Self-Learning Controller Keeps EV Batteries Healthier During Grid Service

Batteries that sell electricity back to the grid earn money for their owners, but the same service quietly eats away at them. Vehicle-to-grid operation asks a pack to swing rapidly between charging and discharging, often at high power, and those direction-reversing cycles accelerate the two processes that permanently damage lithium-ion cells: growth of the solid-electrolyte interphase and lithium plating. A new study reports that a controller which learns how fast its own battery is degrading, and knows when its own predictions are uncertain, can keep significantly more of the battery’s capacity alive after hundreds of punishing grid-regulation cycles than today’s standard approaches.

The work, published in the Springer journal Discover Vehicles, comes from Kun Cai, Qingliang Ma and Chaochun Zhong. Their starting point is a familiar weakness of commercial battery management. Most model predictive control schemes, which plan battery power over a short future horizon, describe aging with a fixed, closed-form degradation law such as an Arrhenius throughput model fitted offline in the laboratory. Such laws handle steady, well-behaved cycling reasonably well, but they underestimate fade under the irregular, partial depth-of-discharge duty cycles that dominate vehicle-to-grid service. Worse, the controller sees a single predicted degradation number with no sense of how much it should trust that number, so it cannot trade grid service against battery health in an uncertainty-aware way.

The researchers’ answer replaces the fixed aging law with a Gaussian process regressor trained online while the battery operates. A Gaussian process is a nonparametric statistical model whose predictions come with a calibrated variance: where training data is dense the model is confident, and where it is sparse the variance grows. That posterior variance is the key ingredient. At every control step, the controller tightens its state-of-charge and temperature limits by an amount proportional to the model’s uncertainty, using a Gaussian chance-constraint back-off with a 5 percent risk level. When the controller is operating in a region of the state space where its learned degradation model is essentially guessing, the admissible operating window shrinks automatically; when the model is confident, the window relaxes.

Feeding the Gaussian process is a digital twin built as a temporal convolutional network. The twin consumes raw 100-hertz streams of terminal voltage, current and surface temperature, filters them in a fast stage running on one-second windows, and produces a refined capacity estimate once per minute. At the end of each equivalent cycle, the difference between the starting and ending capacity estimates becomes a training label for the degradation model. This solves an awkward timing problem: degradation labels arrive once per cycle, orders of magnitude more slowly than the 10-millisecond control steps, and the training set drifts monotonically as the pack ages. The Gaussian process is refit on a sliding window of the 50 most recent cycles, with hyperparameters re-estimated by L-BFGS on a background thread that never blocks the control loop.

Real-time feasibility was a serious engineering constraint in its own right. Exact Gaussian process inference scales cubically with the number of training points, so the team split the computation between a CPU and the programmable fabric of a Xilinx Versal adaptive compute accelerator. The FPGA handles the batched kernel evaluations and the back-substitution against a cached Cholesky factor, while the CPU solves the model predictive control quadratic program with the OSQP solver, warm-started from the previous step’s solution, which cut average solver iterations from 28 to 6. Together the two processors met the 10-millisecond control period demanded by the 100-hertz sensor bus.

The experimental evidence comes from benchtop packs rather than vehicles. Each run used a 20-kilowatt-hour pack of 96 LG Chem NMC cells in a 24s4p configuration, driven by real PJM regulation-D signals resampled to one hertz and scaled to the pack rating. Three intensity regimes were tested, with the aggressive regime adding synthetic one-C bursts every five minutes to mimic contingency reserve deployment. Five independent packs from a single manufacturing lot were run per controller-regime combination for 500 equivalent cycles each, with reference capacity measured every 50 cycles by a coulombic charge-rest-discharge procedure. All statistical comparisons were two-tailed paired t-tests with pack index as the pairing variable.

The headline result concerns the aggressive regime. The proposed controller retained 82.7 plus-or-minus 0.4 percent of nominal capacity after 500 equivalent cycles, against 78.4 plus-or-minus 0.7 percent for a deterministic Arrhenius-based MPC, a paired difference of 4.3 percentage points with p equal to 0.003. It also beat a learning-based LSTM-MPC baseline that used a fixed safety band, retaining 2.6 percentage points more capacity, with p equal to 0.012. The gap over the LSTM baseline widened from 1.4 points in the smooth regime to 2.6 points in the aggressive one, which the authors read as evidence that explicit uncertainty handling matters most precisely when operating conditions are most variable. The price in grid service was small: normalized tracking error rose by under 5 percent relative to the deterministic controller, a difference that was not statistically significant, and the error distribution under the new controller was actually narrower, with a better 5th-percentile error.

Safety metrics improved as well. Across all aggressive runs, the constraint-violation rate fell from 3.1 percent of control steps under the deterministic controller to 1.2 percent, a paired difference of 1.9 percentage points with p below 0.001, and temperature excursions above 45 degrees Celsius dropped from 0.9 percent to 0.26 percent of steps. Ablation experiments attributed the gains to specific modules: removing the chance constraints cost 3.6 percentage points of retention and nearly quadrupled the violation rate, freezing the online learning cost 2.4 points, and replacing the temporal convolutional network with simple coulomb counting cost 2.7 points because the filter’s drift contaminated the training labels. The softened chance constraint itself activated in fewer than 0.1 percent of control steps, so the safety mechanism did not cause optimization failures. The degradation predictor held its one-step-ahead error below 3.5 percent throughout, validated against independent reference capacity checks every 50 cycles.

The method also survived stress tests. Capacity retention stayed within 0.6 percentage points when the prediction horizon varied between 20 and 40 steps, was statistically indistinguishable across sliding-window sizes of 30, 50 and 80 cycles, and degraded only modestly when sensor noise was tripled, because the extra noise inflates the Gaussian process variance and thereby widens the safety back-off automatically. Sweeping the risk level traced a smooth trade-off, with a stricter 1 percent risk level buying 0.3 points of retention at a 7 percent tracking-error cost. A cross-chemistry check on a 15-kilowatt-hour lithium-iron-phosphate pack showed a smaller but still substantial improvement of 3.7 plus-or-minus 0.5 percentage points, with the authors noting that LFP’s flat voltage curve makes state of charge a weaker predictive feature and suggesting feature-set adjustments rather than architectural changes for such chemistries.

The authors are candid about the limits. All packs came from a single manufacturing lot, so the variance estimates reflect within-lot rather than across-lot variability, and validation across at least three lots is needed. The Gaussian assumption behind the chance constraints may fail under heavy-tailed field noise, where one-sided Cantelli or distributionally robust alternatives would apply. No formal closed-loop stability certificate exists for this softened, online-learning configuration, although the controller remained feasible and stable in every 500-cycle trial. The temporal convolutional network was pretrained on laboratory aging data and has not been validated on cells with substantial calendar aging. Deployment also faces practical hurdles: the 0.8-millisecond rig latency is unrepresentative of cellular grid communication, and the Versal hardware adds roughly 600 dollars over an automotive microcontroller, a cost most defensible for frequent regulation service. The team does not claim any monetized return, since cell prices, market rules and warranty terms fell outside the study. What the results do establish is feasibility: an uncertainty-aware, self-updating degradation model can be made fast enough for real-time battery control and delivers measurable, statistically supported gains in battery life under the harshest grid-service conditions tested.

Subject of Research: An uncertainty-aware model predictive control scheme using online Gaussian process learning to reduce battery degradation during vehicle-to-grid service

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

Article References: Cai, K., Ma, Q., & Zhong, C. (2026). Chance-constrained model predictive control with online gaussian process learning reduces battery degradation in vehicle-to-grid service. Discover Vehicles, 2(1), Article 43. https://doi.org/10.1007/s44465-026-00043-y

Image Credits: AI Generated

DOI: 10.1007/s44465-026-00043-y

Keywords: vehicle-to-grid, battery degradation, Gaussian process, model predictive control, chance constraints, digital twin, temporal convolutional network, NMC battery, battery management system, FPGA, capacity retention, grid regulation