Abstract Large-scale wind farm expansion suffers from wake losses of 10%–20%, which significantly erode annual energy production. Existing approaches for wake modeling and layout optimization face trade-offs between accuracy, efficiency, and robustness due to simplified analytical models, costly CFD simulations, lack of physical consistency in data-driven methods, and high-dimensional heuristic optimization bottlenecks. This paper proposes an end-to-end collaborative framework that integrates deep learning with fluid dynamics. The framework jointly trains three modules: a physics-informed neural network that embeds RANS momentum and continuity equation residuals as soft constraints in the loss function to enforce conservation laws in sparse data regions; a graph attention network that represents turbines as nodes and upstream-downstream couplings as directed edge features, learning coupling weights via multi-head attention; and a deep reinforcement learning layer that formalizes layout selection as a Markov decision process, driving PPO training with a multi-objective reward combining power generation and wake uniformity. On an 80-turbine benchmark equivalent to Horns Rev I, the method reduces flow field prediction RMSE by 67% versus Bastankhah-Gauss and by 49% versus a CNN-only model, achieves an R² of 0.97, and compresses single-condition inference time to 41 ms. For layout optimization, AEP increases by 5.7%, wake loss decreases from 11.4% to 6.5%, and consistent improvements of 4.4%–5.9% are observed across five wind condition categories. The approach scales to 200 turbines while maintaining hundred-millisecond-level inference. This study delivers a reusable methodology chain for intelligent large-scale onshore and offshore wind farm layout optimization, balancing accuracy, efficiency, and robustness.