The agricultural Internet of Things (IoT) system is a critical component of modern smart agriculture, and its security risk assessment methods have garnered increasing attention from the industry. Current agricultural IoT security risk assessment methods primarily rely on expert judgment, introducing subjective factors that reduce the credibility of the assessment results. To address this issue, this study constructed a dataset for agricultural IoT security risk assessment based on real-world security reports. A PCARF algorithm, built on random forest principles, was proposed, incorporating ensemble learning strategies to enhance prediction accuracy. Compared to the second-best model, the proposed model demonstrated a 2.7% increase in accuracy, a 3.4% improvement in recall rate, a 3.1% rise in Area Under the Curve ( AUC), and a 7.9% boost in Matthews Correlation Coefficient (MCC). Extensive comparative experiments showed that the proposed model outperforms others in prediction accuracy and robustness.