<p>Presently, Data-Driven Wind Prediction (DDWP) is capable of achieving precise and efficient wind forecasting outcomes. However, the lack of convergence and inefficiency resulting from the extensive parameter space of DDWP pose significant challenges in simultaneously achieving high fidelity and training efficiency. This challenge arises from the imbalance between the complexity of the physical calculations in the current network architecture and the oversimplification of neural networks. Therefore, it is structurally and spatially sophisticated to design an efficient physically guided weather forecasting model. We propose a physics-informed operator, called Dynamic Graph Differential Operators (DGDO), which leverages oriented differentiation on the multi-scale spherical spaces with dynamic graph structure. DGDO utilizes Graph Neural Network (GNN) with multi-order derivatives, and cross multiplications to establish the simulation of the iterative updating process like numerical methods in accordance with physical equations. Furthermore, to address the challenge of capturing the dynamic atmospheric circulation process and to mitigate the complexity of the network parameter search space, we propose a dynamic graph attention mechanism grounded in dynamic equations. This approach aims to develop a computational framework that aligns more closely with numerical methods, thereby ensuring physical fidelity. Through a series of comprehensive experiments, we have demonstrated that DGDO not only enhances prediction accuracy but also significantly reduces computational time compared to traditional models, thereby facilitating its deployment in operational settings.</p>