A Hybrid Approach for Short-Term NO2 Forecasting: Case Study of Bay of Algeciras (Spain)

Abstract

A hybrid model is proposed in this research in order to forecast concentration values of NO2 with one-hour prediction horizon in the air quality monitoring network of the Bay of Algeciras area (Spain). Air pollution is an important environmental problem these days and it requires control. However, it is not an easy task. The main problem is that air pollution data series are non-lineal and non-stationary. Thus, techniques based on regression and simple models are not able to entirely capture the phenomenon behaviour. A LASSO-ANN hybrid model is proposed. The first step has been to predict the linear part of the time-series performing a least absolute shrinkage and selection operator (LASSO) model. Later, an artificial neural network (ANN) model has been performed to predict the residual sequence, the unexplained part of the LASSO model. The chaotic residual behaviour has been smoothed using an autoregressive moving window and applying the window median. The last step has been to aggregate the predicted NO2 value and its predicted residual. The model has been validated and tested using cross-validation based on R correlation coefficient, MSE, MAE and d index of agreement, and also Friedman test and LSD test. In addition, the proposed approach has been compared to a simple ANN model. The results reveal that hybrid model presents a better performance than a multiple linear regression and also a simple ANN model. The main purpose is to develop a forecasting model capable of capturing the non-linear information of the variable and increase the accuracy of the outputs.