A novel differentiable predictive control (DPC) approach for safe and optimal EV charging and discharging scheduling

Differentiable Predictive Control (DPC) has emerged as a data-driven alternative to Model Predictive Control (MPC), attracting growing interest across various research domains. Unlike traditional MPC, which relies on physics-based models, DPC utilizes black-box models—typically neural networks—to learn control policies. The primary advantage of DPC is its computational efficiency. However, a key challenge lies in ensuring constraint satisfaction, as neural network-based policies often struggle with maintaining feasibility. In conventional DPC frameworks for building-related studies, both equality and inequality constraints are enforced through penalty methods. However, tuning penalty parameters is complex, and constraint violations often persist. To address this issue, we propose a novel DPC approach that strictly enforces constraints. Specifically, we incorporate a projection layer based on Karush-Kuhn-Tucker (KKT) conditions to satisfy equality constraints. For inequality constraints, we replace the traditional penalty method with barrier functions, which provide steeper gradient as the solution approaches constraint boundaries, enforcing stricter constraint compliance. To evaluate the effectiveness of our approach, we apply it to an electric vehicle (EV) charging and discharging optimization problem, considering three objectives: minimizing electricity costs, reducing CO2 emissions, and load flattening. We compare our method against conventional penalty-based DPC and the augmented Lagrangian method. The results indicated that the proposed approach reduced constraint violations by over 90 % in occurrence, compared to the conventional penalty method. Additionally, objective performance improved, achieving a 3 % greater bill reduction, 12 % more CO2 reduction, and 29 % greater peak load reduction.