AQI Station: An Outage-Tolerant Edge–Server Architecture for Low-Cost Air Quality Monitoring with National AQI Reporting and One-Hour Particulate Forecasting

Low-cost particulate sensors can fill the space between the sparse reference monitors of Indian cities, but most low-cost designs depend on a cloud service and lose data whenever the network fails. This paper presents AQI Station, a monitor built on a tiered edge–server architecture in which the network can fail without loss of data. An ESP32 microcontroller running MicroPython samples an SDS011 laser particulate sensor, an MQ135 metal-oxide gas sensor and a DHT22 temperature and humidity sensor every 10 s and posts authenticated JSON batches to a FastAPI server on the local network. Because the node has no real-time clock, it never timestamps data: undelivered readings are journalled in flash with a monotonic tick, the node reports each reading’s age on delivery, and the server reconstructs its time; 499 readings (83 min) survive outages and power cuts. The server rejects physically impossible values rather than clamping them, computes India’s National Air Quality Index (NAQI) on read, compares the station with the six nearest Central Pollution Control Board (CPCB) monitors, and forecasts PM2.5 and PM10 one hour ahead. Pre-trained on a year of hourly data from a nearby reference station and evaluated on 1,165 chronologically held-out hours, a Ridge regression on three hourly lags reduced the PM2.5 mean absolute error from 5.66 to 5.20 μg/m³ (−8.1%) relative to persistence, whereas random-forest and gradient-boosting models did not beat the baseline. In a first field run the station delivered 711 readings over 2.3 h; its PM2.5 was 41% above the nearest monitor but within the range of the five monitors within 5.1 km. The core hardware costs about Rs. 3,000.