Non-intrusive load monitoring (NILM) has gained widespread attention for improving residential energy efficiency by analyzing appliance-level energy consumption. Although machine learning-based NILM methods have demonstrated excellent performance, their effectiveness heavily relies on the availability of sufficient training data. Federated learning (FL) has emerged as a promising approach for collaboratively training NILM models by aggregating distributed knowledge across clients, effectively harnessing decentralized data reserves. However, due to significant data distribution discrepancies among clients, the global model aggregated through FL often deviates from the client-specific optimum. To address this challenge, this paper proposes a directed knowledge transfer-based personalized model learning method. In this method, clients acquire eligible models through peer-to-peer communication and perform cross-architecture knowledge transfer via knowledge distillation. Furthermore, a data-driven model trust evaluation mechanism is designed to pre-screen candidate models and guide directed knowledge transfer, thereby reducing communication overhead and improving transfer efficiency. Additionally, consistency learning is introduced to mitigate potential overfitting during personalized model training. Extensive experiments on three public datasets, PLAID, WHITED and HOUIDI, demonstrate that the proposed method achieves superior training efficiency and performance compared to existing methods.