Enhancement of Chaotic Time Series Prediction with Real-time Noise Reduction
S. K. Karunasingha, Sri Lanka · 2007
Short-term prediction of hydrological time series using chaotic dynamical systems approach is gaining popularity. However, noise can severely affect the prediction accuracy of chaos approach for prediction. Several noise reduction attempts have been made in chaotic hydrological time series analysis. Popular simple nonlinear noise reduction techniques have been used in those studies. However, Elshorbagy et al. [2] raised serious doubts on the appropriateness of their noise reduction processes. In addition, it is noticed that the approaches followed by those studies are not appropriate for real-time processing of noisy data. This study identifies the possible ways to improve real-time predictions of noisy chaotic time series, identifies appropriate techniques, and proposes a robust scheme. It also identifies the drawbacks in earlier perceptions and recognizes new areas to be explored. In this study, all the methods and procedures are first tested on a chaotic Lorenz series contaminated with known noise levels. The methods are then applied on two river flow time series. The Extended Kalman Filter (EKF) is demonstrated on the proposed procedure. The validity of the scheme is assessed using two different prediction models: Artificial Neural Networks and Support Vector Machines.