[Paper Review] Delhi air quality prediction using LSTM deep learning models with a focus on COVID-19 lockdown

Animesh Tiwari, Rishabh Gupta+1 more|arXiv (Cornell University)|Feb 21, 2021

Air Quality Monitoring and Forecasting24 citations

TL;DR

The paper develops and compares LSTM-based models (LSTM, BD-LSTM, ED-LSTM) for short-term multi-step and long-term PM2.5 forecasting in Delhi, evaluated across COVID-19 lockdown periods with uncertainty quantification.

ABSTRACT

Air pollution has a wide range of implications on agriculture, economy, road accidents, and health. In this paper, we use novel deep learning methods for short-term (multi-step-ahead) air-quality prediction in selected parts of Delhi, India. Our deep learning methods comprise of long short-term memory (LSTM) network models which also include some recent versions such as bidirectional-LSTM and encoder-decoder LSTM models. We use a multivariate time series approach that attempts to predict air quality for 10 prediction horizons covering total of 80 hours and provide a long-term (one month ahead) forecast with uncertainties quantified. Our results show that the multivariate bidirectional-LSTM model provides best predictions despite COVID-19 impact on the air-quality during full and partial lockdown periods. The effect of COVID-19 on the air quality has been significant during full lockdown; however, there was unprecedented growth of poor air quality afterwards.

Motivation & Objective

• Motivate robust short- and long-term air-quality forecasting for Delhi amid COVID-19 lockdown effects.
• Evaluate LSTM-based architectures (LSTM, BD-LSTM, ED-LSTM) for multivariate and univariate PM2.5 prediction.
• Develop a sliding-window, three-dimensional data representation to enable multi-step ahead forecasting (80 hours) and uncertainty-quantified one-month forecasts.

Proposed method

• Use multivariate time series with 11 features including PM2.5 and meteorological pollutants.
• Apply sliding-window to convert 2D data into 3D for N=5 past steps and Nout=10 future steps (80 hours).
• Compare LSTM, Bidirectional LSTM (BD-LSTM), and Encoder-Decoder LSTM (ED-LSTM) architectures.
• Train with Adam optimizer, 200 epochs, batch size 20, ReLU activations; evaluate via RMSE over 30 runs with mean and 95% CI.
• Investigate univariate (PM2.5 only) versus multivariate inputs; assess effects of COVID-19 lockdown and seasonal variation on model performance.
• Provide one-month-ahead forecasts with uncertainty quantification.

Experimental results

Research questions

• RQ1 How do COVID-19 lockdown periods affect Delhi PM2.5 levels across monitoring stations?
• RQ2 Which LSTM architecture (LSTM, BD-LSTM, ED-LSTM) yields the most accurate multi-step PM2.5 predictions?
• RQ3 Does a multivariate input scheme outperform a univariate PM2.5-only approach, especially during lockdown periods?
• RQ4 Can the model provide reliable one-month-ahead forecasts with quantified uncertainty?

Key findings

• The multivariate BD-LSTM model generally delivers the best short-term PM2.5 predictions across horizons.
• COVID-19 lockdown periods show a significant reduction in PM2.5 during full lockdown, with disparate patterns observed as restrictions ease.
• Uncertainty quantification is provided for long-term forecasts, enabling probabilistic interpretation of one-month predictions.
• Model performance varies with data handling (shuffling windows, seasonal training data) and with whether the input is multivariate or univariate.
• Across stations, PM2.5 correlates with PM10 and NOx metrics, supporting multivariate modeling strategies.