AI for energy optimization &
smart-grid dispatch intelligence.

The transmission and distribution layer of the UK and EU grid was engineered for a world where roughly four hundred large rotating machines produced electricity and several million passive consumers absorbed it. That world is gone. Today, more than thirteen million distributed energy resources — rooftop PV, residential heat pumps, EV chargers, behind-the-meter batteries, aggregator-controlled flexibility — sit on the edge of the network. Each of them imposes a forecast burden, a settlement burden, and a control burden that classical SCADA and EMS systems were never designed to absorb.

The cost of that mismatch is no longer theoretical. National balancing actions in the UK now exceed two billion pounds per year. Curtailment payments for wind plants located behind transmission constraints have grown six-fold since 2018. Operators are increasingly being told to run their assets against contradictory signals: maximise renewable absorption, maintain reserve margins, hold reactive headroom for voltage, respect inertia floors — and do all of it in fifteen-minute settlement windows while the underlying short-run marginal price flips sign twice an hour.

FATHOM exists because that gap cannot be closed by adding more screens to the control room. It is closed by a tightly composed pipeline of probabilistic forecasting, constrained stochastic optimisation, and real-time anomaly detection — all running on field-deployable inference engines that respect the latency, security, and observability that real grid operators require.

Most "AI for grid" pitches collapse into a single LSTM trained on the wrong feature set. We took the opposite path. FATHOM's forecasting layer is an ensemble: gradient-boosted trees on engineered weather and market features carry the median; a temporal convolutional network handles fast-moving sub-hourly dynamics; a small transformer head reconciles them against the most recent ten minutes of SCADA. Quantile regression gives us calibrated 5/50/95 percentile bands, which is what the optimisation layer actually consumes — not point forecasts.

The dispatch layer is a mixed-integer stochastic programme with rolling-horizon recourse. We solve it on commodity hardware with HiGHS and Gurobi backends, but the interesting work is in the warm-start heuristic: a graph-attention policy network trained on five years of solved dispatch instances proposes an initial integer solution that the MIP solver verifies and refines, cutting wall-clock time from minutes to seconds.

The anomaly layer uses physics-informed autoencoders. We do not throw raw waveforms at a generic deep model and hope. We embed the thermal model of a power transformer, the equivalent circuit of an inverter, and the electrochemical model of a Li-ion cell directly into the loss function, so reconstruction error has a physical meaning rather than a statistical one. That is what lets us flag a developing turn-to-turn winding fault four to seven weeks before a dissolved-gas threshold would trip.