Toward Accurate Food Loss Forecasting: GRU Models with Dynamic Time Warping Techniques

Goh, Wei Zheng (2026) Toward Accurate Food Loss Forecasting: GRU Models with Dynamic Time Warping Techniques. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text RDS_Goh Wei Zheng_Fulltext.pdf Restricted to Registered users only Download (3MB)

Abstract

Food loss remains a critical global issue with wide-ranging economic, environmental, and social consequences, particularly within agricultural supply chains where inefficiencies in handling, storage, and distribution lead to substantial wastage. In Malaysia, this challenge threatens food security and resource sustainability, underscoring the need for data-driven solutions. The purpose of this project is to develop a forecasting framework capable of predicting food loss with higher accuracy, thereby providing insights to support better planning, resource management, and policy interventions. The scope of the project covers Malaysia’s agricultural sector from 2013 to 2022, integrating historical production statistics, trade flows, and supply utilisation accounts. Data from the Food and Agriculture Organization (FAO) and the Department of Statistics Malaysia (DOSM) are combined to create a comprehensive dataset encompassing production volumes, imports, exports, and utilisation flows. To enrich the dataset, derived indicators such as per-capita availability, loss ratios, the Self-Sufficiency Ratio (SSR), and the Import Dependency Ratio (IDR) are introduced. The methodology employs a Gated Recurrent Unit (GRU) neural network to capture temporal dependencies, enhanced by Dynamic Time Warping (DTW) for preprocessing to align misaligned and noisy time series. Logarithmic transformation, Min-Max scaling, Robust scaling, and standardisation are applied to ensure comparability across features with different ranges. The model is trained using the Mean Absolute Error (MAE) loss function and optimised with Adam for efficient learning. Testing is conducted using performance metrics such as MAE and Root Mean Squared Error (RMSE), alongside time series cross-validation and walk-forward validation to assess robustness. Results show that the GRU–DTW framework significantly improves forecasting accuracy compared to baseline models, demonstrating its effectiveness in handling real-world agricultural data. The project concludes that this approach is a valuable step towards forecasting food loss, with potential for refinement and scaling to support broader applications in food system resilience.