Predicting　traffic　flow　effectively　alleviates　congestion.　However,　traditional　methods　tend　to　rely　solely　on　historical　traffic　flow　data,　overlooking　the　correlation　between　multimodal　traffic　data,　such　as　speed　and　occupancy　collected　by　sensors　placed　on　the　road.　This　limitation　results　in　low　tolerance　for　abnormal　situations.　Moreover,　the　decentralization　of　multimodal　data　on　edge　devices　may　pose　data　anomalies　or　partial　modal　missing　due　to　equipment　damage　or　absence.　To　address　these　challenges,　we　propose　a　Block-Term　tensor　decomposition-based　multimodal　data　feature　fusion　algorithm　for　traffic　prediction.　This　approach　enhances　the　accuracy　and　robustness　of　traffic　flow　prediction　by　considering　correlations　between　various　modal　data,　such　as　speed　and　occupancy　rate.　In　response　to　the　issues　of　scattered　multimodal　data　anomalies　and　missing　data　on　edge　devices,　and　to　ensure　address　privacy　and　security　issues,　we　employ　federated　learning　methods　to　achieve　adaptive　extraction　and　fusion　of　multi-modal　data　at　the　edges.　Our　method　is　tested　on　a　real　highway　dataset,　demonstrating　superior　prediction　performance　and　robustness　compared　to　traditional　methods,　particularly　in　the　context　of　data　anomalies　or　missing　modalities.