Forecasting for Sustainable Dairy Produce: Enhanced Long-Term, Milk-Supply Forecasting Using k-NN for Data Augmentation, with Prefactual Explanations for XAI

Abstract

Accurate milk supply forecasting for the dairy sector, covering 1000 s of farms with low resolution data, is a key challenge in achieving a sustainable, precision agriculture that can improve farm management, balancing costs, energy use and environmental protection. We show that case-based reasoning (CBR) can meet this sustainability challenge, by supplementing a time series prediction model on a full-year-forecasting task. Using a dataset of three years of milk supply from Irish dairy farms (N = 2,479), we produce accurate full-year forecasts for each individual farm, by augmenting that farm’s data with data from nearest-neighboring farms, based on the similarity of their time series profiles (using Dynamic Time Warping). A study comparing four methods (Seasonal Naïve, LSTM, Prophet, ProphetNN ) showed that the method using CBR data-augmentation (ProphetNN) outperformed the other evaluated methods. We also demonstrate the utility of CBR in providing farmers with novel prefactual explanations for forecasting that could help them to realize actions that could boost future milk yields and profitability.