A Review of Attention-Enhanced GRU Models with STL Decomposition for Food Loss Forecasting

Forecasting food loss with high accuracy is crucial for improving global food security, optimising supply chains, and supporting sustainability goals. However, conventional time series models and standard deep learning techniques, including recurrent neural networks (RNNs), often fall short in handling the irregularity, seasonality, and complexity inherent in food loss data. While Gated Recurrent Units (GRUs) offer advantages over traditional RNNs, such as mitigating vanishing gradients, they still face limitations in modelling long-range dependencies and noisy sequences. This paper reviews recent advancements aimed at overcoming these challenges by enhancing GRU-based models with attention mechanisms and seasonal-trend decomposition using Loess (STL). Evidence from related domains shows that attention mechanisms improve the capture of long-term dependencies and interpretability, while STL decomposition strengthens stability and accuracy by isolating seasonal and trend components. Hybrid GRU models that combine both approaches consistently outperform standalone methods, highlighting their promise for robust and interpretable forecasting. Though underexplored in the context of food loss, this paper identifies the research gap and advocates for domain-specific GRU–attention–STL architectures, offering a foundation for future empirical work to enable timely interventions and foster resilient, data-driven food systems.