Summary Accurate outbound volume (OV) prediction of refined oil depots is vital for efficient energy distribution and logistics management. However, refined oil depots typically store multiple types of oil products, the OVs of which exhibit strong correlations and significant time-varying characteristics, with sparse data availability in some depots. Existing methods fail to deliver accurate predictions as they overlook inherent interdependencies and latent temporal dynamics among product volumes, and struggle with effective knowledge transfer under sparse data conditions. To address these challenges, we propose an innovative multitask learning (MTL) framework. First, a multivariate data decomposition module incorporating fast Fourier transform (FFT) and wavelet decomposition (WD) to optimize multivariate variational mode decomposition (MVMD) is designed, enhancing the capability of extracting multidimensional time-varying properties from OV data. Next, an improved multigate mixture-of-experts (IMMOE) network is introduced to capture dynamic patterns, refine critical temporal features, and effectively model complex dependencies across various refined products, thus achieving accurate OV predictions. Finally, a progressive unfreezing transfer learning strategy is proposed to mitigate prediction accuracy degradation under limited data conditions. Experimental results on real-world data sets demonstrate that the proposed framework achieves more than 50% reduction in error across all evaluation metrics compared with advanced baseline models. Additionally, the transfer learning strategy achieves R2 and Nash-Sutcliffe efficiency values exceeding 0.9 in data-scarce depots. Sensitivity analysis confirms the contributions of each component and emphasizes the more significant importance of multidimensional time-varying feature extraction in performance improvement of OV prediction compared with the temporal feature extraction module.
Li et al. (Thu,) studied this question.