PSVIII-27 Beef Cattle Body Weight Forecast Monitored on Pasture Through Artificial Intelligence

Abstract

Abstract The use of artificial intelligence through forecasting models for the weight of beef cattle allows decision-making regarding the production chain, which makes it more efficient and respects social, environmental and economic aspects, with ever-increasing predictions. Approaches adopting separately stochastic and deterministic components were adopted. Thus, this work, based on deterministic dynamic systems, aims to forecast the body weight of cattle simultaneously with the variables temperature, atmospheric pressure and global radiation, for the first time being monitored on pasture. Data were collected on a commercial farm located in Britânia/GO/Brazil, 158 male animals of crossbread origin,kept in 101 hectares of pasture, in a rotational system, receiving 0.3% BW day-1 of protein supplement and water ad libitum were weighed. Based on Takens' embedding theorem, it was possible to represent the phase spaces and facilitate data entry into the regression model. The models used were Multi-Layer Perceptron (MLP), eXtreme Gradient Boosting (XGBoost), LightGBM, Support Vector Machine (SVM), Long Short-Term Memory (LSTM) and Echo State Network (ESN). To fit the models, the dataset was separated into 110 animals for K-fold cross-validation with 5 subsets and 48 animals for validation. Bayesian optimization was implemented to find the best hyperparameter values for each model. Three scenarios were evaluated: 7, 14 and 28 days of forecast horizon (fp), respectively. The evaluation metric used was the mean squared error (MSE), which verifies the accuracy of the models. As result, we verified according to Table 1 that the SVM presented the least MSE values for the 3 scenarios. After this step, the best model was selected, and the model was adjusted with the 110 animals and validated with the validation data. The result obtained from the MSE was 0.000071 kg². 0.000176 kg² and 0.000478 kg² for 7, 14 and 28 days of fh, respectively. According to Figure 1, the predicted values in relation to the observed values showed good results for the 3 scenarios; however, for the scenarios of 14 and 28 days, a small overestimation was observed for the body weights with the greatest values. The adjusted SMV model demonstrated that it was possible to predict meat production with deterministic models. Thus, the use of artificial intelligence methods is efficient to assist in decision making on the beef cattle production chain.