Temporal Feature Interaction Memory in Air Quality Multi Variate Time Series Prediction

Abstract

Due to the increasing number of vehicles and factories, air pollution has become a critical issue in most urban areas. Accurate prediction of air quality plays a vital role in managing and mitigating pollution, especially in metropolitan cities. Forecasting air pollution levels enables governments to take proactive measures to reduce its adverse effects. In this paper, we propose a novel attention-based approach to predict Particle Matter (PM2.5 and PM10) concentrations using Recurrent Neural Networks (RNNs). The proposed method incorporates two attention mechanisms-temporal attention and feature attention- to enhance predictive accuracy. Temporal attention identifies and emphasizes key time steps within the input sequence, while feature attention assigns greater weight to the most influential air quality variables. This dual-attention strategy is integrated with deep learning and gradient boosting techniques. Unlike conventional attention mechanisms that assign independent importance weights to features and time steps, the proposed Temporal Feature Interaction Memory (TFIM) module explicitly models nonlinear interactions between heterogenous environmental variables across temporal delays. This module enables the network to capture delayed cross-feature dependencies critical for pollution formation dynamics, which are not represented in existing attention-based air quality prediction models. Specifically, the Kaggle Air Quality dataset is employed to train a Bidirectional Long Short-Term Memory (BiLSTM) model augmented with TFIM and Temporal Feature Attention (TFA), and the predictions are further refined using XGBoost to achieve improved accuracy. The proposed method achieves a dramatic reduction in RMSE, MAE and MAPE compared to baseline models. The proposed method achieved 1.46 RMSE and 5.88 MAE for PM2.5 predictions and 1.49 RMSE and 1.37 MAE for PM10 predictions.