시기별 시계열 분할 접근법을 이용한 하수슬러지 발생 및 처리 방법 예측 딥러닝 모델 개발

Abstract

Forecasting sewage sludge generation through complex treatment pathways is uniquely challenging in South Korea owing to severe statistical disconnections caused by major policy interventions, such as the 2012 Ocean Dump Banning. This study accordingly constructed a consistent 21-year (2004–2024) time-series dataset by applying linear interpolation to bridge policy-induced gaps and reclassifying statistical items to ensure longitudinal consistency. The primary innovation of this study lies in its development and validation of a novel two-stage, three-model ensemble gated recurrent unit (GRU) architecture designed to interpret complex, non-stationary data. The constructed dataset exhibited multiple distinct dynamic regimes (e.g., policy shock, transition, and stability) that required specialized modeling approaches. Our hierarchical framework effectively addressed these different regimes by: (1) separating the forecasting of the macro-scale “generated_total” sludge from the seven treatment pathways, and (2) training separate expert GRU models optimized for long-term, mid-term, and short-term time horizons. Model evaluations conducted using the 2024 test set demonstrated highly accurate predictions of macro-trends driven by stable policy and infrastructure, predicting the “generated_total” sludge with a 1.99% mean absolute error (MAE) and “fuelization” sludge with a 2.91% MAE. Furthermore, the model determined that highly volatile sludge treatment pathways such as “composting” (19.04% MAE) are governed by different mechanisms, including external structural factors and regulatory quality standards, rather than simple macro-level drivers. This performance divergence validates the effectiveness of the proposed ensemble approach in disentangling the effects of policy, infrastructure, and external dynamics across different time periods. Consequently, this study established a robust and interpretable modeling framework offering considerable value in data-driven environmental policy planning.