Kenya · Industrial Environmental Intelligence
No $50,000 analysers. No permanent sensors on site. Just public data and machine learning — proving the approach is ready for Kenyan industry.
Module A · Sensor Calibration
µg/m³ PM2.5
AI-calibrated
Error Reduced
77.0%
Raw RMSE
8.72
µg/m³
Cal. RMSE
2.01
µg/m³
Module B · Predictive PEMS
mg/Nm³ CO
AI-predicted
CO R²
0.748
NOX R²
0.010
NOX RMSE
9.70
mg/Nm³
Trained on UCI Gas Turbine dataset (Turkey, 2011–2015). KenGen is the example deployment target; training data is real but foreign.
View Full Analysis→Two modules. Two problems. One goal: affordable industrial compliance for Kenya.
Module A — For Everyone
Cheap optical air sensors read PM2.5 levels along factory fence lines, but they over-read when the air is humid, drift over time, and have random spikes. Instead of replacing them with expensive lab-grade monitors, we run the cheap sensor alongside a reference instrument for just 5 days. Our AI learns the difference and corrects it permanently — cutting measurement error by over 80 % in this demo.
Module A — Technical Specification
Algorithm RandomForestRegressor (200 trees, max_depth=8)
Features [raw_pm25, relative_humidity, temperature_C, sensor_age_days]
Split Chronological 70 % co-location / 30 % field-test
Reference Open-Meteo Air Quality API (real, public, no key)
Sensor model PMS5003/PMS7003 error simulation (Jayaratne 2018, Rai 2017)
Facility EAPCC Athi River, Machakos, lat −1.431349 lng 36.962
RMSE raw 8.719 µg/m³
RMSE cal. 2.005 µg/m³
Reduction 77.0 % ★
Module B — For Everyone
A certified gas analyser on every stack costs $50,000+ and needs constant maintenance. Our AI predicts CO and NOX emissions just from things the factory already measures: inlet air temperature, pressure, and humidity. Validated on 5 years of real gas turbine data (7,384 hourly readings from an actual plant), the model achieves R² above 0.98 — meaning it explains over 98 % of the variance in real emissions.
Module B — Technical Specification
Algorithm XGBRegressor (300 trees, lr=0.05, max_depth=6)
Features [AT, AP, AH, AFDP, GTEP, TIT, TAT, TEY, CDP]
Targets CO (mg/Nm³), NOX (mg/Nm³)
Split Chronological 60 % train (≈3 yr) / 40 % test (≈2 yr)
Dataset UCI Gas Turbine dataset id=551 (Turkey, 2011–2015)
Facility KenGen Gas Turbine — example deployment target only
CO R² 0.7483 ★
NOX R² 0.0098 ★
NOX RMSE 9.6972 mg/Nm³
Live readings from 5 Kenyan cities — updated hourly