Cross-Generational Validation of a Feedforward Neural Network for Milk Yield Prediction in Dairy Cattle

TL;DR

A machine learning model for predicting milk yield in dairy cows remains accurate across generations and can help improve herd management when used alongside other considerations.

Contribution

The study validates a feedforward neural network model's reliability across generations of Holstein cows under the same farm conditions.

Findings

The model achieved a daily RMSE of 5.98 kg/day and a Pearson correlation of 0.64 in predicting milk yield.

Sensitivity analyses showed predicted milk yield increases with later calving ages, but these results reflect training data patterns.

The model's predictions are robust across generations but should be used alongside economic and reproductive factors.

Abstract

Accurately predicting milk production helps dairy farmers improve herd management and efficiency. In this study, we tested a previously developed machine learning model that predicts daily milk yield using genetic information and production data automatically recorded during milking. The model was applied to a new generation of cows, specifically the daughters of the animals used to develop the original model, raised under the same farm conditions. The predictions were accurate and closely matched those obtained in the original study, showing that the model remains reliable across generations. We also explored how predicted milk yield changes when the age and season of calving are modified to better understand the model’s behavior. These findings suggest that data-driven prediction tools can support dairy management decisions, as long as they are used together with economic and reproductive considerations.

Advancements in precision livestock farming and machine learning have expanded the use of data-driven approaches for milk yield forecasting. In this study, a previously developed feedforward neural network (FFNN) model using genomic breeding values, parity, days in milk, month of calving, and age at calving as predictors was validated across one generation of Holstein cows. Specifically, the model was evaluated in first-parity daughters of the animals included in the original training population. Predictive performance was assessed on 228 lactation curves comprising 67,010 daily observations using a train–cross-validation–held-out test framework. On the test set, the model achieved a daily root mean squared error (RMSE) of 5.98 kg/day, with a Pearson correlation of 0.64. Sensitivity analyses were conducted by systematically shifting calving month and age (±1 to ±4 months) while holding other predictors constant. Simulated scenarios suggested increased predicted milk yield with later calving ages; however, these results reflect the structure of the training data rather than prescriptive management recommendations. While the FFNN provides robust milk yield predictions, its practical application for calving strategy decisions should be integrated with economic and reproductive considerations. Overall, the findings support the generational robustness of FFNN-based milk yield forecasting within the studied herd.

Figures