IEEE Conference Submission · 2026

An Explainable CNN-BiLSTM for Forecasting Industrial Air Pollution in Indian Metros.

A hybrid deep-learning framework that pairs 1D convolutional feature extraction with bidirectional recurrent sequence modeling, and decomposes SHAP attributions jointly across seasonal and diurnal axes — turning a black-box AQI forecaster into a policy-grade tool. Evaluated on six years of CPCB hourly data across Delhi, Mumbai, Kolkata, and Chennai.

0.9874
R² on Delhi
13.83
RMSE
23.2%
RMSE↓ from CNN block
08:00
Peak attribution (IST)
4
Indian metros evaluated
6
Pollutants modeled
Overview

What this paper does, in a paragraph.

Industrial air pollution in Indian metropolitan areas remains a serious public health problem, and AQI readings cross hazardous levels in cities like Delhi almost every winter. For policy interventions and citizen advisories to actually work, the underlying forecasts have to be both accurate and interpretable. This paper proposes a CNN-BiLSTM model paired with SHAP GradientExplainer, and decomposes the explainability output across season and hour-of-day to surface actionable intervention windows.

Hybrid Architecture

1D convolution extracts local temporal motifs; stacked bidirectional LSTM models forward accumulation and backward dissipation. Closed by a small FC head.

Explainability

SHAP GradientExplainer gives per-timestep, per-feature attribution. Aggregated globally, then stratified seasonally and diurnally.

Multi-City Honesty

Strong on Delhi and Mumbai; clearly weaker on Chennai because the current six-feature input cannot capture sea-breeze meteorology. We say so explicitly.

Architecture

CNN-BiLSTM with post-hoc SHAP attribution.

Three sequential stages, then a separate explainability pass.

Input
24-hour Multi-Pollutant Window
PM₂.₅, PM₁₀, NO₂, CO, SO₂, O₃  ·  shape (B, 24, 6)
Stage 1 · CNN
1D Conv + BatchNorm + ReLU + MaxPool
64 filters  ·  kernel=3  ·  dropout=0.2
Stage 2 · BiLSTM
Stacked Bidirectional LSTM
128 → 64 hidden units per direction
Stage 3 · Head
Fully Connected Prediction Head
128 → 64 → 1  ·  dropout=0.3
Output
Scalar AQI ŷ
Inverse min-max scaled to AQI range
Post-hoc
SHAP GradientExplainer
φᵢ,ₜ  ·  per-feature, per-timestep attribution
Results

CNN-BiLSTM beats six baselines on the Delhi test set.

We compare against ARIMA, SVR, Random Forest, LSTM, GRU, and CNN-LSTM. All metrics on the inverse-scaled AQI; deep models share a 15% chronological test split.

RMSE comparison (lower is better)

Walk-forward ARIMA shown for completeness; not directly comparable to chronological deep-learning evaluation.

ModelMAERMSEMAPE (%)
ARIMA (2,1,2)2.277.252.080.9336
SVR (RBF)13.7118.727.800.9770
Random Forest13.7417.809.520.9792
LSTM14.4118.279.290.9780
GRU14.7218.539.280.9774
CNN-LSTM10.8914.196.710.9868
CNN-BiLSTM (Ours)9.7413.835.990.9874

ARIMA evaluated via walk-forward one-step-ahead validation on a 30-day subset, feeding the most recent observed value before each prediction.

Multi-City Generalization

Strong on Delhi and Mumbai. Weaker on Chennai — and we say why.

Chennai's R² of 0.68 is not a failure to hide. The current six-feature input doesn't capture sea-breeze meteorology, which is exactly what coastal Tamil Nadu air quality depends on. Adding wind direction and a sea-breeze index is a natural next step.

R² across cities

Climate diversity sharpens the model's limits.

RMSE across cities

Mumbai's low absolute error reflects its lower AQI variance.

CityMAERMSEMAPE (%)
Delhi9.7413.835.990.9874
Mumbai5.136.287.890.9388
Kolkata8.8710.8616.400.8993
Chennai15.4721.6918.910.6808
SHAP Explainability

Why explainability is the actual contribution.

A high-R² forecaster is not the new thing. The new thing is decomposing SHAP attributions jointly across season and hour-of-day, so a policymaker can read off when each pollutant matters most. Three findings stand out.

Global

PM₂.₅ dominates

φ̄ = 0.0033 globally, followed by PM₁₀ at 0.0025. Consistent with PM₂.₅'s heavy weight in the CPCB sub-index formula.

Seasonal

PM₁₀ overtakes during monsoon

Rainfall preferentially scavenges fine PM₂.₅, leaving the coarse fraction dominant. Counterintuitive at first glance, clean once you know the chemistry.

Diurnal

08:00 IST is the intervention window

Both PMs peak sharply at 08:00 IST, exactly when morning traffic is at its worst. NO₂ and O₃ peak between 10:00–14:00 in the photochemical cycle.

Global feature importance

Mean |φ| aggregated over the Delhi test set.

Seasonal attribution shift

Notice the PM₂.₅ → PM₁₀ inversion during monsoon.

Diurnal attribution by pollutant

Hour-of-day SHAP attribution. Shaded morning rush window highlighted in the paper.

Ablation Study

What each component is actually doing.

Component-level ablation on the Delhi test set. Removing the CNN block hurts the most — confirming that local convolutional feature extraction is doing the heavy lifting, not the recurrence alone.

ΔRMSE when component removed

Higher bar = component matters more.

ConfigurationMAERMSEΔRMSE
CNN-BiLSTM (Full)9.7413.830.9874
w/o Bidirection10.8514.000.9871+0.17
w/o BatchNorm12.3816.220.9827+2.39
w/o CNN13.8518.010.9787+4.18
Authors
Akash Nath
Dept. of Computer Science & Engineering
Assam University, Silchar, India
akashnath.aus@gmail.com
Pragyat Jyoti Baruah
Dept. of Computer Science & Engineering
Assam University, Silchar, India
Arnab Paul
Dept. of Computer Science & Engineering
Assam University, Silchar, India
Arun Jyoti Nath
Ecology and Environmental Science
Assam University, Silchar, India
Tirthanka Borah
Dept. of Computer Science & Engineering
Assam University, Silchar, India
Kamalesh Debnath
Department of Management Studies
NIT Silchar, Assam, India
Citation

If you build on this work, please cite us.

BibTeX entry below. Click to copy.

nath2026aqi.bib
@inproceedings{nath2026aqi,
  title     = {An Explainable Deep Learning Architecture for Forecasting
               Industrial Atmospheric Pollutants of Indian Metropolitan Cities},
  author    = {Nath, Akash and Baruah, Pragyat Jyoti and Paul, Arnab and
               Nath, Arun Jyoti and Borah, Tirthanka and Debnath, Kamalesh},
  booktitle = {Proceedings},
  year      = {2026}
}