htm-monitor is an unsupervised early-warning monitor for critical systems, built to stay quiet until something matters. Its power grid results are scored in the open on GridShedWatch, a benchmark of every US load shed since 2020, next to four other methods and the operators’ own public notices.
Scored on GridShedWatch, an open benchmark: six and a half years of public hourly data for each grid, the same rules and alert budget for every method, every alert next to what the grid operator announced at the time, and the limits stated. The engine itself takes any continuous numeric signal.
Every US load shed for lack of supply since 2020, replayed from public data on the open GridShedWatch benchmark, scored next to four other methods, and shown next to what the operators announced.
Early warning of grid emergencies from hourly demand and generation data
GridShedWatch is a separate, method-neutral benchmark: the events, data, shared rules and scoring, with an operator-style rule, a seasonal z-score, an LSTM forecaster, Chronos-Bolt and htm-monitor as reference methods. Anyone can score their own method on it.
GridShedWatch: the full leaderboard and how to add a method →The engine is Hierarchical Temporal Memory — a biologically-inspired sequence-learning algorithm extended with a grouped-consensus decision layer.
Each signal model learns its temporal patterns continuously and adapts to drift without a retraining cycle. No static limits on the raw signal values. No supervised labels.
An alert needs two different kinds of evidence: demand and generation running abnormal, and the forecast or the supply-demand balance breaking down. A watch needs them within a few hours of each other; a warning allows them within the same day but must hold for 12 hours. The consensus layer keeps single-signal noise from becoming an alarm.
Every alert is traceable to specific signal anomalies at specific timesteps. Per-model, per-group visibility lets operators understand exactly what triggered and why.
Runs single-process in Python on a CPU. Scoring needs no external services, so it can run on a small VPS or inside an air-gapped VM.
Most anomaly detection systems generate noise. htm-monitor is built for the opposite: say nothing until something genuinely matters, then say it early enough to act. Foresight is power.
We believe the same unsupervised temporal learning engine can serve any system that produces continuous numeric data: learn what normal looks like, and flag when reality departs from it. That is a design goal, not yet a result. The scored case studies on this site are power grids.
A system that cries wolf loses trust. htm-monitor is designed for rare, high-confidence alerts — not dashboard noise.
No labeled training data required. The system learns from the structure of normal operations and flags departures automatically.
Every alert is fully auditable, traceable to specific signals and timesteps. Operators can see exactly what the system sees.
The engine takes any continuous numeric time series and makes no grid-specific assumptions. Published results cover only what has been scored.
htm-monitor is developed and operated by DataGrip LLC. We're focused on making unsupervised temporal anomaly detection practical and accessible for critical infrastructure operators.
Whether you want a free technical preview, a paid pilot, or just want to learn more — we'd love to hear from you.
Send a 90-day telemetry sample and one past event you'd like scored. You get a preview audit within a week: lead/lag timing, false-positive rate, per-signal analysis. No commitment.
Full historical replay against your event catalog, with a written per-event audit. The fee credits in full against your first subscription month.