Accurate and interpretable multivariate traffic forecasts are essential for congestion mitigation, energy-aware routing, and emergency dispatch in smart cities. Yet existing CNN- or Transformer-based models, while accurate, provide limited insight into the mechanisms that link spatial disruptions with downstream impacts. We propose Threshold-Bound Fuzzy Cognitive Maps (TB-FCM), a hybrid architecture that unifies three elements: (i) learnable dynamic thresholds that control causal activation among traffic variables, (ii) a multi-scale graph-attention network that aggregates adjacent and distant dependencies, and (iii) GRU gates that capture short- and long-term temporal context. This design retains the editable causal semantics of classical FCMs while handling non-stationary demand spikes and network-wide spill-overs. Experiments on three benchmarks—EXPY-TKY (1,843 links) and the Chinese expressways G4 and G50 (three sensors each)—show consistent improvements over strong graph baselines. For 60-min horizons, TB-FCM lowers RMSE by 7.2%, 6.8%, and 4.9% on the respective datasets, with parallel gains in MAE and MAPE across 10- and 30-min windows. Consequently, TB-FCM establishes a rigorously validated, practitioner-ready framework that couples state-of-the-art predictive accuracy with transparent causal insight, advancing interpretable modeling for complex multivariate spatial-temporal systems.