An Evaluation of Machine Learning Approaches for Milk Volume Prediction in Ireland

Milk yield production strongly influences energy consumption, plant utilisation and farm revenue. The competency to identify the annual peak week and the ability forecast daily, weekly, or annual lactation curves in advance is beneficial across the agri-business chain at a management, processor, and farm gate level. The value of a milk yield prediction system depends upon how accurately it can predict inherent milking patterns and its ability to adjust to factors affecting milk volumes. This study presents a review of machine learning (ML) approaches applied to dairy-specific data and meteorological signals to predict milk volumes in Ireland at a national level. This contrasts with existing approaches that forecast for individual cow or herd levels in Ireland. The resulting model performances serve as a benchmark for any future algorithms developed. Adopting the Lewis scale, it was shown that Random Forest and K-Neighbours Regression derived "highly accurate forecasts" with an average Mean Absolute Percentage Error of 8.28 and 8.35 respectively for prediction timelines of 1-52 weeks ahead. 13 other models are shown to produce "highly accurate" week-ahead forecasts which degrade over increasing forecast horizons. This includes three existing ML milk yield forecasting algorithms to facilitate a comparison with the existing state-of-the-art.