딥러닝과 최소분산조정기법을 활용한 계층적 물동량 예측

Traditional parcel volume forecasting primarily targets regional aggregates; however, efficient logistics operations increasingly require granular insights at the product-category level. Forecasting such hierarchical structures presents dual challenges: capturing the high volatility of specific product demands and resolving hierarchical non-coherence. To address this, this study proposes a novel prediction framework combining Long Short-Term Memory (LSTM) networks with the Minimum Trace Reconciliation (MinT) method. Uniquely, our model fosters intrinsic coherence by summing hidden states during the learning phase, establishing a methodological foundation that maximizes the efficacy of the subsequent MinT adjustment We evaluate this approach against nine benchmark models—integrating Multivariate Regression, SARIMA, Random Forest, and LSTM with various hierarchical forecasting strategies—using daily parcel data from Seoul (275 time series). While the Bottom-up LSTM proved most effective among benchmarks, the proposed model demonstrated superior performance, reducing Mean Absolute Error (MAE) by 17.1% at the regional level and 15.4% at the product level. These findings confirm that this stepwise approach—prioritizing intrinsic coherence before reconciliation—significantly enhances predictive accuracy and structural consistency for operational decision-making.