Abstract: Aiming at the timevarying, random and nonlinear characteristics induced delay in the networked control system (NCS), a delay prediction algorithm based on phase space reconstruction and robust extreme learning machine (RELM) is proposed. Firstly, the chaotic property of delay  sequence is detected by 0-1  test, and then the reconstruction delay parameters and embedding dimension are determined by an improved correlation integral method, and then the delay sequence is reconstructed. The new samples can more accurately reflect the delay variation features. Using the reconstructed delay sequence as a training sample, and considering the sparse characteristics of outliers, a robust limit learning machine is used to perform delay sequence prediction. The method has the advantages of fast learning, good generalization performance, effectively reducing the impact of outliers.