As a typical cold-region city in China, Hohhot exhibits significant energy consumption in high-rise office buildings, while local designers lack systematic design strategies for guidance. Previous studies relied on simulation algorithms to adjust urban building designs, but this approach suffers from high time costs and limited generalizability. To address these challenges, this study proposes an innovative data-driven optimization framework for analyzing building energy performance in Hohhot’s cold climate and providing actionable design strategies. The framework integrates parametric modeling, machine learning (XGBoost/Random Forest/LightGBM/SVR), and explainable AI (SHAP) to deliver energy-efficient design solutions for high-rise offices. A case study of Hohhot’s typical high-rise building demonstrated that the XGBoost model outperformed alternatives in predicting heating/cooling/lighting demands (R2 > 0.83). Key findings revealed that heating systems account for 70–78% of total energy consumption, emphasizing compact building forms as the core design principle. SHAP analysis identified critical optimization directions for window-to-wall ratio (WWR): reducing WWR on east/west/ north facades by 30–40% decreases heating loads by 18–25%, while maintaining south-facing WWR at 35–45% balances daylight autonomy without thermal performance penalties. The proposed self-shading configuration (staggered south-facing layouts) strategically reduces cooling demand through solar occlusion. This methodology establishes a design workflow combining simulation-based approaches with explainable machine learning, offering scientific decision-making support for energy-efficient urban development in northern Chinese cities.