Time-series prediction of onion quality changes in cold storage based on long short-term memory networks

This study presents a recurrent neural network (RNN)-based model for predicting physical quality changes in onions during long-term low-temperature storage. Unlike previous studies that primarily used simple regression equations with certain time points, this model utilizes time series-based storage-environment histories. The model was designed to use only easily obtainable, real-time environmental information and storage periods as inputs, thereby enhancing practicality. The onions were stored under controlled conditions for model development, with quality factor data accumulated through regular sampling and destructive testing. Data pre-processing addressed environmental noise and time resolution issues, and bioyield strength was selected as the prediction target considering interrelationships with environmental and quality factors. For comparison, four RNN-based models (stacked RNN, stacked long short-term memory (LSTM), residual LSTM, baseline LSTM) were proposed, along with a conventional multiple linear regression (MLR) model. Given the scarcity of preverified models for agricultural product quality prediction, a grid search was employed to explore various model structures and parameter combinations. The performance was compared using the same independent test data on the retrained models based on the optimized combinations. The residual LSTM model, scored 0.0532 on the normalized test root mean-squared error (RMSE), demonstrating the best performance by predicting up to 336 h ahead. The baseline LSTM model predicted up to 504 h with an RMSE of 0.0601. In contrast, the MLR model showed complete overfitting, with a test RMSE of 0.2814, indicating the unsuitability of utilizing storage history for quality prediction. This study verifies the potential of RNNs in agricultural product storage/distribution and anticipates the utility of the developed model as a foundational technology for future post-harvest process research.