機器學習於沼渣水熱炭加值化之應用

Abstract

厭氧消化可將有機生質物轉化為沼氣，惟過程將同時產生大量副產物-沼渣。基於循環經濟理念影響，傳統被視為廢棄物之沼渣逐漸被視為具有利用價值之生物資源，進而強化沼渣加值化之必要性。水熱碳化可將沼渣轉化為具有多元用途之水炭，係為具有前景之沼渣加值化技術。機器學習(ML)可作為加速開發適宜特定用途生物炭工藝製程之有效工具。本研究首次利用機器學習技術預測沼渣水熱炭產率，其中使用專屬沼渣之資料集進行計算。藉由裝袋法-隨機森林(RF)與提升法-極限梯度提升(XGB)兩種集成樹演算法，本研究藉由沼渣元素分析/近似分析特性與水熱碳化反應參數預測沼渣水炭產率。XGB相較於RF具有更好之預測性能，其中測試R2和均方根誤差(RMSE)分別為0.911和19.15。水熱碳化參數為影響水炭產率之主要因子。本研究另基於機器學習分析結果的啟發，提出整合反應溫度、時間與原料固含量之修正激烈因子（Ro'），相較於傳統激烈因子(Ro)具有較佳之適用性。本研究進一步預測沼渣水炭產率、水碳性質(碳(Cc)、氫(Hc)、氮( Nc)、氧(Oc)、硫(Sc)、灰分(Ashc)、高位熱值(HHVc)與水熱碳化製程參數，包含能量產率(EY)、能量緻密率(ED)與碳回收率(CR)。整體而言，XGB 和 RF 在 Cc、Hc、Oc、Sc、Ashc 和 HHVc之預測展現出合宜性能，測試R2 分別為 0.856-0.942 和 0.864-0.947。本研究另開發可同時預測產率與水炭特性(Cc、Hc、Nc、Oc、Sc、Ashc、HHVc)之多工模型，其中XGB相較於RF可展現較佳之預測效能，其平均測試R2可達0.895，與現行文獻測試成效相仿。SHapley Additive exPlanations (SHAP) 分析結果顯示沼渣碳含量(C)與水熱碳化反應溫度(T)係為影響多工預測成效之主導因子。鏈回歸技術可強化多工模型之預測效能，涵蓋範疇包含EY、ED 和 CR。根據實驗驗證結果，本研究開發之XGB單任務模型對於水炭特性Cc、Oc、ashc 和HHVc 之預測展現可接受的預測性能，顯示所開發之機器學習模型可有效地預測水炭特性，據此有利於優化水熱碳化製程參數與選擇沼渣水炭之適宜應用。

Anaerobic digestion (AD) converts organic biomass into biogas while generating voluminous digestate byproducts, which has been viewed as waste while recently gradually being recognized as a valuable resource based on circular economic concepts, emphasizing the necessity of digestate valorization. Hydrothermal carbonization (HTC) is a promising technique for valorizing digestate into hydrochar with versatile applications. Machine learning (ML) is an effective tool for expediting the biochar engineering process for specific end-use. This research used ML techniques to predict digestate-derived hydrochar for the first time, with the dataset including only digestate as the biomass resource. Two ensemble tree-based machine learning algorithms based on random forest (RF) (bagging) and eXtreme Gradient Boosting (XGB) (boosting) predicted the digestate-derived hydrochar yield from digestate elemental/proximate compositions and HTC reaction parameters. XGB shows better predictive performance than RF, with test R2 and RMSE of 0.911 and 19.15, respectively. HTC parameters are the dominant factors affecting yield prediction. Inspired by the features importance analysis from ML, the new modified severity factor (Ro') integrating the effect of reaction temperature, time, and solid loading was proposed, showing better generalizability than the conventional severity factor (Ro). Moreover, digestate-derived hydrochar yield, properties (Cc, Hc Nc, Oc, Sc, Ashc, HHVc), and HTC process index, including energy yield (EY), energy densification (ED), and carbon recovery (CR) were predicted. XGB and RF showed satisfactory performance in predicting Cc, Hc, Oc, Sc, Ashc, and HHVc, with test R2 of 0.856-0.942 and 0.864-0.947, respectively. The multi-task model for predicting yield and hydrochar properties (Cc, Hc, Nc, Oc, Sc, Ashc, HHVc) was also developed. XGB outperforms RF, with the average test R2 achieving 0.895, comparable to the current published work. The SHapley Additive exPlanations (SHAP) analysis reveals that digestate C content and HTC temperature (T) dominate multi-task predictions. The chain regressor technique enhanced the model performance toward multi-task prediction, including EY, ED, and CR. Based on the experimental validation results, the developed XGB single-task shows acceptable predictive performance toward hydrochar Cc, Oc, ashc, and HHVc, suggesting the developed ML model satisfactorily predicts hydrochar properties, benefiting for optimizing HTC process parameters and determining suitable applications for digestate-derived hydrochar.