Evaluation and analysis of artificial neural networks and decision trees in forecasting of common air quality index in Thessaloniki, Greece

摘要

Air quality management in urban areas requires reliable and accurate computational methods for the forecasting of the concentration levels of pollutants. The Common Air Quality Index (CAQI) has been used by the European Environment Agency for assessing air quality in a harmonized way. We have evaluated the values of this index in Thessaloniki, Greece, in 2001-2003, using a wide range of Computational Intelligence (CI) models. We have applied Artificial Neural Networks (ANN) and Decision Trees (DT) for the forecasting of the CAQI, and we compared the results with those obtained via statistical regression models. An extensive number of computational experiments were performed, in which we evaluated the influences of (ⅰ) different model architectures, (ⅱ) various input datasets, and (ⅲ) the training methods of these models. Model sensitivity was analyzed, in terms of various modelling options. In total, the performance of 1118 different modelling options was investigated. The best of the evaluated models performed well in forecasting both the hourly and daily values of the CAQI. The use of model ensembles (based on the same algorithms and structures), obtained by a k-fold cross-validation, resulted in more accurate forecasts than using individual models. However, the performance of various ANN and DT models was found to be dependent both on their internal structure and on the methods used for their training. The presented results are expected to be useful in developing and implementing operational air quality management and forecasting systems.