Addressing the issues of insufficient accuracy in physical models and lack of physical constraints in data-driven models for real-time wind power estimation, a hybrid estimation method that integrates physical modeling with machine learning is proposed. A three-layer architecture comprising physical modeling, machine learning residual correction, and physics-informed regularization is constructed to unify estimation accuracy and physical credibility. A physics-informed regularization strategy based on sample weight adjustment is designed, which defines three physical constraints including monotonicity, density consistency, and power coefficient. Through iterative optimization, physical constraint embedding is achieved without modifying the kernel of ensemble learning algorithms. Monte Carlo simulation is employed to process power expectation values under turbulence effects, and combined with physical factors such as air density correction, yaw efficiency, and wake effects, a baseline physical model is established. A multi-level feature system is constructed, utilizing random forest and gradient boosting trees ensemble learning framework to capture systematic biases. Experimental results demonstrate that the proposed method achieves significant improvements in estimation accuracy compared with both pure physical models and pure ensemble learning model. Meanwhile, estimation results that violate physical laws are effectively suppressed, providing a new technical approach for real-time wind power estimation.