Evaluation of Sulphur Dioxide Hourly Prediction Using Long Short-term Memory for Summer and Winter Season

Abstract

Increasing air pollution has necessitated the prediction of pollutants over time. Deterministic, statistical, and machine-learning methods have been adopted to predict and forecast pollutant levels. It aids in planning and adopting measures to overcome the adverse effects of air pollution. This study employs long short-term memory (LSTM). This study used the hourly data from a meteorological station in a low-town area, Mohammedia City, Morocco. The model prediction accuracy was evaluated based on hourly, weekly, and seasonal (summer and winter) readings for the summer and winter of 2019, 2020 and 2021. Root mean square error (RMSE), mean absolute error (MAE) and mean arctangent absolute percentage error (MAAPE) were calculated to evaluate the accuracy of the developed LSTM model. The MAE value of 0.026 was observed to be less in winter than 0.029 during summer in 2019. Also, it was observed that MAE values decreased from Year 2019-2021, indicating increased prediction accuracy. MAAPE also observed a similar trend. However, RMSE values indicated the opposite for 2019 and 2020; in 2021, the RMSE value was 0.21 for summer and 0.14 for winter for hourly readings. Based on the error calculation, the study found weekly hourly readings were the most accurate for predicting SO2 concentration. Also, the LSTM model was more accurate in predicting winter SO2 concentration than in the summer season. Further studies must incorporate local incidences affecting the SO2 concentration into the LSTM model to increase its accuracy.