<p>In this paper we consider adaptive deep neural network approximation for stochastic dynamical systems. Based on the continuity equation associated with the stochastic dynamical systems, a new temporal KRnet (tKRnet) is proposed to approximate the probability density functions (PDFs) of the state variables. The tKRnet provides an explicit density model for the solution of the continuity equation, which alleviates the curse of dimensionality issue that limits the application of traditional grid-based numerical methods. To efficiently train the tKRnet, an adaptive procedure is developed to generate collocation points for the corresponding residual loss function, where samples are generated iteratively using the approximate density function at each iteration. A temporal decomposition technique is also employed to improve the long-time integration. Theoretical analysis of our proposed method is provided, and numerical examples are presented to demonstrate its performance.</p>