Modelling of Air Pollution in an Environmental System by use of Nonlinear Independent Component analysis

摘要

In this paper an empirical, non-linear state space model of a metropolitan environmental system is constructed by use of singular spectrum analysis and non-linear independent component analysis. Environmental systems are complex, high-dimensional and non-linear. Conventional modelling techniques demand expensive fundamental models, as well as costly supercomputers to effectively simulate and predict these systems. On the other hand, numerical methods such as empirical state space parameterisation and multiple-layer perceptron neural networks promise simpler models that can be accommodated on affordable desktop computers. The state space model presented in this paper is constructed by embedding and separation of the individual observations of the polluting agents. The observations are regarded as a nonlinear mixture of the underlying process state variables and are classified as deterministic by using a surrogate data technique. It is shown that non-linear separation enhances the ability of the non-linear model to predict the dependent observations, especially in the presence of unknown levels of dynamic and measurement noise. No pre-assumptions are made on the statistical distributions of the original state variables or the noise content. Instead, these distributions are estimated as mixtures of Gaussian distributions. An ensemble learning technique is implemented in the parameter estimation algorithm for the separation model. The results show a reduction in complexity in the attractor and satisfactory one step ahead predictions.