Smart Air Pollution Forecasting Using Deep Learning: Investigating the Role of Meteorological Factors

Air pollution is an important environmental concern in urban areas, and short-term forecasting of fine particulate matter (PM۲.۵) can support timely public-health warnings and environmental decision-making. This study evaluates a Long Short-Term Memory (LSTM) neural network for hourly PM۲.۵ forecasting and examines whether meteorological information improves prediction beyond recent PM۲.۵ observations. The analysis uses ۴۳,۸۰۰ complete hourly observations from a Kaggle dataset covering January ۲, ۲۰۱۰, through December ۳۱, ۲۰۱۴. The multivariate model uses dew point, temperature, atmospheric pressure, wind speed, wind direction, cumulative snow, and cumulative rain together with historical PM۲.۵. A ۲۴-hour input window is used to predict the following hour, and evaluation is performed on a chronological hold-out test set. Performance is assessed using RMSE, MAE, and R². The univariate LSTM achieves lower RMSE and MAE and higher R² than the multivariate model. The results indicate that, for this dataset and model configuration, recent PM۲.۵ history provides strong short-term predictive information, while the available meteorological variables do not improve aggregate point-forecast accuracy. The findings are interpreted cautiously because feature representation and forecasting conditions may affect the contribution of weather variables.

کلیدواژه ها:

Air Pollution Forecasting ، Deep Learning ، Long Short-Term Memory (LSTM) ، Meteorological Factors ، PM۲.۵ Time-Series ، Smart Cities