Prediction boundaries and forecasting of non linear hydrologic stage data

This paper analyzes and forecasts hydrologic stage data in the Loxahatchee National Wildlife Refuge, the northernmost extent of the Florida Everglades. Analysis indicates that the process dynamics are chaotic, for which several attractor invariants are evaluated. A connection is made between the Kolmogorov-Sinai (KS) entropy of the phase-space trajectories and limits of temporal stage predictability. Evaluation of the KS entropy establishes a boundary for forecast time periods. Comparisons are presented between inferences produced by linear statistical models and the nonlinear attractor invariants. A nonlinear estimator (feedforward neural network), along with several linear models are employed to perform temporal forecasts of stage data. The observed degradation of forecast accuracy is consistent with the limits inferred from the attractor entropy. The nonlinear estimator is found to have better prediction accuracy than the linear models for prediction intervals beyond several days.