Abstract: Under the energy transition background, cross-regional distribution network investment faces challenges of data heterogeneity, regional disparities, and model generalization limitations. Traditional methods struggle to handle the complex nonlinear relationships and dynamic uncertainties of multi-source variables, leading to increased decision errors in data-scarce regions. To address this, we propose a domain-adaptive deep transfer learning framework that solves cross-regional investment benefit prediction problems through dynamic feature alignment and multi-task joint optimization. The model incorporates adversarial training and correlation propagation algorithms at the feature layer to enable knowledge transfer of regional load characteristics and policy sensitivity. At the model level, it combines multi-objective optimization with uncertainty quantification to support cost-risk tradeoff decisions in high-penetration renewable energy scenarios.