Hazardous chemical accidents (HCAs) generally stem from multiple interrelated factors, such as equipment failure and inadequate supervision by relevant departments. The relationships between these factors are complex and variable, making accurate modeling challenging. Traditional methods for HCA prediction are insufficient in data utilization and complex relationship modeling. The integration of knowledge graph (KG) and artificial intelligence (AI) technologies can significantly enhance prediction efficiency, accuracy, and intelligence. A knowledge graph-based method for hazardous chemical accident prediction (KG-HCAP) is proposed, leveraging KG to capture and represent the multidimensional risk factors in accidents, thus transforming accident prediction into a semi-inductive link prediction (LP) problem. The KG-HCAP model is designed by combining the structural and type information of entities to strengthen their semantic representations, allowing the model to better distinguish the differences between risk factors across different dimensions. The Estimator and Reducer modules are employed to calculate the vector representations of out-of-knowledge-graph (OOKG) entities based on auxiliary information, achieving accident vector representation without retraining, ensuring prediction efficiency. The proposed method is validated using domestic and international accident investigation reports on hazardous chemical explosions, leakages, fires, and poisoning and suffocation. According to the results, our method achieves an accuracy of 91.7% in predicting the types of accidents and 56.3% in predicting their severity. This approach provides new methods and insights for HCA prediction.