FGT-Net: A multi-granular fuzzy fusion network for stable and explainable traffic flow forecasting

Multivariate traffic flow forecasting requires accurate multi-horizon prediction with interpretability, while maintaining stable performance under non-stationarity. Recent Gaussian linear fuzzy information granules (GLFIGs) provide compact and explainable representations for traffic forecasting, yet the commonly used “segmentation-granulation-prediction” pipeline still suffers from three key bottlenecks in multivariate and multi-granular settings: (i) heuristic and parameter-sensitive time-domain segmentation may produce unstable boundaries and spurious change points, making trend granules unreliable; (ii) granules across variables and granularities are often misaligned, where dynamic time warping (DTW)-style time warping or splitting and merging alignment can rewrite membership semantics and support structures, causing semantic drift and weakening interpretability; and (iii) granule-level forecasting tends to under-react to short-term details, degrading short-horizon performance. To address these issues, we propose FGT-Net, a multi-granular fuzzy fusion network that improves the FIG forecasting pipeline through three component-level innovations. First, dual-domain enhanced trend representation and analysis (DETRA) performs unsupervised trend discovery via frequency-time co-verification: spectral energy concentration in the frequency domain proposes intrinsic long-term and seasonal candidates, and time-domain boundary consistency validates and refines cut points to suppress spurious segmentation. Second, we develop a semantics-preserving granular alignment strategy that aligns multivariate GLFIG representations through convolution-based count alignment and similarity pooling, avoiding destructive split/merge operations and alleviating granularity mismatch. Third, a horizon-aware adaptive fusion head integrates short-term fluctuations with stable trends using exponentially decayed horizon weights, enabling responsive short-horizon forecasting while emphasizing macro-level trends for long horizons. Experiments on six real-world traffic datasets show that FGT-Net improves predictive accuracy while retaining interpretability: compared with representative baselines, root mean square error (RMSE) is reduced by 20.27% for short horizons and 12.65% for long horizons, indicating fewer large deviations and more stable forecasting errors under challenging conditions. Detailed descriptions and access information are available at https://github.com/Ronghua550/Dataset-2 .