AI-powered O&M command crew for utility-scale solar. SCADA telemetry → ML performance-ratio baseline → multi-stage fault detection → EUR impact pricing → LLM-as-judge routing → human-approved action drafts, grounded in EU/DE energy regulation.
Built for the Energy Hack Munich 2026 (Enerparc Open Track).
⚖️ Advisory only: never autonomous control. Every outbound action stops at a human-approval gate. This is a deliberate design choice to stay out of EU AI Act high-risk classification.
Full interactive diagram: FigJam board
Utility-scale solar O&M teams drown in raw SCADA. A 65-inverter plant throws thousands of signals a day, and the hard part isn't seeing a dip; it's knowing why. Is an inverter genuinely faulty, or was it curtailed by the grid operator (which pays compensation, not a truck roll)? Misreading that one distinction wastes field crews on non-faults while real defects (string failures, soiling, AC-side degradation) quietly bleed revenue for weeks. And every response has to respect a thick layer of EU/DE regulation (EEG §15, Redispatch 2.0, §14a, IEC 61724, NIS2, EU AI Act).
beyondWatt turns that firehose into a ranked queue of decisions a human can approve in one click. We solve it in five moves:
- Learn "healthy." An ML model predicts the Performance Ratio an inverter should hit for the current weather, the honest baseline, not a crude fleet average.
- Flag only what's real. A PR cascade (IEC 61724) screens against that baseline with persistence + deviation filters, killing weather noise and one-off sensor blips.
- Name the root cause. A cost-ordered diagnostic separates true faults (string out, soiling, conversion fault) from external interventions (curtailment, §14a dimming) that look identical on a power chart but mean the opposite for action and money.
- Price and route. Each incident is costed in EUR (lost kWh × feed-in tariff) and routed by an LLM-as-judge: dispatch a crew, file a compensation claim, or just log it.
- Draft, cite, and stop. It writes the work order or §15 EEG claim with the exact regulation cited, then parks it at a human-approval gate. Nothing executes autonomously.
The payoff: faster triage, fewer wasted truck rolls, recovered curtailment compensation, and a fully audit-trailed system that stays advisory-only under the EU AI Act.
beyondWatt watches a fleet of PV inverters and turns raw SCADA into decisions a human can approve in one click:
- Predicts the healthy-state Performance Ratio (PR) with a LightGBM model, the baseline an inverter should hit given weather + time.
- Detects deviations with a PR cascade (IEC 61724 pattern), then drills down to root cause.
- Prices the impact in EUR (lost kWh × feed-in tariff).
- Routes each incident (dispatch a truck / file a §15 EEG claim / log only) via an LLM-as-judge.
- Drafts the action + a compliance citation, and parks it in an approval queue.
| Scenario | Signal | Outcome |
|---|---|---|
| Blackout / outage | long zero-output window, irradiation healthy, neighbours fine | fault → priced → work order |
| Curtailment | DV/EVU < 100 % | not a fault → no truck roll → §15 EEG claim drafted |
| Balancing / forecast | Chronos-Bolt 24 h day-ahead | battery charge/discharge narrative |
| Reactive power | cos φ < 0.90 | setpoint recommendation; payload generated only after human approval, never written back |
A LightGBM model deployed on Render predicts expected PR per interval from weather + time features (irradiance, module/ambient temp, hour, day-of-year, month + engineered sin/cos & interaction terms). It never sees I_DC/U_DC; those are reserved for Stage-2 root-cause diagnosis (no leakage).
- Endpoint:
POST https://plant-a-pr-api.onrender.com/predict - Health:
GET /health· Metrics:GET /metrics - Trained: Plant A 2017 (Jan–Sep) · Validated: held-out 2018
2018 validation (500 sampled daytime points, 60 inverters):
| Metric | Value | Reading |
|---|---|---|
| MAE | 0.049 | ~5 PR-points typical error |
| RMSE | 0.080 | outlier-sensitive; RMSE≈1.6×MAE → a few hard low-PR points |
| R² | 0.773 | 77 % of PR variance explained on a held-out year, across unseen inverters |
| bias | +0.008 | effectively unbiased |
The chart (Overview page → Performance Ratio) plots measured vs predicted PR with hover inspection. When the model can't be reached it falls back to a calibrated physics baseline (irr × kWp × temp-derate).
Stage 1: screen: per-inverter daily PR vs the ML-predicted healthy baseline (and fleet-relative deviation), with a 3-day persistence filter and a data-derived absolute floor (~PR < 0.65–0.70).
Stage 2: root cause (runs only when Stage 1 fires, cheapest check first):
| Check | Verdict |
|---|---|
| DV/EVU < 100 % | Curtailment → §15 EEG claim, stop |
| all PRs drop together | Plant-level / weather |
| η = P_AC/(U_DC·I_DC) low | AC/conversion fault |
| I_DC step ≈ 1/n_strings, U_DC stable | String out |
| slow drift, temp-independent | Soiling / degradation |
| deficit locked to sun angle | Shading |
| seen in ticket history | Repeat offender → root-cause visit |
| Agent | Role |
|---|---|
agents/triage.py |
PR cascade: Stage 1 screen + Stage 2 root cause |
agents/impact.py |
EUR pricing (lost kWh × FiT) |
agents/ml_api.py |
LightGBM PR client (Render) + physics fallback |
agents/ml.py |
local per-inverter LightGBM (offline training/eval) |
agents/forecast.py |
Chronos-Bolt 24 h forecast client (AWS Lambda, eu-north-1) |
agents/orchestrator.py |
LLM-as-judge: 2-tier incident routing |
agents/drafter.py |
action / claim / work-order drafts |
agents/rag_compliance.py |
keyword-keyed RAG over 7 EU/DE regulation chunks → citation per incident |
pip install -r requirements.txt
uvicorn app.app:app --port 8080 # from repo rootOpen http://127.0.0.1:8080/draft, the primary dashboard (Overview · Agents · Compliance Chat · Tickets · Approval Queue).
It works out of the box. The committed SCADA sample (data/raw_data/inverters_first5_2023.csv) and the pre-built artifacts in out/*.json are all the dashboard needs; no dataset setup required.
To regenerate those artifacts from scratch (optional, needs the full Enerparc datasets placed locally):
python pipelines/pipeline.py # 5-inverter run → out/incidents.json, timeseries.json, ...
python pipelines/pipeline_2018.py # 2018 validation → out/pr_validation.json (the PR chart)
python pipelines/pipeline_fault_test.py # synthetic fault-injection evalGET /api/inverters · /api/timeseries · /api/incidents · /api/tickets · /api/briefing · /api/forecast · /api/pr_validation · POST /api/decide (audit → out/audit.jsonl) · POST /api/chat
app/ FastAPI server (app.py) + primary UI (dashboard.html)
agents/ the 8 agents above
pipelines/ data pipelines (main / 2018 validation / fault test / synthetic)
eval/ scoring (eval.py), PR chart gen, point-by-point fault-test runner
data/ committed SCADA sample (raw_data/) + organizer xlsx (System_Overview, FiT, Tickets)
docs/ Enerparc-task (system design) · agent · regulations-eu-de · demo-scenarios
assets/demo.gif walkthrough recording
out/ pre-built artifacts the dashboard serves (committed; runtime audit log git-ignored)
- EUR figures are estimates (physics/ML baseline, not certified meter): "estimated revenue impact".
- A 12 MB SCADA sample + the dashboard's
out/*.jsonare committed so the demo runs standalone. The full organizer datasets (100s of MB, full-year/65-inverter files, validation sets) are git-ignored; you only need them to regenerate artifacts via the pipelines. - Runtime paths in
pipelines/*.py//api/ticketsassume the original hackathon folder layout for raw data +2. Additional Data/. - LLM features (Compliance Chat, judge reasoning) call an Anthropic-backed helper and degrade gracefully to deterministic mock/static text when no key/module is present, so the pipelines and dashboard run offline-safe.
- The cos φ = 0.95 reactive-power recommendation is a domain rule, not from the dataset; the forecast price narrative has no live price feed behind it.
- Hard rule: every outbound action stops at the approval queue (EU AI Act advisory posture).
Built by Nico Junkers · Pavan Kumar · Rebecca Riedmayer · Vasu Chukka at Energy Hack Munich 2026.