Methods of predicting milk yield in dairy cows—Predictive capabilities of Wood's lactation curve and artificial neural networks (ANNs)

The study is focused on the capability of artificial neural networks to forecast milk yield for both full and standardised lactations. We used a dataset of 108,931 daily milk yields (dataset A) collected from three lactations of dairy cows managed in a production farm. Using the actual data on daily milk yields and the data recorded on official milk recording test days, a number of neural networks were designed and parameters of Wood's model were estimated. The quality of each network and regression model was measured using coefficients of determination, relative approximation errors (RAE), and root mean square errors (RMS). In order to test the prognostic parameters of the models, we randomly selected a subset of cows from the studied population, which produced in a dataset of 28,576 daily yields (dataset B). For those cows, daily and lactation yield forecasts were generated, which were next compared with their actual (observed) yield records and with the yields calculated by SYMLEK (ZETO Olsztyn Sp. z o.o., www.zeto.olsztyn.pl). The results have shown that the quality parameters of the designed neural networks were better than those of the regression model, for both the daily yields and test-day data (higher coefficients of determination and lower RAE and RMS). The prognostic parameters estimated for the forecasts of the neural networks were characterised by lower errors of prediction for both the daily yields and test-day data and exhibited higher coefficients of correlation between the predicted and the actual data (or the yields produced by SYMLEK). The predictions by the neural networks were more accurate than those by Wood's models. Furthermore, the predictions by both analysed models were closer to reality than the values estimated with the SYMLEK system. Application of neural networks does not require the data meeting the assumptions that must otherwise be met in a regression model. Large datasets are not needed to design a quite reliable neural network and, what is more, it is much easier to work with such a model than with a regression model.