AI Learns to Run Renewable Microgrids 326 Times Faster Than Traditional Solvers

Imagine a power grid that can rethink its entire operating plan in under a tenth of a second — fast enough to react to a cloud rolling over a solar farm, an electric vehicle plugging in, or a sudden swing in electricity prices. That vision has moved a step closer to reality with a new study published in Results in Engineering, in which researchers in China present a framework that couples real-time battery state estimation with a neural network trained to imitate a professional optimization solver. The result is a microgrid scheduler that matches the economic performance of conventional mixed-integer optimization while running hundreds of times faster.

The study, led by Guiyong Wang and colleagues, addresses a bottleneck that has quietly constrained the renewable energy transition: as microgrids become richer in wind, solar, hydrogen systems, and electric vehicles, the optimization problems that keep them running efficiently become harder and slower to solve. Under a rolling-horizon scheme, the scheduler must re-optimize every fifteen minutes as forecasts and system states evolve. Commercial solvers such as CPLEX can handle these mixed-integer problems, but solution times of twenty seconds or more consume a growing share of the available decision window, leaving little computational margin for forecasting, communication, and safety checks.

The researchers’ system models a grid-connected microgrid comprising a 5-megawatt wind farm, a 2-megawatt photovoltaic station, a diesel generator, an electrolyzer and fuel cell linked by a hydrogen storage tank, a fleet of 200 electric vehicles, and a bidirectional connection to the utility grid. The objective function minimizes total operating cost, spanning electricity transactions, diesel fuel, operation and maintenance, pollutant emission penalties for carbon dioxide, sulfur dioxide, and nitrogen oxides, renewable curtailment penalties, and even battery degradation costs from excessive vehicle charging and discharging cycles.

What distinguishes the framework is its insistence that the electric vehicle fleet be treated not as a static block of demand but as a living, state-aware resource. A vehicle’s ability to charge faster or discharge power back to the grid depends directly on its state of charge, its battery capacity, its connection status, and the state of charge it must reach before departure. Rather than assuming these values, the team estimates them online using an enhanced unscented Kalman filter, which reconstructs battery states from noisy current, voltage, and temperature measurements.

The proposed filter, dubbed MI-SVD-NIS-UKF, layers three refinements onto the standard Kalman recursion. Singular value decomposition stabilizes the covariance matrix to keep long-running estimates numerically healthy. A normalized innovation squared statistic continuously compares actual measurements against predicted distributions, triggering adaptive adjustments of the process and measurement noise covariances when conditions change. Finally, a multi-innovation correction blends information from several recent time steps, weighted by a forgetting factor, to suppress the influence of single-step voltage disturbances. Together, these mechanisms cut the mean absolute state-of-charge estimation error to 0.3915 percent under a representative operating condition — a 53 percent improvement over the standard filter — and achieved the lowest mean error among seven filter variants across thirty Monte Carlo scenarios.

The practical payoff of that accuracy becomes vivid when estimation errors are translated into energy terms. An open-loop approach that ignores online state estimation misjudges each vehicle’s dispatchable energy by an average of 3.597 kilowatt-hours, which balloons to 719 kilowatt-hours across a 200-vehicle fleet. The proposed estimator reduces that misperception to 0.294 kilowatt-hours per vehicle, or just 59 kilowatt-hours fleet-wide. These continuously updated per-vehicle flexibility boundaries are then aggregated and handed to the scheduling layer, giving the optimizer a truthful, time-varying picture of how much flexibility the fleet actually offers at every fifteen-minute interval.

With state perception in place, the second half of the framework tackles computation. The team trained a long short-term memory network — a recurrent neural architecture well suited to capturing temporal dependencies — to imitate the optimal decisions produced by CPLEX across three years of fifteen-minute operational data. The network takes nineteen inputs, including wind, solar, and load forecasts, time-of-use electricity prices, system states, and the dynamic electric vehicle flexibility boundaries, and outputs thirteen scheduling variables over a four-step prediction horizon. Training combined imitation loss with feasibility and objective-deviation terms, encouraging the network to respect power balance, device limits, and hydrogen inventory constraints rather than merely mimic surface patterns.

Because neural networks can occasionally produce infeasible outputs, the framework wraps its inference in a layered safety net. Raw network predictions are first projected onto the continuous feasible region through a quadratic program, then binary operating states — such as whether the electrolyzer or fuel cell runs, or whether the microgrid imports or exports power — are repaired to resolve mutually exclusive logic. A final feasibility check gates every action, and any rejected prediction triggers a hard-constrained recovery optimization. In testing, this pipeline achieved a 99.8 percent final feasibility rate, with the long short-term memory architecture outperforming gated recurrent unit and multilayer perceptron alternatives on violation rate, repair success, and economic fidelity.

The speed gains are striking. On fifty unseen testing days, the neural surrogate required a median of 67.9 milliseconds per scheduling decision, compared with 22.33 seconds for CPLEX — a median speedup of 323-fold and a mean of 326-fold. On a representative day, the learned schedule cost 46,920 yuan, only 2.49 percent above the exact optimizer’s 45,780 yuan benchmark, with average root-mean-square and mean absolute errors of just 0.0579 and 0.0349 megawatts across key scheduling variables. Under deliberately harsh generalization tests — continuous rainy days and severe generation fluctuations drawn from a different year — the coefficient of determination for step-cost predictions remained above 0.984.

Scalability results reinforce the approach’s promise. When the vehicle fleet grew from 200 to 400, the exact solver’s runtime climbed from 22.33 to 30.15 seconds, while the surrogate’s inference-and-repair time rose only from 67.9 to 86.4 milliseconds, staying below a tenth of a second throughout. That leaves an enormous computational margin within each fifteen-minute interval for state updating, forecast processing, and communication — margin that future grids, with larger and more heterogeneous fleets and tighter coupling among energy carriers, will likely need. The authors position the framework as a transferable foundation for small- and medium-scale grid-connected microgrids, with extensions toward stronger uncertainty, multi-microgrid coordination, and multi-timescale scheduling on the horizon. If such state-aware surrogate methods mature, the real-time nervous system of a decarbonized energy network may one day think in milliseconds rather than minutes.

Subject of Research: State-aware surrogate learning for fast rolling scheduling of multi-energy microgrids

Article Title: State-aware surrogate learning for fast rolling scheduling of multi-energy microgrids

Article References: Wang, G., He, J., Luo, Z., Chuan, Z., Li, J., & Li, Y. (2026). State-aware surrogate learning for fast rolling scheduling of multi-energy microgrids. Results in Engineering, 32, Article 113317. https://doi.org/10.1016/j.rineng.2026.113317

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113317

Keywords: microgrid, rolling scheduling, state of charge, unscented Kalman filter, LSTM, surrogate model, electric vehicles, hydrogen energy, renewable energy, CPLEX, mixed-integer optimization, energy management