Going zero waste in canteens: Exploring food demand using data analytics

Abstract

Nowadays, almost all of the catering service's food demand management, including its quantitative forecasting, is based either on intuitive managers' guesses or through modeling customers' behavior only as a function of time, which in turn may arise problems such as food menus' underestimation or overestimation, as the latter leads to food waste. Therefore, in order to reduce such waste arising from mismanagement, this paper aims to describe a system capable of, under several circumstances, predicting daily food demand - number of dishes - for a given menu. This system will be firstly designed taking into account the surrounding environment of the Faculty of Engineering of the University of Porto's (FEUP) canteen, from which characteristic factors, influencing food consumption, can emerge. Therefore, factors such as weather conditions, holidays, students' timetable, are included in the model proposed. This study explores the use of advanced data mining techniques - Random Forests (RFs), Support Vector Regression (SVRs) and Artificial Neural Networks (ANNs). In this work, models were built for each type of dish - meat, fish and vegetarian - in order to predict their daily demand. Such models reached a mean absolute error (MAE) - difference between observed and predicted values - around 50 dishes for meat, 30 dishes for fish and 12 dishes for vegetarian. When comparing such results to the effective waste verified each month, it is possible to state that this system fulfills its main purpose, reducing food waste.