One platform. Three lanes. Every dispatch decision a grid operator has to make.

The forecasting lane consumes three classes of signal. Numerical weather prediction ensembles (ECMWF IFS, MetOffice MOGREPS-G, DWD ICON-EU) at fifty-member resolution drive the day-ahead and week-ahead heads. Settlement and balancing data from Elexon BMRS, ENTSO-E Transparency Platform, and EPEX Spot give us market context and a calibration target. SCADA telemetry from the operator's own assets gives us the last ten minutes of ground truth — the only signal that matters for sub-hourly dispatch.

The model stack is deliberately heterogeneous. Long horizons (D+1 to D+7) lean on a gradient-boosted ensemble with engineered weather features and explicit market regime classifiers. Medium horizons (1–24 hours) use a temporal convolutional network that handles the diurnal-plus-ramp dynamics of net demand. Short horizons (sub-hour to ten minutes) use a small transformer that consumes raw SCADA at 100ms resolution and emits forecasts at 1s resolution. A meta-learner reconciles them and emits a single calibrated quantile distribution per node, per horizon.

Crucially, we do not stop at point forecasts. The optimisation lane consumes the full 5/25/50/75/95 percentile bands, which means dispatch decisions are made against a probability distribution of futures, not a single guess. That single architectural choice is the difference between a portfolio that earns the median and a portfolio that earns the upper quartile of revenue available in any given trading day.

The dispatch lane is where the money is made and lost. We formulate the problem as a rolling-horizon mixed-integer stochastic programme. The objective is portfolio P&L net of degradation cost. The constraints are: per-asset ramp and energy limits, per-cell cycle-life budgets, per-feeder network capacity, per-product contracted envelopes (DC-L, DC-H, FFR, Capacity Market), and per-portfolio reserve obligations. The decision variables are continuous power set-points per asset per settlement period, plus binary commitment for products that have on/off discontinuities.

What makes this tractable in production is the warm-start policy. We trained a graph-attention network on five years of historically solved instances; given the current portfolio state, market context, and forecast bands, it emits an initial integer solution in under twenty milliseconds. The MIP solver then verifies and tightens — usually in two to three branch-and-bound iterations rather than thousands. For a typical 200 MW / 400 MWh portfolio across four market products and a 48-hour horizon, end-to-end solve time is under 1.2 seconds at P99.

The lane is co-optimisation in the literal sense. We do not run wholesale, then run DC-L on the residual, then bolt capacity market obligations on top. We solve all of them jointly, which means the optimiser can choose to give up a wholesale trade because the cycle it would consume is more valuable held as headroom for a likely frequency event two hours later. This is the single biggest revenue uplift our customers see versus their incumbent stacks.

The third lane watches the asset, not the market. For each class of grid asset — power transformer, switchgear, inverter, Li-ion module, OLTC, capacitor bank — we maintain a physics-informed autoencoder. The encoder compresses the recent telemetry window into a low-dimensional latent state; the decoder reconstructs it. The reconstruction loss is augmented by a physics term: for a transformer, the difference between observed top-oil temperature and the IEEE C57.91 thermal model prediction; for a Li-ion module, the difference between observed terminal voltage and a coupled electrochemical + thermal model; for an inverter, the difference between observed AC current harmonics and the modulation-strategy reference.

This is not a marketing claim. It is the architectural reason FATHOM Sentinel can flag a developing turn-to-turn winding fault four to seven weeks before the dissolved-gas analysis threshold would trip, and the reason it can flag a cell-level imbalance in a 4 MWh battery rack two weeks before the BMS would isolate it. The model is not pattern-matching; it is being told what physics looks like, and complaining when reality deviates.

The lane outputs are designed for the maintenance and asset-management workflow, not just the dashboard. Every event carries a risk score, a confidence interval, a suggested intervention window, and a full audit trail of the input window, the reconstruction, the physics residual, and the model version that produced the verdict. If your reliability engineer wants to argue with the model, they can — the evidence is in front of them.