Introduction
In the transition to electric vehicles (EVs), energy management extends far beyond the powertrain. The heating, ventilation, and air conditioning (HVAC) system represents the most significant auxiliary load, capable of reducing a vehicle's driving range by up to 40% in extreme climates. Optimizing these systems is critical for extending range, yet the underlying physics—specifically the refrigerant dynamics of the vapor compression cycle—present a formidable control challenge. These dynamics are highly nonlinear and operate on much faster timescales than the slow thermal response of the vehicle cabin.
The Control Challenge
Traditional control methods often struggle to balance energy efficiency with passenger comfort. While Model Predictive Control (MPC) offers a framework for multi-objective optimization under constraints, its implementation in automotive hardware is hindered by computational costs. To account for slow cabin temperature changes, a standard MPC requires a long prediction horizon, which, when combined with nonlinear refrigerant models, leads to optimization problems too complex for real-time execution on standard Electronic Control Units (ECUs).
System Architecture and Modeling
The researchers focused on a standard vapor compression cycle integrated with a battery thermal management system (TMS). The system includes a compressor, condenser, evaporator, and expansion valves (TXVs), as illustrated in Figure 1.
The control objectives are twofold: maintaining the evaporator pressure ($p_e$) at a setpoint to ensure cabin comfort and minimizing the total energy consumption of the compressor and radiator fan.
Methodology: 1-Step MPC via Value Function Approximation
To overcome the computational bottleneck, the authors proposed a novel strategy: reducing the MPC horizon to a single step ($N=1$) by incorporating a "cost-to-go" term, or value function, as a terminal cost.
The process began offline, where Dynamic Programming (DP) was used to solve the global optimization problem across various driving cycles. By analyzing the resulting value function, the authors observed that it could be accurately approximated by a quadratic function centered around the target operating points.
As shown in Figure 2, the fitted quadratic curves closely match the actual cost-to-go data points derived from DP. By using this mathematical approximation as the terminal cost in a 1-step MPC, the controller effectively "sees" the long-term energy consequences of its immediate actions without needing to simulate the entire future horizon.
Performance and Real-Time Viability
The proposed 1-step MPC was evaluated using the EPA's SC03 driving cycle, which simulates real-world driving with air conditioning active. The results demonstrated exceptional efficiency and speed:
- Computational Efficiency: The controller solves in just 5-6 milliseconds on a standard CPU, making it easily compatible with the 20ms cycle times typical of automotive rapid-prototyping hardware.
- Energy Savings: In the SC03 Full cycle, the MPC achieved a 23.7% energy reduction compared to a baseline production controller. Interestingly, it outperformed the DP benchmark in energy savings (911 kJ vs 1386 kJ) by strategically allowing a slightly higher temperature tracking error (2.0% vs 1.08%), a trade-off that remains imperceptible to passengers.
- Robustness: The quadratic terminal cost acts as a natural stabilizer, ensuring the system remains robust against external disturbances such as varying vehicle speeds or ambient temperatures.
Conclusion and Future Directions
This research provides a viable path for implementing advanced optimal control in production vehicles. By leveraging the "cost-to-go" from offline optimization, the 1-step MPC achieves near-optimal performance with minimal computational overhead. Future work will focus on developing adaptive laws to automatically tune control weights for varying environmental conditions and conducting in-vehicle experimental trials to validate these findings in real-world scenarios.
Summary updated