Fuzzy greenhouse climate control
A fuzzy-PI controller with a digital twin, model predictive control and AI fault detection for greenhouses, vertical farms and cold stores.
- Origin
- Started as a MATLAB fuzzy controller
- Built with
- Python
- Fuzzy logic
- MPC
- ROS 2 Humble
- Modbus TCP
- MQTT
- IEC 61131-3
- scikit-learn
A MATLAB Mamdani controller that turned a sensed and a target temperature into two PWM signals, one for a heater and one for a cooler. Fuzzy logic suits this job because it works with terms such as “cold”, “warm” and “hot” the way an operator would, and moves smoothly between them.
What is the main project focus
- A redesigned controller. An incremental fuzzy-PI with a 7 × 7 rule base that works on the error, so one rule base serves every crop and setpoint.
- Split-range actuation. Heating, free cooling through vents, then pad or compressor cooling, with no heating and cooling at the same time.
- A safety supervisor. Redundant sensor validation, frost and heat limits, interlocks and a fail-safe mode.
- A digital twin. A grey-box thermal model with solar gain, crop transpiration, and realistic sensors with injectable faults.
- Model predictive control that uses weather forecasts, with fuzzy-PI as fallback.
- Anomaly detection that finds drifting, frozen, or spiking sensors and removes them from the fusion.
- Field integration. A ROS 2 package, a Modbus TCP link to a PLC, MQTT telemetry, and export of the controller as IEC 61131-3 Structured Text or a C99 header.
Results on the digital twin

| Scenario | Controller | RMSE | Time in band | Heat + cool overlap |
|---|---|---|---|---|
| Tomato greenhouse, 72 h | Legacy FIS | 5.27 K | 19.4 % | 72 h |
| Fuzzy-PI | 0.29 K | 98.3 % | 0 h | |
| MPC | 0.34 K | 98.5 % | 0 h | |
| Sensor and boiler faults | Fuzzy-PI, rules only | 1.71 K | 65.2 % | 0 h |
| Fuzzy-PI with AI monitor | 0.35 K | 97.6 % | 0 h |
Fuzzy-PI moves the actuators about 2.5 times less than a tuned PID for nearly the same tracking, which means less wear on valves, burners and compressors. These numbers come from the digital twin, not from a real site.