

The adaptivity of the limiting algorithm is encoded in three mechanisms:

a) Adaptive Cost Weights and Soft Limits:

The QP penalty matrix $m=16$5 can be made a nonlinear function of predicted current margin:

$m=16$6

This enforces greater penalty for corrective action in directions that may breach critical currents, thereby adaptively reshaping the control’s feasible action set as operation nears hard bounds.

b) Sequential/Conditional Engagement:

The QP allocator remains dormant ($m=16$7) while all predicted currents are safely inside their margins. As a bound is approached, the QP solution seamlessly "switches on," reallocating actuation effort to prevent violations—requiring only the QP’s built-in constraint-activation logic, with no extrinsic mode switches or hard discontinuities (Frattolillo et al., 31 Jan 2025).

c) Time-Varying Constraints and Online Margin Functions:

The bounds themselves (e.g., $m=16$8) may be updated as a function of real-time diagnostics, operational objectives, or reliability policies (e.g., gradual ramp-down of permissible current every 0.2 s to model superconducting coil safety procedures).

3. Implementation and Real-Time Performance

Realization of adaptive current-limiting policies imposes strict requirements on computational efficiency and system observability:

• Sampling and Solution Rate: CLA on TCV is executed at 1 kHz, matching the plasma shape control loop.
• Solver Selection: The use of the Goldfarb–Idnani active-set QP solver (quadprog++ in C++) ensures deterministic solve times (typical problem: 16 variables, 32 constraints, <3 iterations, <200 μs maximum) well within real-time constraints.
• Measurement Cycle: Each frame entails acquisition of coil currents ($m=16$9), shape and position ($u_{CLA} \in ℝ^m$0), computation of predicted currents ($u_{CLA} \in ℝ^m$1), adaptation of active bounds, QP solution for $u_{CLA} \in ℝ^m$2, and distribution of corrections to shape-control and hybrid references.
• Continuity and Stability: The continuous, piecewise-linear response of $u_{CLA} \in ℝ^m$3 ensures Lipschitz continuity (absence of chattering), with overall closed-loop stability inheriting from the outer loop design—guaranteed formally via Lyapunov arguments linked to the continuous reference modifications (Frattolillo et al., 31 Jan 2025).

4. Experimental Validation and Comparative Performance

Empirical assessment of the adaptive QP-based current-limiting scheme in TCV covered several scenarios:

• Single-Circuit Limit Reduction: Upon lowering the E5 upper bound during plasma operation, CLA became active within 10 ms, returning the current within limit, with shape error RMS increase <0.5 mm.
• Simultaneous Multi-Circuit Restrictions: Concurrent reduction on E5, F3, and F7 saw the CLA reallocate current across circuits, preventing saturation at the expense of only mild, localized shape error (<1 mm) and no oscillation.
• Dynamic Time-Varying Limit: E5 upper bound stepped down by 250 A every 0.2 s. CLA maintained coil compliance at each step, with only minor, transient shape error increases.

Compared to a static limiting policy (e.g., brute-force voltage clamping), the QP-based approach provided approximately 50% higher utilization of the current margin with shape deviations in the sub-millimeter regime and zero disruptions or violation events.

5. Scalability, Robustness, and Implications for Large-Scale Systems

The QP-based current limit avoidance paradigm is directly extensible to large-scale superconducting coil arrays such as those in ITER. With $u_{CLA} \in ℝ^m$4, the resulting optimization problem remains of dimension amenable to 100–200 μs solves on commodity CPUs or FPGAs, operating comfortably at ≥100 Hz. The method requires only steady-state (or sampled) gain matrices—not higher-order model-predictive constructs—so the computational burden grows polynomially (as $u_{CLA} \in ℝ^m$5), not exponentially.

Robustness is ensured because the current allocator acts upon predicted steady-state gains, with dynamical model uncertainties managed by the principal hybrid and shape-control loops. The global piecewise-linear feedback law maintains closed-loop continuity, precluding control-induced instabilities or chattering as coil limits are approached or receded.

6. Generalization and Broader Impact

Adaptive current-limiting via online QP allocation is a generalizable strategy applicable in any multi-actuator system where hard current/effort limits must be enforced in real time, without sacrificing primary control objectives. The approach integrates naturally as an "envelope-protection" layer wrapping existing high-performance controllers and requires only measured (or estimated) system states and accurate, linear steady-state gain mapping.

The deployment in TCV supports the conclusion that a light-weight, single-step quadratic optimizer embedded in the fast control loop suffices for automatic, adaptive current-limit avoidance, achieving near-maximal performance utilization with robust safety compliance. For next-generation fusion reactors or critical infrastructure systems relying on superconducting or power-electronic actuators, these mechanisms provide a foundation for scalable, formally verifiable, and minimally invasive current-limiting solutions (Frattolillo et al., 31 Jan 2025).