<p>Future prediction can be developed to exploit the dynamic feature for a preview of nonlinear system status and performance tendency. However, since the most existing prediction methods focus on the temporal feature while lacking inherent spatial feature, it is difficult for these methods to describe comprehensive dynamic feature of nonlinear systems, especially in disturbance conditions, which can result in the degradation of prediction accuracy and robustness. Thus, to solve this problem, a data-driven prediction method, using a self-organizing fuzzy neural network with temporal-spatial feature (TSF-SOFNN), is proposed. First, a spatial-temporal feature extraction strategy based on self-attention mechanism is introduced to obtain the temporal-spatial feature of nonlinear system with disturbance. This feature can be fully used to design TSF-SOFNN. Second, an adaptive feature optimal algorithm with the temporal trend and spatial correlation is developed to adjust the fuzzy rules of TSF-SOFNN. Meanwhile, a dynamic learning algorithm is employed to update parameters. As a result, the prediction performance of TSF-SOFNN can be achieved with an enhanced accuracy in nonlinear systems under the disturbance environment. Third, the convergence analysis of TSF-SOFNN is presented by Lyapunov function, which can provide a theoretical guarantee for its application. Finally, the experimental results demonstrate that TSF-SOFNN can obtain better prediction performance in series of nonlinear systems with disturbance.</p>