建構基於集成式深度學習解析臺灣循環經濟發展之關鍵因素

Abstract

循環經濟已日益成為永續發展與資源治理的核心策略。然而，其發展涉及多重要素的交互作用與高度非線性的系統動態，使得如何有效掌握影響因素與發展成效之間的關聯成為挑戰。為此，本研究提出一套整合特徵工程與類神經網路的資料驅動建模方法，以辨識影響循環經濟發展指標的關鍵因素，並進行情境模擬與優化分析，以支援政策規劃與策略調整。本研究採用 2013 至 2022 年臺灣公開之物質流與統計資料，以貝葉斯優化（Bayesian Optimization, BO）、注意力機制（Attention Mechanism, AM）與自編碼器（Autoencoder, AE）優化人工神經網路（Artificial Neural Networks, ANN）、深度神經網路（Deep Neural Networks, DNN）及循環神經網路（Recurrent Neural Networks, RNN）。並以均方根誤差（Root Mean Squared Error, RMSE）、均方誤差（Mean Squared Error, MSE）及平均絕對誤差（Mean Absolute Error, MAE）選擇最佳模型，並透過隨機搜尋（Random Search）與前向連結交叉驗證（Forward‐Chaining Cross‐Validation）確保模型穩健性。因素貢獻度則藉由 SHapley加法解釋（SHapley Additive Explanations, SHAP）進行詮釋。結果顯示，鋁箔包容器回收清除處理費費率、鋁容器回收清除處理費費率及植物纖維容器回收清除處理費費率為臺灣循環經濟發展之主要驅動因素，彰顯經濟誘因在推動循環經濟發展的重要性。並且，在「擴大生產者責任（2023_E）」情境中，將容器處理費率以 2022 年之數據為基準，提高 10%、政府於環境保護投資金額比例提升 0.7% 以及提高綠色採購金額 10%。對比 2022 年各發展指標之數據，將可以分別提升臺灣資源生產力、廢棄物回收率和再生物料使用率 5.3 ％（預估 86.1 元/公斤）、11.2%（預估 61.6%） 和 3.7%（預估 27.8%），亦可降低人均廢棄物產量 11.3%（預估 427.39 公斤/人）以及進口端之環境負荷密度（預估 0.83）。最後，本研究提出一套資料驅動的決策支援系統，以協助政策制定者評估政策成效。

Circular economy has progressively emerged as a core strategy for sustainable development and resource governance. However, its evolution entails the interplay of multiple determinants and highly nonlinear system dynamics, rendering the elucidation of the relationships between driving factors and developmental outcomes challenging. To address this, the present study proposes a data-driven modeling framework that integrates feature engineering with neural network architectures to identify key drivers of circular economy development indicators, and to perform scenario simulation and optimization analyses in support of policy planning and strategic adjustment. Utilizing publicly available material flow and statistical data for Taiwan from 2013 to 2022, we optimize artificial neural networks (ANN), deep neural networks (DNN), and recurrent neural networks (RNN) via Bayesian Optimization (BO), attention mechanisms (AM), and autoencoders (AE). Model selection is based on root mean squared error (RMSE), mean squared error (MSE), and mean absolute error (MAE), and robustness is ensured through Random Search and forward-chaining cross-validation (FCCV). Factor contributions are interpreted using SHapley Additive Explanations (SHAP). Results indicate that the treatment-fee rates for aluminium-foil packaging containers, aluminium containers, and plant-fiber containers constitute the principal drivers of Taiwan’s circular economy development, underscoring the importance of economic incentives. Under the “Extended Producer Responsibility (2023_E)” scenario—wherein treatment-fee rates are increased by 10% relative to 2022 levels, the government’s share of environmental protection investment is raised by 0.7%, and green procurement expenditures are increased by 10%—resource productivity, waste-recycling rate, and recycled-material utilization rate in Taiwan are projected to improve by 5.3% (to NT$ 86.1 kg⁻¹), 11.2% (to 61.6%), and 3.7% (to 27.8%), respectively. Concurrently, per-capita waste generation is expected to decrease by 11.3% (to 427.39 kg capita⁻¹) and the import-side environmental impact density to decline to 0.83. Finally, we present a data-driven decision-support system to assist policymakers in evaluating policy effectiveness.