Three improved neural network models for air quality forecasting

Adapt radial basis function (ARBF) network and improved support vector machine (SVM) are presented in atmospheric sciences and can automatically determine the size of network and parameters, fasten the learning process and achieve good generalization performances in prediction of pollutant level.

Abstract

Artificial neural networks (ANN) are appearing as alternatives to traditional statistical modeling techniques in many scientific disciplines. However, the inherent drawbacks of neural networks such as topology specification, undue training expense, local minima and training unpredictability will overlay their merits in engineering applications, especially. In this paper, adaptive radial basis function (ARBF) network and improved support vector machine (SVM) are presented in atmospheric sciences. The principle component analysis (PCA) technique is employed to the ARBF network as well, namely, ARBF/PCA network for the convenience of expression and comparison, so as to fasten the learning process. Comparing with traditional neural network models, the proposed models can automatically determine the size of network and parameters, fasten the learning process and achieve good generalization performances in prediction of pollutant level. The simulation results based on a real‐world data set demonstrate the effectiveness and robustness of the proposed methods.