BiLSTM Air Quality Forecasting

Project Overview

An applied deep-learning research project: a Bidirectional LSTM (BiLSTM) network trained on a 5-year NASA air-quality dataset for Baghdad (2020–2025), forecasting five pollutants — CO, SO₂, SO₄, PM2.5, and NO₂ — and mapping predictions to US-EPA Air Quality Index health categories. Results are served through a standalone, single-file interactive dashboard rather than a static notebook.

Problem

Air-quality pollutant levels are noisy, seasonal, and interdependent, which makes reliable short- and medium-term forecasting genuinely difficult — small time series make standard train/test splits unreliable, outlier spikes distort ordinary loss functions, and different pollutants respond very differently to the same modeling approach.

Solution

A BiLSTM sequence model per pollutant, trained with Huber loss for robustness to concentration outliers, circular harmonic (sine/cosine) month encoding to capture seasonality, and walk-forward cross-validation appropriate for a small time series. 95% confidence bounds come from Monte Carlo Dropout. Every result is backed by a full statistical battery — MSE, RMSE, MAE, MAPE, R², Durbin-Watson, Pearson and Spearman correlation — plus ADF/KPSS stationarity testing on the underlying series before modeling.

Key Features

• Overview: at-a-glance best-performing pollutant, MAPE, Pearson r, and training configuration (60 epochs, early stopping).
• Pollutant Predictions: observed-vs-predicted charts on the held-out test set, per pollutant.
• AQI Results: predicted Air Quality Index mapped to US-EPA health categories (Good / Moderate / Unhealthy for Sensitive Groups / Unhealthy), with observed-vs-predicted comparison.
• 12-Month Forecast: forward-looking pollutant forecasts with Monte-Carlo-Dropout uncertainty bounds and a month-by-month value table.
• Full Metrics Table: the complete evaluation matrix for all five pollutants side by side.
• Stationarity Tests: ADF and KPSS test results per pollutant, checked before modeling.

Architecture

A two-script pipeline, run in sequence: airquality_lstm_publication.py handles data loading, stationarity testing, seasonal feature engineering, per-pollutant BiLSTM training/evaluation, and forecasting — writing results out as CSVs and figures. generate_dashboard.py then compiles those outputs into a single standalone HTML5 file with Chart.js, so the dashboard has no server dependency at all — it's a static artifact that happens to be interactive.

Technology Stack

Engineering Highlights

• Huber loss over MSE: chosen specifically for robustness to concentration outliers in pollutant readings, rather than defaulting to a squared-error loss that outliers would dominate.
• Seasonal encoding done right: month is encoded as circular harmonic sine/cosine features rather than a raw integer, so the model doesn't see December and January as maximally far apart.
• Walk-forward cross-validation: used instead of a single random train/test split, which matters for a genuinely small time series where a lucky or unlucky split can badly mislead the evaluation.
• Uncertainty quantification: 95% confidence bounds on the 12-month forecast come from Monte Carlo Dropout, so the forecast isn't presented as a single false-precision line.
• Rigorous, honest evaluation: the dashboard reports full metrics for every pollutant — including the ones where the model underperforms — rather than only showcasing the best result.
• Standard-compliant AQI: AQI is computed per US-EPA methodology for NO₂ and CO, with an explicit unit-conversion note in the code for the remaining species (the raw data ships as mol/mol mixing ratios, not the μg/m³ or ppm the standard expects).
• Zero-dependency delivery: the dashboard is a single generated HTML file — no backend, no build step — which is also why it deploys as a static Render site with no server-side moving parts to keep running.

Results

Model performance varies meaningfully by pollutant. NO₂ forecasting performed strongly: R² = 0.9515, MAPE = 1.70%, Pearson r = 0.977 on the test set, with the derived NO₂ AQI achieving 0.7% MAPE and correct health-category classification across all 10 test months. PM2.5 showed moderate fit (R² = 0.561), while CO, SO₂, and SO₄ were harder to model reliably in this run (negative R² on the test set — reported transparently rather than omitted).

Links