Surrogate-Accelerated Multi-Objective Optimization for Energy Scenario Analysis A Python framework for large-scale energy performance optimization under climate scenario uncertainty, combining high-fidelity surrogate modelling with evolutionary multi-objective search across high-dimensional parameter spaces. Overview This framework addresses a core computational bottleneck in scenario-based energy analysis: physics-based simulation models are too expensive to evaluate at the population scales required for robust Pareto-front exploration. The solution implemented here replaces direct simulation calls with a Bayesian-optimized CatBoost surrogate model, enabling efficient sampling across high-dimensional design and policy spaces under both historical and future climate forcing. The optimization targets three coupled performance objectives — energy demand intensity, thermal discomfort risk, and operational carbon emissions — and resolves trade-offs between them using NSGA-II evolutionary search. Results are analyzed under two climate scenarios derived from regional climate projections, demonstrating non-linear performance divergence across scenario space under changing forcing conditions. This methodology is domain-agnostic: the surrogate-optimization pipeline is structurally applicable to any setting where expensive simulation models constrain scenario ensemble size — including power system models, integrated assessment models, and industrial process optimization. Methods Surrogate modelling: CatBoost gradient boosting with Bayesian hyperparameter optimization (Optuna). Model selection benchmarked against Random Forest, XGBoost, and neural baselines. Evaluated on RMSE, R², and out-of-sample generalization across scenario conditions. Interpretability: SHAP-based global sensitivity analysis to identify which input parameters drive performance divergence between climate scenarios — analogous to variance-based sensitivity analysis in uncertainty quantification. Optimization: NSGA-II multi-objective evolutionary algorithm for Pareto-front resolution across three objectives simultaneously. Non-dominated solution sets analyzed via parallel coordinates and density distributions. Climate scenario integration: Baseline and mid-future climate forcing derived from regional climate projections. Scenario-specific boundary conditions propagated through the surrogate to quantify performance uncertainty under non-stationary climate conditions. Repository Structure surrogate-accelerated-multiobjective-optimization/ ├── data/ │ ├── raw/ │ ├── processed/ ├── src/ │ ├── surrogate/ │ ├── optimization/ │ └── analysis/ ├── notebooks/ ├── results/ │ ├── plots/ │ └── metrics/ ├── requirements.txt └── README.md Requirements pip install -r requirements.txt Core dependencies: pandas, numpy, scikit-learn, catboost, optuna, pymoo, shap, matplotlib Usage python# Train surrogate model python src/surrogate/train.py --scenario baseline --config configs/default.yaml
python src/optimization/nsga2_run.py --surrogate models/catboost_baseline.pkl Key Results Surrogate model achieves R² > 0.97 across all objectives with RMSE well within acceptable bounds for optimization use. NSGA-II identifies non-dominated solution sets demonstrating significant trade-offs between energy demand reduction and thermal resilience under future climate forcing. Climate scenario comparison reveals non-linear performance divergence that coarse statistical approaches fail to capture.