Hourly counts from bicycle and pedestrian sensors are increasingly important for daily operations, safety monitoring, and near-term planning, yet these series are sparse, highly zero-inflated, and sensitive to time-of-day context. This study develops a text-augmented Transformer-GNN workflow for short-term active-mobility forecasting on 2024 California hourly bicycle and pedestrian counts. Directional movements are aggregated to mode-location node series, a 5-nearest-neighbor geographic graph is built from site coordinates, and event texts are generated from calendar and operating-context tags such as overnight low demand, weekday commute access, weekend recreation, and holiday schedules. The forecasting model combines scaled dot-product temporal attention, graph-diffused count values, structured text tags, and a validation-calibrated decoder. Baselines include persistence, daily and weekly seasonal forecasts, temporal Ridge regression, graph Ridge regression, and an Elastic Net text model. Experiments use a chronological split with January-August for fitting, September for calibration, and October-December for testing at 1, 3, and 6 hour horizons. Across 46 high-coverage nodes and 101,568 node-hours in the test period, the Text-Transformer-GNN achieves the lowest MAE at all horizons: 2.393 for 1 hour, 2.613 for 3 hours, and 2.656 for 6 hours. Relative to the strongest seasonal baseline, these values reduce MAE by 12.5%, 14.7%, and 13.3%, respectively. Error analysis shows that high-volume pedestrian sites dominate residual risk, while event-tagged explanations clarify whether forecasts are driven by recent persistence, prior-day seasonality, graph-neighbor demand, or text-defined operating context.

Keywords: Active transportation counts; bicycle and pedestrian forecasting; Transformer; graph neural network; event text; traffic operations; California mobility data; explainable forecasting.