Methods for reconstruction chaotic dynamical system structure directly from experimental time series are described. Effectiveness of general methods is illustrated with the results of numerical simulation. It is of common interest that from the single time series it is possible to reconstruct a set of interconnected variables. Predictive power of dynamical models, provided by the nonlinear dynamics inverse problem solution, is limited firstly by the noise level in the system under study and is characterized by the horizon of predictability. New physical results are presented, concerning nonstationary and bifurcation nonlinear systems: 1) algorithms for revealing of nonstationarity in random‐like chaotic time‐series are suggested based on discriminant analysis with nonlinear discriminant function; 2) an opportunity is established to predict the final state in bifurcation system with quickly varying control parameters; 3) hysteresis is founded out in bifurcation system with quickly varying parameters; 4) delayed correlation ⟨noise‐prediction error⟩ in chaotic systems is revealed. ©1996 American Institute of Physics.