Abstract: An improved discrete physics-informed neural network method incorporating both warm-start techniques and double-precision computing was proposed to address the temporal causality failure encountered in continuous physics-informed neural networks for long-time integration of partial differential equations. The temporal causality compliance of neural network was enhanced through the introduction of a time-discretization strategy, while computational resource consumption was reduced by employing a warm-start scheme that utilized model parameters obtained from previous time-step training as initial parameters for subsequent time steps. In terms of computational accuracy, the solution precision for linear partial differential equations was improved through the implementation of double-precision floating-point arithmetic. To validate the effectiveness of the proposed method, comparative numerical experiments were designed involving both convection equations and wave equations. The results demonstrate that the application of warm-start techniques achieves a 2–3 fold improvement in computational efficiency while providing approximately 20% enhancement in computational accuracy. For long-time integration tasks, the implementation of double-precision computing is shown to yield about 10 times higher computational accuracy, though accompanied by a 7–8 fold increase in computation time. Furthermore, when additional physical constraints are incorporated, the computational accuracy is observed to be further improved by 5–6 times without significant additional computational overhead. Through the combined application of warm-start schemes and double-precision computing, high-accuracy long-time integration is successfully realized by the discrete physics-informed neural networks within controllable computational costs, thereby offering a novel solution for prolonged simulations of complex systems.