Being able to have a sufficiently rich model of underlying traffic situations to allow accurate road accident prediction is a critical challenge in a smart city. Yet, few studies took into account the complex relationships between external factors and the road network characteristics. This is particularly pertinent when using Open Street Map (OSM) data. Providing detailed information about the street network helps predictive models understand the accessibility, traffic flow, and general location characteristics. This paper investigates how to provide road accidents predictions based on building a deep learning framework supported by a knowledge graph for geospatial knowledge representation that integrates semantic correlation of multisource and heterogeneous data. In order to create knowledge representations and capture the correlations between semantic and spatiotemporal factors of a road accident, we first build a knowledge graph for accident prediction. Then we use a knowledge representation learning approach to encode both entities and relations in a low dimensional semantic space. To integrate the extracted knowledge of OSM and other traffic factors, We suggest using a spatial-temporal graph convolutional backbone network (GCN-GRU) with a fusion cell as the input.