This paper proposes a regime-switching recurrent network model (RS-RNN) for non-stationary time series. The RS-RNN model emits a mixture density with dynamic nonlinear regimes that fit flexibly data distributions with non-Gaussian shapes. The key novelties are: development of an original representation of the means of the component distributions by dynamic nonlinear recurrent networks, and derivation of a corresponding expectation maximization (EM) training algorithm for finding the model parameters. The elaborated switching dynamic nonlinear regimes make the RS-RNN especially attractive for describing non-stationary environmental time series. The results show that the RS-RNN applied to a real-world wind speed time series achieves standardized residuals similar to popular previous models, but it is more accurate distribution forecasting than other linear switching (MS-AR) and nonlinear neural network (MLP and RNN) models.Regime-switching (RS) models [18] become increasingly popular in practice for density forecasting in environmental sciences. They are found useful for modeling the distributions of sequentially arriving data, like meteorological time series [4,6,12,16,20,32], hydrological time series [42], wind power series [1,13,34,38,48], etc.. These switching models are essentially mixture density models that can approximate well data having distributions of various shapes, which makes them especially attractive for fitting real-world time series contaminated by large amount of noise with unknown character [24,40]. Similar tools generating probabilistic predictions along with probabilistic of their uncertainty are the mixture density neural networks [30], and the kernel density models [47].The application of the switching models to wind speed time series is currently of particular importance for the supply of renewable wind energy and its integration into the power systems. Wind speed forecasts are required to provide information about future wind energy generation and to reduce instabilities in the energy distribution. Finding accurate solutions to the task of short-term wind speed prediction addressed here is also of special interest for improving power plant scheduling and grid operations management [25].The main difficulty in this task is the continuous fluctuation of the wind speed due to the stochastic character of the atmospheric processes [8,31]. It is known that pressure, frontline passages and cyclonic conditions lead to variabilities, like nonstationarity and sudden changes in atmospheric series [1,34,35]. The problem is to find good descriptions of such time-varying data distribution with which the wind dynamics can be explained and reconstructed adequately. Although a lot of research have been conducted on describing wind series using various statistical, neural network and hybrid approaches [15,17,47], many of them do not have enough potential for capturing complex variabilities beyond simple periodicities.The statistical approaches to wind speed prediction build models using time-lagged explanat...