Using Ensembles of Artificial Neural Networks to Improve PM₁₀ Forecasts

• Main Authors: R. Souza, G. Coelho, A.E. Silva, S.A. Pozza
• Format: Article
• Language: English
• Published: AIDIC Servizi S.r.l. 2015-05-01
• Series: Chemical Engineering Transactions
• Online Access: https://www.cetjournal.it/index.php/cet/article/view/5128

High concentrations of atmospheric pollutants provoke negative effects that range from respiratory problems in humans to altered growth in crops due to the reduction of solar radiation. In this context, the study of suspended particulate matter (PM) in the atmosphere is especially relevant. Several works in the literature are dedicated to evaluate PM impacts and to develop models to forecast PM concentrations. Among these models, artificial neural networks (ANNs) are often employed mainly due to the facts that they are capable of learning from a set of training data samples and that they are known to be universal function approximators. However, most ANN training algorithms are susceptible to initial conditions, so the resulting models of distinct training phases may present different accuracies for the same problem. It is known from the machine learning literature that the ensemble approach, which basically combines a set of slightly different high-accuracy predictors, tends to lead to more accurate forecasts. Therefore, in this paper an ensemble of ANNs is proposed to forecast the daily concentrations of PM10 (? = 10 µm) in the city of Piracicaba, Brazil. The ensemble was trained with daily samples collected from 07.2009 to 06.2013 and evaluated with one-day-ahead forecasts from 07.2013 to 06.2014. Experiments with distinct ANN configurations were made and an average reduction of 8.85 % was obtained in the Mean Squared Error. The ensembles were compared to individual ANNs that led to the best accuracy in the training dataset. It was also verified that, when compared to distinct single ANNs, the ensemble-based approach facilitated the generation of high accuracy models, as it increased the robustness of the development process. It is important to highlight that the proposed approach can be directly applied to other scenarios related to the prediction of PM concentrations, such as different atmospheric pollutants and meteorological data.