Prediction model for dairy cow feed consumption status based on machine learning

Feed is essential to provide nutrients to dairy cows. An accurate prediction of feed consumption can greatly contribute to ensuring the health of dairy cows and improving production efficiency. However, the feed consumption status often exhibits nonlinear and non-stationary patterns, resulting in low prediction accuracy. In this study, based on the empirical mode decomposition (EMD) and long short-term memory (LSTM), a prediction model of feed consumption was proposed to combine the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), random forest (RF), and improved LSTM (ILSTM), namely ICEEMDAN-RF-ILSTM. Among them, ILSTM, an enhancement of LSTM, was used to adjust the output range of the forgetting gate to retain more data features, thereby improving feature learning. This model decomposed the original data into multiple relatively stationary components. Then, the different features of each component were considered to predict. The first stage was data decomposition. Since ICEEMDAN can decompose nonlinear and nonstationary time sequences into several relatively stationary components, the model decomposed the original data into multiple intrinsic mode function (IMF) components using ICEEMDAN. Each component was arranged from the high to the low frequency. The component with the highest frequency showed complex fluctuation patterns, whereas, the component with the lowest frequency represented the overall change trend of the data, regarded as the trend component. The remaining components reflected the periodic pattern of the data, regarded as the periodic components. The second stage was data prediction for the different components. RF, an ensemble learning method, was used to predict the component with the highest frequency, benefiting from its ability to construct and integrate multiple models. Additionally, the robustness and generalization of the RF model were improved by randomly selecting features. Meanwhile, ILSTM was used to predict the periodic and trend pattern components. The final stage was data integration. The predictions were summed from all components for the final prediction. The model was trained and tested on a self-built feed dataset. The results show that ICEEMDAN-RF-ILSTM achieved high accuracy with the coefficient of determination (R²), mean absolute percentage error (MAPE), and root means square error (RMSE) indicators of 0.993, 2.576%, and 0.596%, respectively, indicating that the methods proposed can effectively predict feed consumption status. At the same time, its performance was better than ICEEMDAN-LSTM and other mainstream models. This research also confirms that LSTM’s learning ability was enhanced by adjusting the output value range of the forgetting gate, in order to retain more features of the data. Meanwhile, by decomposing non-stationary raw data into multiple relatively stationary components, the prediction accuracy was improved. Moreover, by considering the different features of each component and using different methods to predict different components, the prediction accuracy can be further improved. This finding can provide a practical approach to assessing feed consumption, aiding in informed decision-making for intelligent animal husbandry.