智慧酪農業—應用機器學習於牛乳產量預測

Abstract

牛奶和乳製品在人類社會中扮演著重要的角色。然而，酪農業目前正面臨許多挑戰，例如需求增加、勞動力短缺和氣候變遷等。為了應對這些問題，牛奶生產系統有待改進。智慧酪農業將物聯網（IoT）、大數據分析、人工智慧（AI）等技術應用於牧場，為這些挑戰提供創新的解決方案。智慧酪農業包含四個要素：環境控制、個體動物資訊、自動飼養系統和牧場管理，涉及熱緊迫控制、個體動物辨識、動物健康監測、自動榨乳系統、精準餵飼和精準育種等主題。乳量預測是智慧酪農業中的關鍵技術，它可以幫助農民預測農場收入、監測動物健康並優化育種選擇。本研究的目標是建立機器學習模型來預測乳量，並評估乳量與其他特徵之間的關係。本研究的資料來源為台灣乳牛群性能改良（DHI）計畫的資料庫，包含了33,185筆自2013年至2018年、共1,818頭荷蘭牛的榨乳記錄。經過數據清理和分析後，本研究使用不同的機器學習演算法來建立預測模型，包括支持向量機（SVM）、隨機森林（random forest）和XGBoost。研究結果顯示，最終的XGBoost模型在三種演算法中表現最佳，其對未來乳量預測的準確率達到76.33%。此外，本研究還發現影響乳量的重要因素為過去平均產量、泌乳天數、與前一胎間隔和月齡。這些結果對乳量預測的發展具有重要意義。

Milk and dairy products play a significant role in human society. However, milk production is currently facing several challenges, such as an increasing demand, labor shortages, and climate changes. To address these issues, improved milk production systems are required. Smart dairy farming incorporates the Internet of Things (IoT), big data analytics, artificial intelligence (AI), and other technologies in dairy farms, offering innovative solutions to these challenges. Smart dairy farming is made up of four elements: environmental control, single animal information, automatic rearing systems, and management, involving topics like heat stress control, individual animal identification, animal health monitoring, automatic milking systems, precision feeding, and precision breeding. One crucial aspect of smart dairy farming is milk yield prediction, which allows farmers to get a projection of farm income, monitor animal health, and optimize breeding selection. The objectives of this study were to develop machine learning models for milk yield prediction and to assess the relationship between milk yield and other features. The data used in this study were obtained from Dairy Herd Improvement (DHI) database in Taiwan, which included 33,185 milking records from 2013 to 2018 involving 1,818 Holstein cattle. After data cleaning and analysis, prediction models were built using various machine learning algorithms, including support vector machine (SVM), random forest, and extreme gradient boosting machine (XGBoost). The findings of this study demonstrated that the final XGBoost model exhibited the highest performance among the three algorithms, attaining an impressive 76.33% accuracy in predicting future milk yield. Moreover, it was revealed that specific factors, including average yield, days of lactation, calving interval, and age in months, significantly influenced milk yield. These insights serve as valuable contributions to the advancement of milk yield prediction.