PSX-23 Use of time series data in the prediction of beef cattle dry matter intake using machine learning

Abstract Improvements in technology that facilitate the estimation of individual animal dry matter intake (DMI) in any setting, whether in confinement or on pasture, will improve beef cattleproduction and management. Big Data approaches could further enhance DMI estimation. In this work, we improved upon our predictive models by developing variable time series features and forgoing the use of full-test Average Daily Gain, a variable that provides unwanted intake information during model training, in DMI prediction. We implemented the improved time series analysis techniques of first differences, moving averages, and seasonal decomposition to create new predictive variables from daily-level data. These new variables were used in k-fold cross-validation to improve the variable selection process and prediction accuracy of the Random Forests Regression (RFR) and Repeated Measures Random Forests (RMRF) models. Reduced models for the RFR and RMRF algorithms were identified, assessed, and compared with estimated DMI calculated using the current NASEM equation for estimation of beef cattle DMI. Data on feed intake, animal weights, and climate conditions were collected from 25 November2019 to 20 February 2020 for bulls and from 19 March 2020 to 7 May 2020 for steers. Feed intakes were recorded using Vytelle Feed Intake Nodes. Animal body weights (BW) and water intakes were measured daily using Vytelle In-Pen-Weighing Positions and metered waterers. Climate data included information on temperature, humidity, precipitation, and wind speed and were recorded at 30-min intervals. We found that seasonally-adjusted and moving averaged features generally outperformed first difference features. The baseline models presented here performed similarly to our previous models, despite using ‘Average Daily Gain to Date’ rather than the ‘Full Test Average Daily Gain’. The range of R2 was 0.44 to 0.69 and the range of RMSE was 0.95 to1.67 kg. Additionally, use of Average Daily Gain to Date in place of full-test Average Daily Gain improves the field deployability of our models. Lastly, while the demonstrated models can both over- or under- predict daily DMI, both RFR and RMRF showed more accurate predictive ability than the current NASEM equation. This study further demonstrates the potential of RFR and RMRF models DMI prediction of beef cattle. Future research will focus on including a wider range of cattle breeds, climatic conditions, and geographic locations, as well as assessment of suitability of RFR and RMRF for DMI prediction of cattle on pasture.