Food Loss Forecasting with Recurrent Neural Networks Using Gated Recurrent Units

Abstract

Food loss forecasting is critical for global food security and sustainability. This study compares Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) architectures for predicting post-harvest food loss using Malaysian agricultural time-series data. Results demonstrate that GRU outperforms LSTM across all metrics, achieving 10.7% lower RMSE, 5.0% lower MAE, and 61.8% reduction in MAPE. GRU also exhibits superior computational efficiency with 23.9% fewer parameters and 19.5% faster inference. We release a reproducible setup that can be directly extended with Attention-GRU and seasonal-trend decomposition GRU (STL-GRU) to better capture long-range and seasonal dependencies. This work provides a practical foundation for next-generation models aimed at reducing losses and strengthening resilience in agri-food systems.