Paper Accepted to Journal of Forecasting!!!

Deep learning-based long time series forecasting (LTSF) has achieved high accuracy by effectively capturing the underlying trends, seasonality, and temporal dependencies within time series data. However, at the individual entity level, termed the low aggregation level (LAL), intermittency, irregularity, and data sparsity undermine the inductive biases of neural networks, leading to a significant decline in performance. To address the challenges of LAL-LTSF, this study introduces a similarity matching (SM)-based method that leverages historical time series patterns from analogous entities. Recognizing that the efficacy of the SM algorithm in LTSF is highly dependent on embedding quality, we present a hybrid temporal autoencoder (HTAE) architecture tailored to effectively capture LAL-specific time series features. The HTAE integrates conventional self-attention mechanisms with elements derived from a newly developed contextual multihead self-attention module, facilitating the creation of robust representations, even in noisy and intermittent LAL scenarios. The proposed approach outperformed 13 baseline models, including state-of-the-art techniques, across two real-world time series datasets. This study offers a promising solution for enhancing LAL-LTSF accuracy and has the potential to improve long-term demand forecasting and service delivery in real-world industrial applications.