The food system is currently under stress at all levels, from food service providers to farms. This requires that the system be robust and flexible enough to allow for rapid changes due to rapidly changing consumption patterns, to avoid creating large imbalances in the system that end up generating food shortages in some places and food waste in others. Only by reducing these imbalances the food system can develop in a sustainable way, ensuring food security without unnecessary consumption of resources. Indeed, current population and consumption growth trajectories further increase the importance of finding solutions to meet food demand in a sustainable way. This contribution focuses on the sustainable management of food resources in Sweden, with specific reference to food produced and consumed in public service canteens. Since food production can be considered a complex process where uncertainty is relevant, mainly due to stochastic supply and demand and the variability of raw materials and ingredients, we place our analysis into statistical forecasting framework using micro-data at kitchen level to define food canteen management models and use them as a tool for efficient food planning and management. More specifically, we explore the potential of machine learning-based models to rationalize food procurement and obtain accurate predictions of food production planning in Swedish public-school canteens, thereby improving organization and food management and ultimately reducing food waste. The availability of daily time-series data at the kitchen level allow us to validate training and testing data as well as estimate future values and trends in order to provide management bodies at the school level at a broader perspective policy makers and stakeholders with an effective tool to plan the right amount of food to be served and to adjust staffing levels, so that food can be prepared and served efficiently and the optimization of economic and environmental resources is achieved. We test different machine learning methods and in particular neural networks, Poisson auto-regressive models and random forests while comparing them in terms of prediction accuracy. Lastly, with the aim of evaluating the practical implications and the impact of our models and their usefulness in reducing the economic impacts related to food waste, we estimate the economic costs and the associated savings for the individual kitchen in three different situations non-design (no-plan), actual and forecasted (model assisted) scenarios.
