Frequency49.995 HzCarbon120 gCO₂/kWhSourceLive · degraded
Investor demo modeLive software demonstrator running on mock and simulated data. The same panels are designed to swap to live grid, DER and market API feeds when LEM moves into pilot deployment.

Training

Physics-supervised training history. Conservation residual and constraint violations are first-class metrics, not nested telemetry.

Production model
v0.7.1
Started 2026-07-07 17:26:24
Conservation residual
6.9MW
Max: 54.1 MW
Constraint violations
19 (0.09% of samples)
Physics loss
0.013
λ = 0.32

Physics-supervised model — internal state across training runs

Run 1 / 8 · v0.1.0
residual 48.2 MW · violations 342
coherence 0%
Mock / seeded — illustrative
chaotic → settled

Each connection's brightness and thickness is driven by conservation_residual and constraint_violations_count for the selected run. Early runs render as jittery amber traces; as physics violations fall, the network settles into a steady emerald lattice — a visual proxy for the model's grid representation becoming more physically consistent, not a real activation trace.

Model Anatomy — LEM-Former schematic

Mock / illustrative
LEM-FORMERSpatiotemporal Physics-Supervised TransformerLEM · PATENT SCHEMATIC · REV 0.7.101Voltage (V)N/I02Frequency (f)03Current (I)N/I04Power (P = V×I)05Reactive Power (Q)N/I06GPS Position07Timestamp / Local Clock08Weather Conditions09Grid Topology ContextN/I10Control Actions / Load ModulationsLEM-FORMERSpatiotemporal Physics-Supervised TransformerPositional Encoding (GPS, Time)Physics-Informed Attention LayersCausal State-Action-Effect ModellingFederated Training / AggregationAdaptive Control Policy InferenceEncoderAPhysics ConstraintsBCausal InferenceCDecoderDLearned SpatiotemporalEnergy RelationshipsTRAINEDLarge Energy Model(LEM)→ predictive stabilisation → PAVLINAinstrumented signalnot yet instrumented (N/I)
Open PAVLINA →

Left: raw signal streams flowing into the LEM-Former. Voltage, current and reactive power are shown greyed — planned instrumentation, not yet ingested. Centre: transformer stack with physics-informed attention and federated aggregation. Right: predictive stabilisation signals routed to PAVLINA for actuation.

Training runs

VersionStatusStartedLoss totalLoss physicsλCons. residualViolationsNotes
v0.7.1succeeded2026-07-07 17:26:240.2070.0130.326.919Stable — current production model
v0.7.0failed2026-07-06 17:26:24——0.40——Divergent — physics weight too high
v0.6.0succeeded2026-07-04 17:26:240.2190.0150.358.227Conservation loss curvature improved
v0.5.0succeeded2026-07-02 17:26:240.2410.0180.3011.744Wider training window
v0.4.0succeeded2026-06-29 17:26:240.2620.0220.2515.478Frequency-bounds constraint added
v0.3.0succeeded2026-06-25 17:26:240.2980.0270.2021.8132Ramp-rate constraint added
v0.2.0succeeded2026-06-21 17:26:240.3510.0330.1532.1210Increased physics weight
v0.1.0succeeded2026-06-17 17:26:240.4120.0320.1048.2342Baseline physics weight

Data provenance

  • Mock seeded dataActive (live)

    Historical grid, carbon, node telemetry, events, PAVLINA and training series seeded into the demo database.

  • Simulated autonomous nodesActive (live)

    12 representative nodes stand in for a synthetic fleet of 10,000 distributed assets.

  • Live upstream feedsUpstream error

    Elexon BMRS (metered generation) + NG ESO carbon + Open-Meteo, refreshed every 20 min via cron. Embedded solar/wind (PV_Live, NESO) disabled pending source-aligned merge. carbonintensity_uk: ok · elexon_demand: ok · elexon_generation: ok · neso_embedded_wind: disabled · openmeteo_weather: error · pvlive_solar: disabled.

  • Live control layerPlanned

    Future PAVLINA outbound actuation to real DER, EV and BESS operators via signed control APIs.