Research On Prediction Capability Of Forecasting Method Based On Phase Space Reconstruction And Its Application In Storm Flooding Prediction

To evaluate the predictive ability of nonlinear chaotic models based on reconstructed phase space of Lorenz system more precisely, we introduce a new evaluation method based on the nonlinear local Lyapunov exponent, instead of Root-mean-square Error which may include much more uncertainty. In this paper, the authors evaluate the predictive ability of nonlinear chaotic models in light of their predictability which can be determined according to the saturation property of the mean relative growth of initial error. We find that the predictability of reconstructed Lorenz phase space is similar to original Lorenz system, which testifies the feasibility of prediction methods based on phase space reconstruction, to certain extent. Besides, for reconstructed Lorenz phase space k-Nearest Neighbors method is not superior to zeroth-order approximation for the poor performances of some members. In addition, compared with univariate chaotic model, multivariate chaotic model have almost parallel predictive ability.To meet the requirement of embedding theory, we introduce wavelet filter and moving average filter to remove peoredical trajectraries whose existence can been proved by ln(err(t))-t graphs. However, wavelet filter is effective than moving average filter.The storm surge prediction model of BP neural network based on phase space was purposed in this paper, through combining reconstruction phase space with BP neural network. A phase space is reconstructed with the storm surge data and fitted with BP neural network model. The model is used to predict storm surge in Cuxhaven. The predictability of the model has been estimated to be 36 hours through calculating nonlinear local lyapunov exponent, while the pure BP neural network model only 21 hours. Besides, the prediction accuracy can been markedly improved through data de-noising.