Forecasting of fine particulate matter based on LSTM and optimization algorithm

• Main Authors: Zaini N., Ahmed A.N., Ean L.W., Chow M.F., Malek M.A.
• Other Authors: 56905328500
• Format: Article
• Published: Elsevier Ltd 2024
• Subjects: Air pollution; Artificial intelligence; Deep learning; Optimization; Particulate matter; Air quality; Empirical mode decomposition; Energy utilization; Forecasting; Health risks; Learning algorithms; Long short-term memory; Particles (particulate matter); Population statistics; Sustainable development; Traffic congestion; Air pollution forecasting; Air quality monitoring stations; Fine particulate matter; Forecasting accuracy; Optimisations; Optimization algorithms; Particulate Matter; Performance; PM 2.5; Particle swarm optimization (PSO)

Accurate air pollution forecasting may provide valuable information for urban planning to maintain environmental sustainability and reduce mortality risk due to health problems. The city with higher industrial activities, traffic congestion, population density, and energy consumption is most likely to produce higher pollution than the other cities. Therefore, this study uses hybrid deep learning models to forecast air pollution based on the concentration of particulate matter with diameter size of less than 2.5 ?m (PM2.5) for two air quality monitoring stations in Kuala Lumpur, Malaysia. The proposed models predict the hourly air pollutant concentration based on 4-h historical input based on six air pollutant data, meteorology parameters, and PM2.5 concentration data from the neighboring air quality monitoring stations. Long short-term memory based on metaheuristic algorithms, namely particle swarm optimization and sparrow search algorithm (PSO-LSTM and SSA-LSTM), are first developed and applied to determine the significance input combination to the changes of PM2.5 concentration at respective target stations. Then, the input configuration that gives the best forecasting accuracy was selected for subsequent experiments using enhanced approaches based on ensemble empirical mode decomposition (EEMD-PSO-LSTM and EEMD-SSA-LSTM). Subsequently, this study also analyzed the contributions of the neighboring PM2.5 dataset to the fluctuation of PM2.5 concentration at target stations. It is found that EEMD-SSA-LSTM model of M5 improves other models in Batu Muda and Cheras by 2.65% and 20.00% for RMSE and 9.31% and 25.30% for MAE, respectively. The results indicate that the proposed model yields the highest forecasting accuracy compared to the other models, and additional information on neighboring PM2.5 significantly improves the forecasting accuracy at both target stations. Besides that, comparing the performance of the two optimization approaches, SSA provides better performance compared to PSO in optimizing LSTM hyperparameters to forecast PM2.5 concentration. � 2023 Elsevier Ltd