Univariate and multivariate time-series methods to forecast dairy income

Vahid Naghashi, Gabriel Dallago, Abdoulaye Banire Diallo, Mounir Boukadoum

Keywords: Univariate time-series, Multivariate time-series, Dairy income, Attention mechanism

Abstract:

Forecasting the income from milk sales can be addressed as a time-series problem since the sequence of multiple dairy attributes during lactation cycles are inter-related and temporally dependent. In this paper, we provide a framework to forecast the income from milk sales during the third lactation of the dairy cows based on dairy attributes recorded through the first and second lactation. We modeled the problem as univariate and multivariate time-series predictions. We propose several state-of-the-art implementations with ARIMA, N-BEATS, transformer and an original method, MuMu+attention, that combines Long-Short Term Memory neural network and attention mechanism to capture the temporal dependencies.
To benchmark the implemented methods, we curated data from 147,749 dairy cows from 5,844 Canadian herds. The monthly income from milk sales ($CAD) measured at each cow during their third lactation was treated as the prediction target. The dataset was composed of dairy attributes of milk quality, production, season, year, and health, recorded over the first and second lactation of the dairy cows. The results highlighted that most of the methods can achieve relative good performance with the best prediction accuracy obtained by MuMu+attention. MuMu+attention results were 43% better over the classic ARIMA model. By forecasting the income from milk sales, our model could help farmers to early identify less profitable animals and better allocate resources.