Air Pollution forecasting using Long Short-Term Memory Networks in the Bay of Algeciras (Spain)

Abstract

Continuing with similar studies developed in the Bay of Algeciras, a new method from deep learning has been used. Long Short-Term Memory (LSTMs) are applied to predict future concentrations of SO2, NO2and PM10in the port of Algeciras (Spain). Maritime data from the port of Algeciras are kindly provided by the Algeciras Bay Port Authority and meteorological and pollution data are hourly collected in twenty-one monitoring stations at different points of the Bay of Algeciras from 1st January 2017 to 31stDecember 2019 provided by Andalusian Regional Government. The structure of an LSTM consists of a deep net with feedback connections that can process entire sequences of data which makes it better to forecast relevant pieces of data in the sequence and preserve it for several instants of time due to it can therefore have both short-term (like basic Recurrent Networks) and long-term memory. In this work, LSTM models have been used to predict future values of SO2, NO2, and PM10. A cross-validation strategy was adopted to obtain generalization results for unseen patterns. Furthermore, a Bayesian regularization procedure for hyperparameter searching was applied to avoid overfitting. The results have been very promising, obtaining R values above 0.75 for all pollutants in the 2h-ahead predictions as well as above 0.6 in the 8h-ahead forecasts. These results have been more accurately obtained than using shallow artificial neural networks as authors used in previous studies.