Deep-Learning models for daily PM10 forecasts using feature selection and genetic algorithm

Abstract

With the continuous development of the economy and its industrial activities, air pollution has become a serious problem. Therefore, it is absolutely necessary to develop a very accurate air quality forecasting model. In This paper, ten years of records of air pollution parameters and meteorological observations were used to forecast one-daily ahead of PM10 (particulate matters with a diameter less than $10 \mu\mathrm{m}$) for two stations in Casablanca city, Morocco. Recurrent deep learning models namely: Long short-term memory (LSTM), Recurrent Neural Network (RNN), and Gated Recurrent Unit (GRU) are proposed. All of these nonlinear models were tuned using the genetic algorithm (GA) technique, which performed well. Among various combinations of predictors, the EFS (Exhaustive feature selection) method selected the best combination of predictors based on statistical scores mainly MSE. The analysis of the three prediction results shows approximately a similar performance. Interestingly, good scores were observed in terms of Pearson correlation coefficient (r), coefficient of determination (R), mean absolute error (MAE), and root mean squared error (RMSE), allowing decision-makers to anticipate the PM10 ground-level accurately.