IEEE Access (Jan 2024)
Effective Multi-Step PM2.5 and PM10 Air Quality Forecasting Using Bidirectional ConvLSTM Encoder-Decoder With STA Mechanism
Abstract
Effective prediction of PM2.5 and PM10 levels is essential for preserving public health and informing governmental actions. Nevertheless, the unpredictable behavior of air fluxes makes it difficult to forecast these concentrations accurately. The objective is to create advanced technology for accurately evaluating the air quality in Delhi, which is essential for implementing effective measures to reduce pollution. Meteorological and air quality data were gathered hourly from 6 monitoring sites in Delhi, as part of the Central Pollution Control Board (CPCB) initiative. The data collection period spanned from June 1, 2018, to October 1, 2019, with measurements conducted at hourly intervals. Address air quality prediction as an issue of predicting spatiotemporal sequences, since individual models have difficulties in successfully capturing both spatial and temporal relationships effectively. This study proposed the hybrid of an innovative encoder-decoder model based on a Bidirectional Convolutional Long-Short term network (BiConvLSTM) with a Spatial-Temporal Attention mechanism that could achieve this by developing spatial-temporal characteristics and, a prediction system that accurately forecasts PM2.5 and PM10 concentrations across several time steps. This model performed extremely well in forecasting PM2.5 and PM10 levels, achieving impressive error metrics. For PM2.5 MSE of 0.0298, MAE of 0.1511, RMSE of 0.1728, and the coefficient of determination (R2) of 0.999. Similarly, for PM10 MSE of 0.0582, MAE of 0.1833, RMSE of 0.3134, and the coefficient of determination (R2) was 0.999. Analyze the model’s efficacy over prediction periods of 6,12,24,36,48 hours concerning the distribution of errors and the level of accuracy. This study emphasizes the effectiveness of the proposed method using an extensive comparison with state-of-the-art models such as Transformer, GNN+LSTM, and Seq2Seq with attention.
Keywords
