<p>Direction of Arrival (DoA) estimation has important applications in radar, satellite, and wireless communication systems. However, real-world signal reception systems often suffer from array mismatches due to factors such as antenna pattern and channel delay errors, significantly affecting DoA estimation accuracy. Traditional single-snapshot methods face high computational costs and insufficient robustness to array mismatch, limiting their scope of application. This paper proposes a deep learning method named Joint Loss and Array Mismatch Compensation Network (JLAMCNet), which combines a Direction-Aware Contrastive Loss function (DACL) and a spatial spectrum loss function to simultaneously optimize the accuracy of DoA estimation and feature representation capabilities. The proposed Antenna-Aware Spatial Attention mechanism (AASA) generates adaptive spatial weights to focus on key information and compensate for feature biases caused by array mismatch. Additionally, JLAMCNet introduces an adaptive optimization mechanism that dynamically adjusts the loss function weights based on different mismatch conditions, enhancing system robustness. Experimental results show that JLAMCNet achieves higher DoA estimation accuracy and significantly improved robustness to array mismatches compared to traditional methods and existing deep learning approaches, particularly in multi-target scenarios.</p>