To address the trade-offs among energy savings, thermal comfort, and retrofit cost in existing multi-story residential buildings in cold climate regions, this study developed a retrofit optimization framework integrating building performance simulation, artificial neural network (ANN) surrogate models, and multi-objective optimization. A typical six-story residential building in Suzhou, Anhui Province, was selected as the case building. The DesignBuilder model was calibrated against one month of measured summer data, and the design variables were screened using Morris sensitivity analysis. Latin hypercube sampling was then employed to generate 5000 samples for training separate surrogate models of annual net electricity-use intensity (ANEUI), annual discomfort hours (DH), and initial retrofit cost per unit floor area (IRC). The optimization performance of NSGA-III, MOPSO, and NSGA-III-PSO was compared, and a compromise solution was identified using a combined subjective-objective weighting TOPSIS method. NSGA-III-PSO showed a better balance between convergence and objective-space coverage. EnergyPlus re-simulation indicated that the compromise solution reduced ANEUI from 60.29 to 35.98 kWh/(m²·a) and DH from 5413.2 to 4012.5 h, corresponding to reductions of 40.3% and 25.8%, respectively, while the IRC was 398.9 CNY/m². The proposed framework provides quantitative decision support for integrated residential retrofits under comparable climatic conditions.