人工智慧整合系統：用於浮生植物監測與農畜廢棄物共消化沼氣生產

Abstract

隨著全球對再生能源的重視，能將有機廢棄物轉化為甲烷沼氣的「厭氧共消化」技術，已廣泛應用於農畜廢棄物處理與循環經濟。然而，其消化後的放流水仍富含營養物質，易引發水體優養化，因此結合浮生植物進行淨化的植物修復技術也日益受到重視。傳統上，厭氧系統的操作與植物採收多仰賴人工，常因主觀判斷而導致系統運行不穩與採收時機失準；此外，目前多數研究仍僅侷限於實驗室規模，缺乏於戶外環境之應用。 為此，本研究開發了一套人工智慧與物聯網技術的智能監測與決策管理系統，依序探討「厭氧共消化沼氣生產」與「浮生植物淨化廢水」兩大階段。首先，在沼氣生產階段，本研究採用實驗室的歷史發酵數據，利用長短期記憶網路 (LSTM) 與自回歸移動平均模型 (ARIMA) 為未來實時沼氣產量建立了完善的預測流程。隨後，在浮生植物淨化階段，本研究將應用拓展至戶外，於桃園弘智畜牧場建置獨立的物聯網監測系統，同步收集多種浮生植物的連續生長影像與淨化過程中的多項水質數據。 在數據分析與預測方面，植物淨化端的水質電導度 (EC) 趨勢預測運用了支持向量迴歸 (SVR)，其在未來 1 至 8 天的預測中，平均絕對百分比誤差 (MAPE) 均維持在 10% 以下；在影像分析上，分別以 YOLO11 與 SAM3 的兩階段語意分割架構，有效分割水體區域 (mAP50: 0.9839) 與植物範圍 (IoU: 0.9114)，以此建立精準的植物覆蓋率趨勢。 最後，本系統將上述預測與辨識模型整合至兩個圖形化使用者介面 (GUI) 中：分別為「自動覆蓋率計算」與「多源數據整合」介面。透過進一步結合大型語言模型 (LLM)，依據植物生長影像與水質數據生成符合實際操作的操作建議。本研究期望透過多源數據的自動化整合，提供適當的進料與採收策略，以此提升農牧廢棄物處理的監測與管理效率。

With the global emphasis on renewable energy, anaerobic co-digestion (AcoD), which converts organic waste into methane-rich biogas, has been widely applied in agro-livestock waste treatment. However, the nutrient-rich effluent from AcoD can cause water eutrophication if not properly managed, bringing increasing attention to phytoremediation using floating aquatic vegetation. Traditional anaerobic operation and vegetation harvesting rely heavily on manual labor, leading to system instability and imprecise timing. Furthermore, most studies remain confined to lab scales, lacking outdoor applications. To address these challenges, this study developed an intelligent monitoring and decision-management system integrating AI and the Internet of Things (IoT), sequentially exploring the two main stages of AcoD biogas production and wastewater phytoremediation by floating aquatic vegetation. First, this study utilized lab-scale historical AcoD datasets and employed long short-term memory (LSTM) and autoregressive integrated moving average (ARIMA) models to establish a robust forecasting pipeline for future real-time biogas production. Subsequently, this study expanded the application to outdoor environments by deploying an independent IoT monitoring system at the Taoyuan Hongzhi Livestock Farm to synchronously acquire continuous growth images of various floating plants and water quality data during the purification process. Regarding data analysis and prediction, support vector regression (SVR) was utilized to predict water quality electrical conductivity (EC) trends at the plant purification end, maintaining mean absolute percentage errors (MAPEs) below 10% across 1–8 day forecasting horizons. For image analysis, a dual-stage YOLO11 and SAM3 segmentation architecture precisely extracted water body regions (mAP50: 0.9839) and vegetation coverage (IoU: 0.9114) to establish precise vegetation coverage trends. Finally, these predictive and segmentation models were integrated into two distinct graphical user interfaces (GUIs): the “Automated coverage calculation” and “Multi-source data integration” interfaces. By further incorporating a large language model (LLM), the system automatically generates optimal feeding and harvesting recommendations based on the acquired multi-source data, significantly enhancing the monitoring and management efficiency of agro-livestock waste treatment.