Chance-constrained model predictive control with online gaussian process learning reduces battery degradation in vehicle-to-grid service

Vehicle-to-grid (V2G) duty cycles are noisy and high-rate, and the capacity fade they cause is poorly described by the empirical, fixed-parameter degradation laws built into today’s model predictive control (MPC) battery managers. The controller sees a single number where it should see a distribution, and so it cannot trade off grid service against aging in an uncertainty-aware way. This paper develops an MPC scheme in which the degradation cost comes from a Gaussian process (GP) regressor trained online, and the predictive variance of that regressor is used to tighten the state-of-charge (SOC) and temperature constraints through a Gaussian chance-constraint back-off. A temporal convolutional network acts as a digital twin, filtering raw sensor streams at 100 Hz and delivering per-cycle capacity-fade samples that keep the GP training window current without offline retraining. The GP posterior is evaluated on an FPGA and the MPC quadratic program on the CPU, which together meet a 10 ms control period on a Xilinx Versal ACAP. We tested the method on five independent 20 kWh NMC packs per scenario, driven by PJM regulation-D signals for 500 equivalent cycles. Under the aggressive scenario the proposed controller retained 82.7 ± 0.4% of nominal capacity, against 78.4 ± 0.7% for a deterministic Arrhenius-based MPC (paired t-test, p = 0.003); tracking RMSE rose by under 5%. Ablations showed that the chance-constrained tightening was the dominant contributor to the safety margin, and the online-updated twin to the capacity-retention gain. Sensitivity sweeps indicated that the controller is stable with respect to sensor-noise scaling, to the MPC horizon within 20–40 steps, and to the sliding-window size within 30–80 cycles.