Early prediction of milk yield per lactation of Holstein-Friesian cows in Queretaro, Mexico

Abstract

Objective: To obtain a machine learning (ML) model to predict milk yield for the same lactation and adjusted to 305 d (MY305).

Design/methodology/approach: We used a test-day (TD) database consisting of 11,892 lactation records with more than 150 days in milk (DIM), coming from 19 dairy farms located in Querétaro, Mexico. We standardized milk yield to specific DIM (5, 10, 20, 30 and 40) to obtain estimates of MY305 using ML modeling. The herd, month of birth of the cow, month of start of lactation, lactation number, number of days to three daily milking and the first two linear somatic cell scores were also incorporated as explanatory variables.

Results: The best goodness of fit was achieved with model ensembles, obtaining a deviance of 1503584 for the training data (with 80% of data chosen at random), while with 20% of the data reserved to evaluate the model’s deviance was 1576776. The relationship between observed data and MY305 predictions from the ensemble of models, had a coefficient of determination of r² = 0.79 and RMSE of 1256. In the best model (deviance of 2281420) of the 'deep learning' type, the most important variables were the milk yield at 30, 10, 5 and 20 DIM (19.9, 16.6, 16.2 and 12.8%, respectively).

Limitations on study/implications: RMSE values are high. TD databases are generated regularly and following systematic measurement procedures but not many dairy farms are represented.

Findings/conclusions: For the database examined, milk yield in the early lactation phase together with an ensemble of learning models showed adequate prediction of milk yield adjusted to 305 d.

Keywords: machine learning, somatic cells, lactation curve, test day