ABSTRACT
This study aims to produce accurate predictions of the NO2 concentrations at a specific station of a monitoring network located in the Bay of Algeciras (Spain). Artificial neural networks (ANNs) and sequence-to-sequence long short-term memory networks (LSTMs) were used to create the forecasting models. Additionally, a new prediction method was proposed combining LSTMs using a rolling window scheme with a cross-validation procedure for time series (LSTM-CVT). Two different strategies were followed regarding the input variables: using NO2 from the station or employing NO2 and other pollutants data from any station of the network plus meteorological variables. The ANN and LSTM-CVT exogenous models used lagged datasets of different window sizes. Several feature ranking methods were used to select the top lagged variables and include them in the final exogenous datasets. Prediction horizons of t + 1, t + 4 and t + 8 were employed. The exogenous variables inclusion enhanced the model's performance, especially for t + 4 (ρ ≈ 0.68 to ρ ≈ 0.74) and t + 8 (ρ ≈ 0.59 to ρ ≈ 0.66). The proposed LSTM-CVT method delivered promising results as the best performing models per prediction horizon employed this new methodology. Additionally, per each parameter combination, it obtained lower error values than ANNs in 85% of the cases.
ABSTRACT
The 'Campo de Gibraltar' region is a very industrialized area where very few air pollution studies have been carried out. Up to date, no model has been developed in order to predict air pollutant levels in the different towns spread in the region. Carbon monoxide (CO), Sulphur dioxide (SO(2)) and suspended particulate matter (SPM) series have been investigated (years 1999-2000-2001). Multilayer perceptron models (MLPs) with backpropagation learning rule have been used. A resampling strategy with two-fold crossvalidation allowed the statistical comparison of the different models considered in this study. Artificial neural networks (ANN) models were compared with Persistence and ARIMA models and also with models based on standard Multiple Linear Regression (MLR) over test sets with data that had not been used in the training stage. The models based on ANNs showed better capability of generalization than those based on MLR. The designed procedure of random resampling permits an adequate and robust multiple comparison of the tested models. Principal component analysis (PCA) is used to reduce the dimensionality of data and to transform exogenous variables into significant and independent components. Short-term predictions were better than medium-term predictions in the case of CO and SO(2) series. Conversely, medium-term predictions were better in the case of SPM concentrations. The predictions are significantly promising (e.g., d (SPM 24-ahead) = 0.906, d (CO 1-ahead) = 0.891, d (SO2 1-ahead) = 0.851).
