Machine Learning based Models for Fresh Produce Yield and Price Forecasting for Strawberry Fruit
Loading...
Date
2020-06-02
Authors
Okwuchi, Ifeanyi
Journal Title
Journal ISSN
Volume Title
Publisher
University of Waterloo
Abstract
Building market price forecasting models of Fresh Produce (FP) is crucial to protect retailers and consumers from highly priced FP. However, the task of forecasting FP prices is highly complex due to the very short shelf life of FP, inability to store for long term and external factors like weather and climate change. This forecasting problem has been traditionally modelled as a time series problem. Models for grain yield forecasting and other non-agricultural prices forecasting are common. However, forecasting of FP prices is recent and has not been fully explored. In this thesis, the forecasting models built to fill this void are solely machine learning based which is also a novelty.
The growth and success of deep learning, a type of machine learning algorithm, has largely been attributed to the availability of big data and high end computational power. In this thesis, work is done on building several machine learning models (both conventional and deep learning based) to predict future yield and prices of FP (price forecast of strawberries are said to be more difficult than other FP and hence is used here as the main product). The data used in building these prediction models comprises of California weather data, California strawberry yield, California strawberry farm-gate prices and a retailer purchase price data. A comparison of the various prediction models is done based on a new aggregated error measure (AGM) proposed in this thesis which combines mean absolute error, mean squared error and R^2 coefficient of determination.
The best two models are found to be an Attention CNN-LSTM (AC-LSTM) and an Attention ConvLSTM (ACV-LSTM). Different stacking ensemble techniques such as voting regressor and stacking with Support vector Regression (SVR) are then utilized to come up with the best prediction. The experiment results show that across the various examined applications, the proposed model which is a stacking ensemble of the AC-LSTM and ACV-LSTM using a linear SVR is the best performing based on the proposed aggregated error measure. To show the robustness of the proposed model, it was used also tested for predicting WTI and Brent crude oil prices and the results proved consistent with that of the FP price prediction.
Description
Keywords
machine learning, deep learning, forecasting, prediction, yield, price, ensemble, fresh produce, time series, oil price