本研究以本實驗室已建立之116 kW煙管式熱水鍋爐試驗平台為對象，針對低負載操作條件下之排放與效率權衡問題，提出一套結合「量測端資料驅動代理建模」與「保守式多目標最佳化」之工程流程。受限於現場尾氣再循環(Exhaust Gas Recirculation, EGR)與燃燒空氣供應條件難以直接以質量分率精準溯源，本研究以燃料流量與兩階段風速量測結果作為核心可觀測量，建立可表徵有效燃燒空氣供應程度與等效尾氣再循環影響之代理指標，並以此作為模型輸入，以降低控制端設定值與實際燃燒狀態可能存在之不一致所致之建模偏差。代理模型採用高斯過程回歸(Gaussian Process Regression, GPR)，同時預測校正至6% O2之氮氧化物(NOx)與一氧化碳(CO)的濃度，以及水側進出口溫差(\Delta T)，並利用其預測不確定性進行風險保守設計，以提升在小樣本(N = 30)與量測波動條件下之決策可靠度。多目標搜尋則採用非支配排序基因演算法II(Non-dominated Sorting Genetic Algorithm II, NSGA-II)，在操作限制(如風量可操作範圍與代理尾氣再循環上限等)下生成帕累托非支配解集(Pareto Non-dominated set)，以解析NOx、CO與效率之相對權衡與可行操作帶。最後選取具代表性之帕累托解進行實機驗證。結果顯示，所提方法可在既定可操作域內產生具一致性之帕累托解集，並提供最低NOx與折衷解之操作候選點，作為後續操作調校與實驗規劃之決策參考。;This study investigates a 116 kW fire-tube hot-water boiler test platform established in our laboratory and addresses the trade-off between emissions and efficiency under low-load operating conditions. An engineering workflow that integrates measurement-based, data-driven surrogate modeling with conservative multi-objective optimization is proposed. Because the on-site exhaust gas recirculation (EGR) level and the combustion-air supply condition cannot be accurately traced as mass fractions, fuel flow rate and two-stage air-velocity measurements are adopted as the key observable variables. Proxy indicators are constructed to represent effective oxygen supply and the equivalent recirculation effect, and are used as model inputs to reduce modeling bias caused by potential discrepancies between controller setpoints and the actual combustion state. A Gaussian Process Regression (GPR) surrogate model is developed to jointly predict NOx and CO concentrations corrected to 6% O2, as well as the inlet–outlet water temperature difference. The predictive uncertainty of GPR is incorporated into a risk-averse, conservative design to improve decision reliability under a small sample size (N = 30) and measurement fluctuations. Multi-objective search is performed using the non-dominated Sorting Genetic Algorithm II (NSGA-II) under practical operating constraints (e.g., feasible air-flow range and an upper bound on the proxy recirculation level), producing a Pareto-optimal non-dominated solution set. This set is used to analyze the trade-offs among NOx, CO, and efficiency and to identify feasible operating bands. Representative Pareto solutions are validated on the test rig. The results indicate that the proposed approach can generate a consistent Pareto set within the feasible operating domain and provide candidate operating points for a minimum-NOx solution and a compromise solution, serving as a practical reference for subsequent tuning and experimental planning.
