Advancing an existing project.
Forecast → detect → decide. Prediction tells the system what is coming, anomaly detection tells it what is wrong, and the dispatch layer decides what to do about both.
IoT Integration
The measurements have to exist before anything can be learned from them. We instrument the plant and get the data out reliably.
- Edge sensing & metering
- Secure MQTT / TLS uplink
- Gateway to PLC, inverter and BMS
- Historian & time-series storage
Forecasting
Load, generation and price predicted with modern sequence models — recurrent and transformer architectures with temporal attention.
- Day-ahead generation forecasting
- Short-term load prediction
- Long-term seasonal outlook
- Benchmarked against measured output
Anomaly Detection
Machine-learning models that flag faults, degradation and abnormal behaviour while they are still cheap to fix.
- Condition monitoring
- Drift & degradation tracking
- Fault classification
- Alarm ranking, not alarm floods
Uncertainty Management
A single forecast number is not a plan. We quantify the spread and dispatch against it.
- Probabilistic & interval forecasts
- Scenario generation & reduction
- Reserve sizing under uncertainty
- Deep reinforcement learning and MILP dispatch
Web-based UI
An interface an operator can read at a glance — in the style of our TCSA forecasting simulator, where every prediction can be replayed hour by hour against what actually happened.
- Live single-line diagram
- Prediction versus actual replay
- Cumulative error as it evolves
- Exportable logs & reports
Hardware where it is missing
When the instrumentation does not exist, we build it — embedded firmware, sensor integration and custom PCB assemblies designed for the measurement the model needs.
- Microcontroller firmware
- Schematic capture & PCB layout
- Design for manufacture
What an operator interface looks like
Not a picture of one — the interface itself, on measured data. Pick a plant and an architecture family, then play the horizon. The error figures are cumulative over the hours played so far, the same way an operator would watch accuracy build through the day.
Edge to closed-loop control
Got data and a control problem?
We will scope what is learnable, what is optimisable, and what should run on the edge.
Start a project →