Abstract

Fresh produce (FP) is one of the most important items traded in most retail outlets due to its nutritional benefits. However, the downside of these products is that their quality depreciates over time which may mean losses for the retailers. In a bid to curb profit decline incurred by product waste, retailers have managed to leverage traditional time-series forecasting to determine customers’ demand. This approach, however, is limited because it cannot capture the non-linearity in customers’ purchasing patterns, which makes deep learning techniques a viable alternative.

The fresh produce supply chain network (FPSC) is complex and this thesis only focuses on the post-harvest marketing operations of the network by leveraging deep learning (DL) models to determine the likely demand of FP and, specifically, strawberry because of its difficulty in modelling.

These demands forecast technique intricacies were explored systematically premised on the established characteristics of two major deep learning models: the Long Short-Term Memory (LSTM) neural network - which uses the temporal relationship of data to make predictions and the Convolutional Neural Network with LSTM (CNN-LSTM) - which uses the spatio-temporal relationship of data to predict.

Furthermore, a self-attention layer was added to the LSTM model and the LSTM portion of the CNN-LSTM in a bid to optimize them before deciding on the better model. To further strengthen the decision process, three key things were factored in determining the models’ (both the optimized and non-optimized) ability: their respective error levels on the validation dataset using the mean absolute error as the metric, predictive ability with the test data using the coefficient of determination (R2-score) as the metrics and forecasting (across 28 weeks horizon) skill using the R2-score.

The outcome revealed the optimized version of LSTM, that is the self-attention LSTM outperforms other models in all ramifications. Although not part of the decision process, a two end of an argument was explored to observe the findings of a manual hyperparameter optimization or an automatic approach using the Bayesian optimization technique. This was done for the LSTM model alone

Evaluation of Deep Learning Models for Forecasting Fresh Produce Demand