A comparative study of deep learning architectures for particulate matter prediction in Lugbe, FCT Abuja, Nigeria

This study developed and evaluated three deep learning architectures: a recurrent neural network (RNN), a long short-term memory (LSTM) network and a convolutional neural network--bidirectional LSTM (CNN--BiLSTM) for predicting fine (PM2.5) and coarse (PM10) particulate matter concentrations. The models used 2-h (60-observation) sliding-window inputs and air quality and meteorological data comprising 1,211,810 entries. Hyperparameters were optimized using GridSearchCV with time-series cross-validation. Performance was assessed using mean absolute error (MAE), root mean square error (RMSE), relative absolute error (RAE), mean absolute percentage error (MAPE) and the coefficient of determination (R2). All three models demonstrated strong PM2.5 predictive capability (R2 > 0.90), explaining more than 90% of the observed variability. Model 2 showed the most balanced performance, with MAE of 2.898 μg/m3, RAE of 0.1697, RMSE of 6.453 μg/m3, MAPE of 7.7013% and R2 of 0.9166; Model 3 achieved a marginally higher R2 of 0.9188. Model 2 was considered the most reliable PM2.5 estimator because its error metrics were generally lower than those of Model 3. Prediction was weaker for PM10 across all architectures (R2 = 0.7037--0.7348), reflecting greater temporal variability and complexity. For PM10, Model 2 achieved the lowest RMSE (12.7097 μg/m3) and highest R2 (0.7348), indicating superior generalization. The results highlight the effectiveness of hybrid deep learning approaches for air quality forecasting, particularly for fine particulate matter.