What I'm Doing Now

The Problem: Traditional ML flattens grid nodes into flat vectors, ignoring physical connections (buses & transmission lines), causing severe errors as variable renewables increase.

The Core Idea: Model physical topology explicitly: Nodes = buses, Edges = lines. Combines GraphSAGE (spatial) + GRU (temporal) across 48 past time steps to predict voltage, angle, active power (P), and reactive power (Q).

One-line Takeaway: Knowing *who is connected to whom* substantially improves physical power-grid state forecasting.

The Problem: Global AI weather models (GraphCast, AIFS, Pangu, GenCast) are evaluated against ERA5 reanalysis data—which they were trained on—creating an evaluation feedback loop.

The Benchmark: Evaluated 7 AI weather models against 458 ground stations, rain gauges, and satellites over the South Asian Monsoon.

Major Finding: When evaluated against actual ground observations, model errors are 15–45% larger than when evaluated against reanalysis data. AIFS proved strongest overall deterministic model, while GraphCast underestimated extreme rainfall.

One-line Takeaway: AI weather forecasting is powerful, but we shouldn't trust benchmark numbers until tested against real ground observations.

The Discovery: Previous papers claimed ML warm-starts reduced IPOPT iterations by 30-46%, but they compared against an artificial flat start (Vm=1, Va=0) instead of IPOPT's true default midpoint x0 = (l+u)/2. When compared to default midpoint, primal-only ML warm starts actually made convergence worse!

The Solution: Predict the complete primal-dual-barrier state (x_hat, lambda_hat, z_hat, mu_hat) including constraint multipliers and barrier parameters.

One-line Takeaway: Predicting the solution isn't enough—you must predict the solver's entire primal-dual state.

Core Discovery: Prediction distance error e0 = x0 - x* = ρv is NOT a good measure of Newton-Raphson convergence. The error direction v dictates convergence far more than distance magnitude ρ, especially near voltage collapse where Jacobian σ_min(J) → 0.

The RL Pipeline: Replaces supervised distance loss with GRPO Reinforcement Learning where Reward = negative solver iteration count.

Result: Produces starting points that are mathematically further from true solution but converge significantly faster. Only method to converge 100% on all 2000-bus test grids.

One-line Takeaway: Replaces "predict mathematically correct answer" with "learn the starting point that makes the numerical solver behave best."

The Problem: Solar and wind generators require hourly updates to avoid severe financial deviation penalties (up to 48% revenue loss under Grid-India rules), while global NWP models update only 2-4 times daily.

The System: Fuses geostationary satellite imagery + ground stations + operational NWP into an hourly re-initialized model predicting GHI, 2m temp, 10m wind, and 100m wind up to 48 hours ahead.

Impact: Outperforms ECMWF IFS-HRES and NOAA GFS across full 48-hr horizon over Indian solar parks & wind clusters.

One-line Takeaway: Turns AI weather forecasting into an hourly operational tool specifically designed around India's renewable energy problem.