01The problem
Energy telemetry is most valuable exactly when the network is failing. Devices need to keep observations, replay them safely and let operators trust what an alert is based on.
02The approach
- Keep observations in a SQLite WAL outbox on the edge and in IndexedDB in the browser; remove entries only after the server acknowledges their event IDs.
- Ingest through authenticated batches and an MQTT QoS 1 subscriber, detecting duplicate and conflicting events by UUID.
- Forecast load and solar with a seasonal-median baseline scored on a time-ordered holdout against persistence, and flag load spikes with rolling median/MAD.
03The evidence
- The repository runs a full demo with 72 hours of deterministic synthetic data and scripted outage, restoration, load-spike and low-battery scenarios.
- Forecasts are reported with holdout MAE next to a persistence baseline instead of on their own.
04Scope & status
A runnable reference implementation. Demo readings and replay fixtures are synthetic, not field measurements, and the forecasting baseline is deliberately simple.
Public source is available for independent inspection.