Abstract
Technological advances are transforming sustainable cattle farming practices. Electronic feeding systems, for instance, generate big longitudinal datasets on individual animal feed intake, offering the possibility for autonomous precision livestock systems. However, the literature still lacks a methodology that leverages these longitudinal big data to accurately predict feed intake accounting for environmental conditions. To fill this gap, we developed an AI-based framework to accurately predict feed intake of individual animals and pen-level aggregation using experimental longitudinal big data. Data from 19 experiments (>16.5M samples; 2013-2024) conducted at Nancy M. Cummings Research Extension & Education Center (Carmen, ID) feedlot facility and environmental data from AgriMet Network weather stations were used to develop two novel environmental indices: InComfort-Index, based solely on meteorological variables, showed good predictive capability for thermal comfort but had limited ability to predict feed intake; EASI-Index, a hybrid index integrating environmental variables with feed intake behavior, performed well in predicting feed intake but was less effective for thermal comfort. Using the environmental indices, Machine learning models were trained and the best-performing machine learning model (XGBoost) accuracy was RMSE = 1.38 kg/day for animal-level and only 0.14 kg/(day-animal) for pen-level. This approach provides a robust AI-based framework for predicting feed intake of individual animals and pens, with potential applications in precision management of feedlot cattle, through optimizing feed delivery, reducing feed waste, and providing climate-adaptive livestock management.
From electronic feeders, we obtained 16.5M samples that contained feed intake (FI, kg), FI duration (min), and timestamp of starting FI. Individual FI data separated by less than 20 min (dt < 20 min) were aggregated into meal data (1,1M samples), considering a maximum meal duration of 2h and maximum meal intake (MI, kg) of 10 kg. First, outliers removal used Generalized Additive Models (GAM) and substituted samples that had MI deviation from predicted (MI) greater than 3 interquantile range (IQR). The same outlier removal procedure was repeated using Nonlinear Monotonic Regression (NMR) with spline transformation. Clean data included meal data and weather data (from Corvallis, MT, weather station from 2013-2019; from Salmon, ID, weather station from 2020-2024): air temperature (Ta, °C), daily maximum Ta (Ta,max, °C), daily minimum Ta (Ta,min, °C), relative humidity (RH, %), daily maximum RH (RHmax, %), daily minimum RH (RHmin, %), mean vapor pressure (Pv, kPa), mean dew point temperature (Tp, °C), mean wind direction (Wd, °), mean wind hourly speed (Ws, m/s), wind peak gust (Wg, m/s), solar irradiation (SI, W/m2), and precipitation (Rain, mm). Using weather data, InComfort index was developed. Using clean data, EASI index was developed. Finally, daily individual animal’s FI and pen-level FI (penFI, kg; 149k samples) were predicted using machine learning algorithms (ML).