Artificial Intelligence-Based Tools for Fresh Produce Procurement Price Decisions in Canadian Distribution Centres

Market price forecasting models for Fresh Produce (FP) are crucial to protect retailers and consumers from highly priced FP. However, utilization of the data for forecasting is obstructed by the occurrence of missing values and limited by conventional machine learning or simple deep learning models with moderate performance. Therefore, it is imperative to develop imputation models for those missing instances thereby enabling effective forecasting. Moreover, most of the current forecasting models are for yield forecasting and they are either conventional machine learning or simple deep learning models with moderate performance. These models are usually univariate, i.e. do not consider many of the factors that affect the yield, or multivariate models that rely on historical station-based data that are not readily available for many croplands around the globe. In addition, fresh produce price prediction is rare to find and the models are univariate considering previous prices as input. All available yield and price forecasting models tend to be crop dependent and most of the work consider one crop per model leaving out any framework for generalization to other crops.

Fahkri Karray’s research team, which includes WICI core member Dawn Parker, proposes a deep learning (DL) application to be used for decision support by procurers, which forecasts the fresh produce yield and prices for any county using extensive DL models to enhance performance. Hence station-based data, imputed by DL imputation models to avoid missing data, as well as satellite images are used to overcome the scarcity of data in some geographic areas. Compound deep learning forecasting models such as SeriesNet with Gated Recurrent Unit (GRU) and Convolutional Neural Network LSTM with Attention layer (Att-CNN-LSTM) are trained and tested with both: station-based parameters and the same parameters extracted from satellite images. An averaging ensemble of both outputs is built for better forecasting performance. Strawberry is chosen as initial FP and the microeconomics theory is used to estimate the supply curve of strawberries based on Loblaws procurement dataset. A generalization framework should be provided to transfer the learning to similar crops or retrain for dissimilar ones. The application should be tested with complete functional and usability test scenarios to ensure accuracy and user friendliness. This project is supported by Loblaws and a Natural Sciences and Engineering Research Council (NSERC) Collaborative Research and Development Grant ($600,000).