本研究基於 14 頻道 LED 燈箱系統，開發出一套可根據目標光學參數即時生成對應光譜之生成器。研究流程首先導入基因演算法 (genetic algorithm， GA)，於照度(illuminance， Ev) 1000 lux 條件下，模擬產出涵蓋 3000 K 至 6500 K 多種相關色溫 (correlated color temperture， CCT) 之多元數據集，數據維度整合了 14 頻道權重、色偏差值 (Delta u-v， Duv)、平均演色性指數 (general color rendering index， Ra) 及黑視素日光效能比 (melanopic daylight efficacy ratio， mel-DER) 等關鍵光學參數。以此數據集訓練神經網路，並再次透過 GA 針對網路超參數 (hyperparameters) 進行優化搜索，以建構預測性能最佳化之光譜生成模型。 在模型訓練程序中，本研究將數據集劃分為訓練、驗證與測試集，並針對 Ra 與 mel-DER 實施二維過採樣 (two-dimensional over-sampling) ，藉此處理數據分佈不均對預測精度之干擾。為在模型準確度與運算成本間取得平衡，本研究變更隱藏層層數與神經元數目，評估不同架構對訓練時間、預測誤差及泛化能力之影響。研究結果顯示，適度增加網路深度即便在減少神經元數量的條件下，仍可獲得較佳之訓練誤差，然而深層與寬層架構間之誤差差異微小；此外，訓練時長主要受超參數配置影響，而不同網路結構之泛化能力則表現相近。 研究的光譜生成器分成無拓展輸入與拓展輸入兩種，無拓展輸入情況下，光譜生成時間平均小於0.98秒，而拓展輸入時，光譜生成時間為15.94秒，時間比無拓展輸入久但可以預測較大的範圍。;This study developed a spectrum generator based on a 14-channel LED light box system， capable of real-time synthesis of corresponding spectra according to target optical parameters. The research methodology initially employed a Genetic Algorithm (GA) to simulate and generate a diverse dataset under an illuminance (Ev) of 1000 lux， covering various correlated color temperatures (CCT) ranging from 3000 K to 6500 K. The dataset dimensions integrate critical optical parameters， including 14-channel weights， chromaticity difference (Delta u-v， Duv)， general color rendering index (Ra)， and melanopic daylight efficacy ratio (mel-DER). This dataset was used to train an neural network， followed by a secondary GA-based optimization of the network′s hyperparameters to construct an optimized spectrum prediction model. During the model training procedure， the dataset was partitioned into training， validation， and test sets. To address the impact of data imbalance on prediction accuracy， two-dimensional over-sampling was applied to Ra and mel-DER， thereby mitigating interference from non-uniform data distribution. To achieve a balance between model accuracy and computational cost， this study varied the number of hidden layers and neurons to evaluate their respective impacts on training time， prediction error， and generalization ability. The results indicate that appropriately increasing the network depth yields superior training error even with a reduced number of neurons; however， the difference in error between deep and wide architectures remains marginal. Furthermore， training duration is primarily influenced by hyperparameter configurations， while different network structures exhibit similar generalization performance. The developed spectrum generator features two modes: "Non-extended Input" and "Extended Input." In the non-extended mode， the average spectrum generation time is less than 0.98 seconds， demonstrating high computational efficiency. In the extended mode， the generation time is approximately 15.94 seconds; although slower than the non-extended mode， it allows for a significantly broader range of parameter prediction.
