TinyML Air Quality Forecaster

Most student air-quality projects stop at displaying the current PM2.5 reading — useful, but reactive. The genuinely valuable step is forecasting: pollution follows daily rhythms driven by traffic, cooking hours, temperature inversions and wind, and those patterns are learnable. This project builds a forecaster that runs on the sensor node itself. An ESP32 reads a PMS5003 laser particulate sensor plus temperature, humidity and pressure, keeps a rolling history in RAM, and feeds it to a tiny quantized regression network that predicts PM2.5 a few hours ahead — no cloud, no API key, no subscription. Training uses the well-documented Beijing multi-site air-quality dataset from the UCI repository: hourly pollutant and meteorological readings from 12 stations across four years, the standard public dataset for this task. The report is honest about the limits: a single node's forecast is approximate and is evaluated against the dead-simple persistence baseline ("tomorrow equals today") that any forecaster must beat to be worth anything.

1. The ESP32 polls the PMS5003 (PM2.5/PM10) and BME280 (temperature, humidity, pressure) on a fixed schedule and appends readings to a rolling history buffer in RAM.
2. Every forecast cycle, firmware assembles the recent history window (design: trailing 24 hourly-equivalent samples, resampled) plus time-of-day features into the model input vector.
3. The quantized regression network outputs the predicted PM2.5 for the forecast horizon (design: 1–3 hours ahead), entirely on-device.
4. Firmware maps the predicted value to an air-quality band (Good / Moderate / Poor per standard breakpoints) and updates the OLED with the current reading, forecast band and trend.
5. Crossing into the Poor band drives the alert LED/buzzer; the band thresholds are firmware constants the student can adjust.
6. The model is trained offline on the Beijing dataset with the identical windowing, quantized to int8, and flashed with the firmware; the report includes the persistence-baseline comparison.

Tech stack:

• ESP32 (sensing + inference)
• PMS5003 laser PM2.5/PM10 sensor
• BME280 temperature/humidity/pressure sensor
• TensorFlow Lite Micro (int8 regression model)
• Python training pipeline (pandas, TensorFlow)
• Beijing multi-site air-quality dataset, UCI ML Repository
• 0.96-inch OLED display, buzzer, LEDs
• Arduino IDE / PlatformIO (C/C++ firmware)

Dataset & model details

• Dataset: Beijing Multi-Site Air-Quality Data — hourly air-pollutant (PM2.5, PM10, SO2, NO2, CO, O3) and meteorological (temperature, pressure, dew point, rain, wind) readings from 12 nationally-controlled monitoring sites in Beijing, March 2013 to February 2017, hosted on the UCI Machine Learning Repository. Public research dataset.
• Task: Short-horizon regression — input = trailing window of PM2.5 plus temperature/humidity/pressure and time-of-day features, output = predicted PM2.5 at the forecast horizon (1–3 hours ahead).
• Model: Compact regression network — a small LSTM or MLP (design target ~10k–50k parameters), trained in Python with the identical windowing used on-device, post-training quantized to int8 and deployed with TensorFlow Lite Micro.
• Metrics: MAE and RMSE on a held-out time period, compared against the persistence baseline; forecast skill is a design target reported from the actual training run in the delivered report — never claimed as measured in advance. On-device forecast behavior is characterized by the student's own logging procedure.

• [On-device PM2.5 forecasting] A quantized regression network predicts near-term PM2.5 from the sensor history and local meteorology — the forecast is computed on the ESP32, not fetched from an API.
• [Real multi-year training data] The model is trained on the Beijing multi-site air-quality dataset (UCI ML Repository): hourly PM2.5, gases and weather from 12 monitoring stations over four years.
• [Honest baseline comparison] Forecast skill is evaluated against the persistence baseline, and the report states plainly where the tiny model beats it and where it does not — the kind of honesty examiners respect.
• [Forecast band display] The OLED shows current PM2.5, the predicted band (Good / Moderate / Poor) and a simple trend arrow; a buzzer/LED flags Poor-band forecasts.
• [Multi-sensor node] A PMS5003 laser PM sensor plus BME280 temperature/humidity/pressure gives the model the meteorological context that drives pollution dynamics.
• [Rolling history buffer] Firmware maintains the trailing sensor history in RAM with graceful handling of gaps (sensor warm-up, power cuts), documented in the report.
• [int8 TinyML deployment] The network is quantized to 8-bit integers with TensorFlow Lite Micro, running inference on the ESP32 within its memory budget.

• Assembled air-quality node (ESP32 + PMS5003 + BME280 + OLED + buzzer)
• Trained int8 forecasting model and the training notebook (Beijing dataset pipeline)
• Firmware source with sensor drivers, history buffer and inference
• Wiring diagram and forecast-logging test procedure
• Baseline-comparison analysis (model vs persistence) in the report
• Project report PDF (theory, dataset, model, quantization, results)
• PPT presentation and viva Q&A document