EnglishDeutsch/api/energyGitHub · iceccarelli
Grimaldi EngineeringCluster control · Frankfurt
Energy Intelligence · Cluster 1

Target architecture

GridOS is the one product. Nine modules, each with a reason to be separate. Shared primitives only where they remove duplication a named consumer actually has.

GridOS — module mapexistsmerging / specified / researchmissing
GRIDOS KERNELDERIM — DER integration — Protocol adapters (OPC UA, Modbus, IEC 61850, EEBus/SunSpec) and device onboarding.DERIM — DER integrationintegration · merging · derim-middlewareEnergie Teilen — §42b EnWG energy-community economics, tariff logic, reports and paid pilot intake.Energie Teilendata · exists · energie-teilenAsset / SCADA integrations — Read paths into existing SCADA / asset systems and asset-health features for the anomaly module.Asset / SCADA integrationsintegration · missing · Advanced Asset Insight*Digital twin — Network model, asset parameters and state — the single schema the simulator, optimizer and agents read.Digital twindata · specified · GridOSForecasting — Probabilistic load, PV, price and constraint forecasts with tracked error.Forecastingprobabilistic · exists · GridOS (MPC loop)Optimization — Dispatch, flexibility and curtailment optimization against explicit constraints; MILP first, learned surrogates only with benchmarks.Optimizationdeterministic · exists · GridOS (MILP dispatch)Physics engine — Power flow (Newton baseline), contingency, inertia / RoCoF checks; PINN surrogates when they beat the baseline.Physics enginedeterministic · research · physics-informedAI planning — LLM / agent planner: reads state, proposes candidate actions and explanations through the tool interface only.AI planningprobabilistic · specified · gridforge (analyst)NeuralBridge safety runtime — Policy engine, risk evaluation, authorization, actuator gateway, hash-chained audit.NeuralBridge safety runtimedeterministic · missing · NeuralBridge (unlocated)

Dashed edge = probabilistic module; it has no path to an actuator. Solid edge = deterministic or data/integration.

Target tree

GRIDOS
├── DERIM — DER integration
├── Energie Teilen
├── Asset / SCADA integrations
├── Digital twin
├── Forecasting
├── Optimization
├── Physics engine
├── AI planning
└── NeuralBridge safety runtime

Modules

ModuleFromPurposeSeparate becauseKindState
GridOS kernelGridOSTelemetry ingestion, asset identity, time-series storage, API surface, tenant boundary.It is the product. Everything else plugs in here through versioned contracts.dataexists
DERIM — DER integrationderim-middleware, GridOS (OPC UA adapter)Protocol adapters (OPC UA, Modbus, IEC 61850, EEBus/SunSpec) and device onboarding.Adapters touch OT networks. Their blast radius is isolated from the analytics and from the planner.integrationmerging
Energie Teilenenergie-teilen§42b EnWG energy-community economics, tariff logic, reports and paid pilot intake.A different buyer (Vermieter, WEG, Stadtwerke) and a regulatory surface that changes on its own clock.dataexists
Asset / SCADA integrationsAdvanced Asset Insight*, Advanced Mobile SCADA Suite*, BatteryTrack*Read paths into existing SCADA / asset systems and asset-health features for the anomaly module.Read-only OT access needs IEC 62443 zoning independent of any write path. Nothing here may write.integrationmissing
Digital twinGridOS, physics-informed (schemas)Network model, asset parameters and state — the single schema the simulator, optimizer and agents read.CIM-derived schemas outlive any solver or model; they are a contract, not a feature.dataspecified
ForecastingGridOS (MPC loop), gridforge (feeds)Probabilistic load, PV, price and constraint forecasts with tracked error.Models are swapped and re-trained on their own cadence; the optimizer only sees a distribution interface.probabilisticexists
OptimizationGridOS (MILP dispatch)Dispatch, flexibility and curtailment optimization against explicit constraints; MILP first, learned surrogates only with benchmarks.Deterministic, auditable, reproducible — it is the last stage before the policy engine that may propose an action.deterministicexists
Physics enginephysics-informed, Renewables_Migration_*_Proof_EnginePower flow (Newton baseline), contingency, inertia / RoCoF checks; PINN surrogates when they beat the baseline.The constraint check must be independent of the model that proposed the action.deterministicresearch
AI planninggridforge (analyst), GridOS (agent tools)LLM / agent planner: reads state, proposes candidate actions and explanations through the tool interface only.Probabilistic intelligence is isolated from deterministic safety enforcement. It never touches an actuator.probabilisticspecified
NeuralBridge safety runtimeNeuralBridge (unlocated), mcp-foundry (governance kernel)Policy engine, risk evaluation, authorization, actuator gateway, hash-chained audit.The only component allowed to talk to an actuator. Certifiable in isolation; no ML dependency; own release cadence.deterministicmissing

Hard separation triggers

Components stay isolated when coupling would create any of these:

cybersecurity riskOT riskcertification riskdeployment riskvendor lock-inunnecessary dependenciesexcessive blast radiusunclear ownership

Shared primitives

Each primitive answers the five questions. One without answers does not get built.

PrimitiveWho uses itEliminatesReduces costSecurity benefitUnlocksFromState
Authorization & audit (governance kernel)AI planning, NeuralBridge, APIThree separate permission checks and two log formatsOne audit format to review, one place to revokeHMAC token gate, hash-chained append-only logAI-Act traceability, certifiable authorization pathmcp-foundryextract
Asset identity & digital-twin schemakernel, DERIM, simulator, optimizer, agentsPer-module asset tablesOne migration when a device type changesOne place to scope tenant access to assetsContingency analysis, hosting-capacity productsGridOS + CIM (RWTH thesis lineage)specified
Telemetry ingestion & time-series abstractionkernel, forecasting, anomalyTwo backends (Timescale, Influx) addressed twiceBackend swap without touching modelsOne ingress to rate-limit and authenticateMW/assets-connected KPI becomes measurableGridOSexists
Event & candidate-action schemasplanner, simulator, policy, NeuralBridgeAd-hoc JSON between stagesContract tests replace integration debuggingSigned, versioned, replayableCross-cluster events (GridOS anomaly → work order) without couplingnew, smallspecified
Simulation & policy interfacesoptimizer, physics, agentsSolver-specific call sitesNewton vs PINN swap is one adapterConstraint check independent of proposerBenchmarks with identical harnessphysics-informedspecified
Experiment tracking & model registryforecasting, physics, researchNotebook-only resultsReproducible runs, no re-derivationProvenance of every deployed modelHonest "state of the art" claimsnew (MLflow or equivalent)specified

As of 2026-09-05Machine-readable: /api/energy/architectureAll resources: /api/energy/index

Written and maintained by Vincenzo Ceccarelli Grimaldi, Frankfurt am Main.