Predictive Torque and Drive Control — MPC at the Microsecond Timescale

Why This Matters (Executive Summary)

• Electric drives and power converters are critical assets in oil & gas compression, industrial turbomachinery, and precision manufacturing. Unplanned trips cost hours to days of downtime.
• Model Predictive Control at the drive level optimizes switching sequences in real time (10—100 microsecond cycles), delivering measurable improvements in torque quality, efficiency, and fault tolerance.
• Reference implementations report: partial torque ride-through during grid voltage dips on 42 MW compressor trains, up to 50% reduction in switching losses, 27% torque ripple reduction in reluctance motors, and servo bandwidth increases from 147 Hz to 208 Hz.
• The pattern applies to any system with electric motors, inverters, or contactless bearings where torque quality, efficiency, or disturbance rejection directly affects product quality or uptime.

The Design Pattern Explained

Drive-level MPC operates at the fastest timescale in the MPC family. Instead of computing smooth continuous setpoints, the controller selects from a finite set of inverter switching states (Finite-Set MPC) or solves a fast QP to determine optimal voltage vectors — every 10 to 100 microseconds.

The key difference from process-level MPC is the discrete nature of the control action: a three-phase inverter has a limited number of valid switching combinations. Finite-Set MPC enumerates these states and evaluates a cost function (torque error, flux error, switching losses) to pick the best one at each step. For longer horizons, tree-search or sphere-decoding methods keep computation tractable.

The architecture typically consists of: current/flux estimation from electrical measurements, a predictive model of motor electromagnetics, a cost-function optimizer running on dedicated hardware (DSP, FPGA, or industrial controller), and protection logic that coordinates with process-level systems.

Three distinct optimization approaches in drive-level MPC: Drive-level predictive control spans three fundamentally different formulations — and it is important not to conflate them. Finite-Set MPC (FS-MPC / MPTC / MPDTC): an inverter has a discrete, finite set of valid switching states (typically 8 for a two-level three-phase inverter). The optimizer enumerates these states over the prediction horizon and selects the best sequence by evaluating a cost function — no QP or NLP is solved. This makes FS-MPC extremely fast and naturally suited to microsecond timescales, but the discrete action space limits the achievable torque ripple compared to continuous-input controllers. Linear MPC (QP): for applications with continuous voltage control (active magnetic bearing position, PMSM with continuous modulation), the linearized electromagnetic or mechanical dynamics yield a convex Quadratic Program — fast, globally optimal, and directly deployable on DSP hardware. Nonlinear MPC (NMPC): when dynamics are inherently nonlinear — open-loop unstable magnetic levitation, where the force-gap relationship is strongly nonlinear — the full nonlinear model is used in a non-convex NLP. Solver performance and initialization are critical; results outperform PID but at significantly higher computational cost. Applications below are labelled by type.

Applications & Reference Implementations

Application 1: MPTC on ABB MEGADRIVE-LCI — Gas Compressor Ride-Through and Heavy Industrial Drives

ABB’s MEGADRIVE-LCI is a commercially available medium-voltage drive covering synchronous motor loads from 2 to 150 MW, with Model Predictive Torque Control (MPTC) as a core feature. MPTC optimizes the switching sequence of the load-commutated inverter every 1 ms on an ABB AC 800PEC controller, delivering converter efficiency exceeding 99% and handling input-voltage disturbances of +20% / −50% with derating. The critical oil & gas application is partial-torque ride-through during grid voltage dips: at Statoil’s Kollsnes (2 × 42.2 MW) and Karstoe (3 × 7.5 MW) gas compressor facilities, conventional zero-torque ride-through risked compressor surge and restart downtime of hours to days. With MPTC, the controller maximizes feasible partial torque while enforcing current limits — in a documented on-site voltage dip event (0.3 pu for approximately 80 ms), MPTC recovered approximately 0.23 pu rated torque, successfully protecting compressor operation. This single application class — high-power synchronous motor drives with voltage-dip ride-through — demonstrates predictive control delivering measurable uptime value at the MW scale. ¹²³⁴⁵

Application 2: Enhanced FS-MPC for Switched Reluctance Motors — Torque Ripple Reduction

Switched Reluctance Motors (SRMs) are attractive due to their magnet-free design but suffer from significant torque ripple, especially at low speeds. An enhanced Finite-Set MPC approach with a single cost function (no weighting-factor tuning required) was experimentally evaluated and reported 27.14% torque ripple reduction at 600 rpm compared to conventional techniques. The torque ripple standard deviation was measured at 1.074 Nm. This demonstrates that predictive control can address one of the main barriers to SRM adoption in industrial applications.⁶

Application 3: Long-Horizon MPDTC for PMSM — Switching Loss and THD Optimization

Model Predictive Direct Torque Control (MPDTC) was evaluated for a Permanent Magnet Synchronous Motor with a three-level voltage source inverter, using prediction horizons up to 150 time steps. The simulation study reported switching losses and switching frequency reduced by up to 50%, and torque Total Harmonic Distortion (THD) reduced by 25%, while current THD remained unchanged. Long-horizon predictive control enables the optimizer to plan switching sequences that minimize losses over extended windows rather than being greedy at each step.⁷

Application 4: Online MPC Torque Control on Embedded DSP — PMSM Drives

An online MPC torque controller for a PMSM was formulated as a Quadratic Program and solved in real time on a low-cost Digital Signal Processor, demonstrating that constraint-aware predictive torque control does not require expensive computing hardware. The Processor-In-the-Loop evaluation confirmed feasibility in both compute and memory terms, with improved torque dynamics versus a conventional controller. This is significant because it shows MPC can be deployed on the same class of hardware already present in standard industrial drives.⁸

Application 5: Offset-Free MPC for Active Magnetic Bearing — Turbomachinery

An offset-free MPC formulation was applied to a cone-shaped Active Magnetic Bearing (AMB) system on a turbocompressor test bench. The centralized position controller runs at 4 kHz with ADC conversion triggered by a 20 kHz PWM carrier. Active magnetic bearings levitate rotating shafts without contact, and predictive control enables vibration suppression and offset-free position tracking under disturbances — critical for high-speed turbomachinery where bearing failure means immediate shutdown.⁹

Application 6: NMPC for Magnetic Levitation — Experimental Comparison vs. PID

A real-time NMPC approach using an evolving linearized model inside the optimization loop was experimentally demonstrated on a fast nonlinear magnetic levitation plant. Compared against a PID baseline, NMPC reported better control performance at the cost of higher computational complexity. This laboratory result illustrates the fundamental advantage of predictive control for nonlinear, open-loop unstable systems where PID tuning is inherently limited.¹⁰

What This Means for Your Operations

• Any system with high-value rotating equipment (compressors, turbines, precision spindles) can benefit from predictive control at the drive level.
• Fault ride-through is especially valuable where unplanned trips have high consequential cost — gas compression, continuous chemical processes, semiconductor manufacturing.
• Torque quality improvements directly reduce vibration, noise, and mechanical wear, extending equipment lifetime.
• Embedded feasibility means MPC can often be deployed on existing drive hardware or with minimal compute upgrades.

How We Deliver This (Engagement Model)

• Phase 0: NDA + data request — characterize drive topology, motor parameters, grid conditions, and protection logic.
• Phase 1: Fixed-scope discovery — simulation-based feasibility, control architecture concept, and expected KPI improvements.
• Phase 2: Implementation + validation + commissioning — controller design, Hardware/Processor-In-the-Loop testing, field commissioning.
• Phase 3: Monitoring + training + scaling — integration with process-level protection, operator training, and fleet rollout.

Typical KPIs to Track

• Torque quality: torque ripple (Nm, %), THD, vibration levels
• Efficiency: switching losses (W), converter efficiency (%)
• Fault tolerance: ride-through success rate, avoided trip count, mean time to recover
• Availability: unplanned downtime hours per year, restart frequency

Risks & Prerequisites

• Real-time compute budget: drive-level MPC requires sub-millisecond solve times. Explicit MPC (pre-computed lookup tables) or FPGA-based solvers may be needed for the fastest applications.
• Model accuracy at speed: motor parameters (inductance, flux linkage) must be characterized accurately. Parameter estimation or adaptive approaches may be required.
• Protection coordination: MPC ride-through must be coordinated with existing protection relays and process-level safety systems. Timer-based protection may trip before MPC can act if not reconfigured.
• Regulatory and safety certification: drive-level changes in safety-critical applications (oil & gas, nuclear) require functional safety assessment and may need SIL-rated implementations.

FAQ

Q: Can drive-level MPC be retrofitted to existing motor drives? In many cases, yes. If the drive controller has sufficient compute headroom (DSP or dedicated FPGA), MPC can be implemented as a firmware update. For older drives, an external compute module may be needed.

Q: How does Finite-Set MPC compare to Field-Oriented Control (FOC)? FOC uses a modulator to generate switching patterns from continuous voltage references. Finite-Set MPC eliminates the modulator and directly selects switching states, enabling tighter constraint handling and often lower switching losses, but requires more computational effort per control cycle.

Q: Is this relevant for smaller motors (below 1 MW)? Absolutely. The PMSM and SRM examples on this page are in the sub-100 kW range. The principles apply across the power spectrum — from precision servo motors to 150 MW compressor drives.

Book a 30-Minute Discovery Call

Ready to explore whether this pattern fits your system?

Dr. Rafal Noga — Independent APC/MPC Consultant

Email me | noga.es

Fixed-scope discovery — NDA-first — DACH on-site available

Public References

Footnotes

1. ABB, “ABB’s Model Predictive Torque Control (MPTC) protects against gas supply interruptions” (ABB Highlights, 2016). Link ↩

2. ABB, “Partial Torque Ride Through with Model Predictive Control” (PCIC Europe, 2016). PDF ↩

3. ABB, “Electric driven gas compressor control (MPTC + DT2S)” (ABB Review, 2017). PDF ↩

4. ABB, “MEGADRIVE-LCI product page” (ABB). Link ↩

5. ABB, “MEGADRIVE-LCI technical data” (ABB). Link ↩

6. Deepak et al., “Experimental analysis of enhanced FS-MPC and DTC in SRM drives” (Scientific Reports, 2024). Link ↩

7. Geyer et al., “Model Predictive Direct Torque Control of PMSM” (ECCE, 2010). PDF ↩

8. Cimini et al., “Online Model Predictive Torque Control for PMSM” (ICIT, 2015). PDF ↩

9. Castellanos et al., “Offset-Free MPC for a Cone-shaped Active Magnetic Bearing System” (Politecnico di Torino, 2021). PDF ↩

10. Novotny, “Magnetic levitation: real-time NMPC demonstrated experimentally vs PID baseline” (Springer, 2025). Link ↩