Effect of parity weighting on milk production forecast models

Highlights

• The NARX model and Ali-Schaeffer model were compared at individual cow level.

• Six input treatments including parity weight combinations were tested and compared.

• The NARX Model was more accurate than the Ali and Schaeffer model.

• The effectiveness of the parity weight treatment varied between cow groups.

• Parity weight trends were a determining factor in the success of the treatments.

Abstract

The objectives of this study were to compare the prediction accuracy of two milk prediction models at the individual cow level and to develop, compare and evaluate six input data preprocessing treatments designed to factor parity information into the milk prediction model configuration process. The two models were a nonlinear auto-regressive model with exogenous input and a polynomial curve fitting model. These were tested using six different parity data input treatments. Different combinations of static parity weight, dynamic parity weight and removal of the first lactation data were selected as input treatments. Lactation data from 39 individual cows were extracted from a sample herd of pasture-based Holstein-Friesian cattle located in the south of Ireland and situated in close proximity. The models were trained using three years of historical milk production data and were employed for the prediction of the total daily milk yield of the fourth lactation for each individual cow using a 305-day forecast horizon. The nonlinear auto-regressive model with exogenous input was found to provide higher prediction accuracy than the polynomial curve fitting model for individual cows using each input treatment. An improvement in forecast accuracy was observed in 62% of test cows (24 of 39). However, on average across the entire population, only part of the treatments delivered an increase in accuracy and the success rate varied between test groups. Prediction performance was strongly influenced by the cows' historical milk production relative to parity and also the prediction year. These results highlighted the importance of examining the accuracy of milk prediction models and model training strategies across multiple time horizons. Removal of the first lactation and applying static parity weight were shown to be the two most successful input treatments. The results showed that historical parity weighting trends had a substantial effect on the success rate of the treatments for both milk production forecast models.

Introduction

Accurately predicting the milk production of a dairy cow throughout lactation is an important factor in milk production forecasting. Precisely recording large volumes of detailed data pertaining to each cow is not practical on a commercial farm. However, variables such as daily milk yield (DMY), days in milk (DIM), and parity are known on most modern dairy farms. The effect of parity on dairy cow milk yield has been presented in several previous studies and the corresponding findings are in consensus. Parity has a significant effect on milk yield (Collins-Lusweti, 1991, Rémond et al., 1997, Silvestre et al., 2009, Ríos-Utrera et al., 2013, Storli et al., 2014, Otwinowska-Mindur and Ptak, 2016) and DIM at peak yield varies in respect to parity (Rekik et al., 2003). The profile of the first lactation curve is not consistent with subsequent lactations. Total and peak milk production of dairy cows in the first parity is lower than those of cows in the second parity and the third parity (Hansen et al., 2006, Stanton et al., 1992, Tekerli et al., 2000). The highest total yield is typically presented in the third and subsequent parities (Dematawewa et al., 2007, Friggens et al., 1999, Rekik et al., 2003, Ríos-Utrera et al., 2013) and the first lactation has a slightly delayed peak DIM and lower peak yield (Jamrozik et al., 1998, López et al., 2015). The profile of the first lactation has been shown to introduce difficulties for lactation curve fitting models, in comparison with the second and later parities (Guo and Swalve, 1995). Hence, first lactation is substantially different in profile and magnitude of yield in comparison to the second, third and later parities which display similar lactation profiles. Hence in this study, milk yield of the first lactation was tested as a treatment parameter of the control groups to check whether removing the first lactation would improve the model prediction accuracy.

Many studies have been undertaken for the purpose of describing a lactation curve at the herd and individual cow level including: curve fitting models, regression models and auto-regressive models (Ali and Schaeffer, 1987, Green and Silverman, 1993, Lacroix et al., 1995, Murphy et al., 2014, Durón-Benítez and Huang, 2016, Bangar and Verma, 2017, Salehi et al., 1998, Sharma et al., 2007, Sikka, 1950, Wilmink, 1987, Wood, 1967). The Ali and Schaeffer model (Ali and Schaeffer, 1987) has been shown to outperform other curve fitting models for milk production forecasting (Olori et al., 1999). Recently, new modeling techniques have been applied to milk production forecasting. The NARX (Non-linear Auto-Regressive with exogenous inputs) model (Murphy et al., 2014) has been shown to produce accurate milk prediction with high degrees of flexibility, and adaptability at the herd level (Murphy et al., 2014, Zhang et al., 2016). The Ali and Schaeffer model has been applied on the herd level before. In spite of that, these models have never been compared at individual cow level, and parity information pertaining to the individual cow has not been tested for these models. Milk yield forecasting at individual cow level could be beneficial to these applications in the dairy industry: monitoring health conditions and disease detection by monitoring individual cow milk yield, i.e. mastitis (Andersen et al., 2011); decision support for advanced milking parlors and milking machines (Thomas and DeLorenzo, 1994) and precision input for herd simulation models (Petek and Dikmen, 2006). While previous studies achieved accurate milk production predictions using large neural network models, they required a very high number of input parameters. The Model developed by Lacroix et al. (1995) required 16 input parameters such as energy fed on test day, protein fed on test day, logarithm of somatic cell count, and dry matter fed on test day and many others. Sharma et al. (2007) built a neural network trained using 12 individual cow traits such as season of birth, season of calving, period of birth, genetic group, birth weight, age at maturity, weight at maturity, period of calving, age at calving, weight at calving, peak yield and days to attain peak yield. The high number of parameters and large amounts of data pertaining to each individual cow that are required to train these models restricts their adoption on commercial dairy farms. However, time series models and curve fitting models require very few parameters for each cow. Therefore, they may be more suited to commercial applications.

There are two primary objectives of this study: 1: To compare prediction accuracy of both the Ali and Schaeffer model and the NARX model at the individual cow level. 2: To develop, compare and evaluate six input data preprocessing treatments designed to factor parity information into the model configuration process to predict individual animal milk production.