TL;DR
Automotive air conditioning heavily drains electric vehicle batteries, but optimal control algorithms are usually too computationally heavy for real-time use. This paper introduces a Model Predictive Control (MPC) strategy that uses a "cost-to-go" approximation derived from Dynamic Programming to reduce the controller's prediction horizon to a single step. This achieves near-optimal energy savings with a computation time of just 5-6 milliseconds, making it highly viable for real-time automotive hardware.
What Problem Does This Paper Address?
In electric and hybrid vehicles (BEVs/HEVs), the air conditioning (A/C) system is the single largest auxiliary energy consumer. In extreme weather, heating and cooling loads compete directly with the powertrain for battery power, capable of slashing vehicle range by up to 40%.
Engineers can use advanced algorithms like Model Predictive Control (MPC) to optimize A/C energy use while maintaining cabin comfort. However, they face a major mathematical roadblock: the refrigerant dynamics in an A/C system are fast and highly nonlinear, while cabin temperatures change slowly. To optimize this using standard MPC requires a long "prediction horizon" (looking far into the future to make current decisions). This demands too much computing power for standard automotive control units to handle in real-time, forcing engineers to either simplify the physics too much or abandon MPC altogether.
Methods
The authors start with a lumped-parameter nonlinear model of the vehicle's vapor compression cycle, focusing on two main states: evaporator pressure and condenser pressure. They first solve the energy optimization problem offline using Dynamic Programming (DP) to find the absolute global optimal control policy for the compressor and radiator fan.
To make this usable in real-time, they employ a clever trick based on Bellman’s Principle of Optimality. Instead of running a computationally heavy MPC with a long horizon, they extract the "cost-to-go" (also known as the value function) from the offline DP solution. They found this cost-to-go can be accurately approximated as a quadratic mathematical function.
By plugging this approximated quadratic function into the MPC as a terminal cost, they can shrink the MPC prediction horizon down to a single step ($N=1$). This drastically reduces computational complexity while retaining the complex, nonlinear dynamics of the physical A/C system.
Datasets
The authors evaluate their controller using standard regulatory driving cycles, specifically variations of the SC03 Supplemental Federal Test Procedure (SC03 Mid, SC03 Low, and SC03 Full). The SC03 cycle is specifically designed by the EPA to represent driving with the air conditioning turned on. No new datasets were introduced.
Key Insights and Ideas
- The "Cost-to-Go" is Predictable: The authors demonstrated that the value function from the DP solution forms a neat, predictable quadratic shape around the target reference pressure. This smooth shape is what allows a simple equation to replace a long, complex prediction horizon.
- One Step is Enough: By accurately capturing the future cost of current actions via the terminal cost approximation, a 1-step MPC can perform almost as well as a controller looking minutes into the future.
- Inherent Robustness: Because the quadratic terminal cost heavily penalizes deviating from the target evaporator temperature, the controller naturally resists external disturbances. It maintains stability despite sudden changes in vehicle speed, engine speed, or outside air temperature.
Key Results
- Massive Speedup: The 1-step MPC framework solves in just 5-6 milliseconds on a standard desktop CPU. This allowed the authors to deploy it on Rapid-Prototyping hardware with a 20ms cycle time—well within the strict real-time requirements of automotive control units.
- Significant Energy Savings: On the SC03 Full cycle, the baseline production controller used 1646 kJ of energy. The DP (global optimal) used 1386 kJ (15.8% savings). The proposed MPC used 911 kJ (23.7% savings).
- The Trade-off: The MPC actually saved more energy than the DP benchmark by allowing a slightly higher tracking error (deviating slightly more from the exact target temperature). The DP had a tracking error of 1.08%, while the MPC had an error of 2.0%—a negligible difference for human comfort, but a massive win for battery range.
Limitations and Future Work
- Limitations: The value function approximation was derived from a single specific drive cycle (SC03 Mid). While it generalizes well, the weighting parameter ($\alpha'$) that balances energy savings against temperature tracking currently requires manual tuning to adapt to different conditions. Furthermore, by approximating the value function, the MPC loses the strict mathematical guarantee of "recursive feasibility" (the guarantee that a solution will always exist) inherent to pure DP.
- Future Work: The authors plan to design an adaptation law to automatically tune the $\alpha'$ weight online based on driving conditions. They also plan to move from hardware-in-the-loop simulations to actual in-vehicle experimental verification to prove the system's real-world viability.
Summary updated