Introducing Forecastream — probabilistic forecasting for any plant signal
We're launching Forecastream, the probabilistic forecasting module of the Industream AI Engine. Forecastream projects how any process signal will evolve over a short horizon, with a confidence interval around every prediction — so operators catch drifts before they turn into alarms. No training, no fine-tuning, no model tuning per asset. We've wrapped it in a FlowMaker box. Drop it on any tag, press deploy.
Why this matters for heavy industry
Every plant floor runs on time-series: tag histories, sensor streams, vibration traces, OEE curves. Historically, forecasting them meant one of two things:
- Train a custom model per asset — long, expensive, hard to maintain at scale.
- Use brittle rule-based thresholds — fast but blind to seasonality and regime shifts.
Neither answers the real question an operator asks: where is this signal going, and how sure are we? Forecastream flips the economics: one engine, zero training on your data, probabilistic forecasts at inference cost. For plants with thousands of tags, this means forecasting every signal by default — not just the critical few you had budget to model.
What the probabilistic output buys you
Forecastream doesn't return a single predicted value. Every prediction comes with three quantiles —
p10, p50, p90 — which together form an 80% confidence band
around the forecast median.
In practice, that changes the operator's job:
- A narrow band means the model is confident. Act on the prediction.
- A widening band is itself a signal that something unusual is happening — regime change, sensor drift, unseen operating point.
- A band drifting toward the alarm threshold is an early warning, often 1 to 2 minutes ahead of the classic threshold-based alarm. Enough time to adjust, not enough to wait.
Point forecasts hide uncertainty. Forecastream surfaces it and makes it actionable.
What we built
A Forecastream box in FlowMaker. Here's a real pipeline running on a customer site:
The flow pulls a signal from InfluxDB, feeds it to Forecastream, and writes the forecast
(p10 / p50 / p90) back with the horizon tag. Operators see observed
vs predicted side-by-side in their Grafana dashboards — and downstream alert rules can fire on
predicted excursions, not just current ones.
We've deployed Forecastream first on cooling panel temperatures of an EAF — 18 segments, forecast refreshed every 10 seconds, 128-second horizon — and it's now rolling out to chemistry, energy, and vibration signals across our customer base.
How it fits the Industream philosophy
- Open — runs fully on-premise. No black box, no SaaS lock-in, no data leaves your plant.
- No-code first — drag the box, pick the tag, set the horizon. Zero Python required.
- Edge or cloud — deploys on CPU, GPU, or NPU. Same box, any runtime.
- Hybrid-ready — chain it with physics models or Pattern Studio detectors when raw forecasting isn't enough.
Try it
The Forecastream box is shipping in the next Industream release. If you want early access
on a specific asset — vibration, temperature, chemistry, energy, tuyere camera embedding —
drop us a line.