Accurate short-term forecasting of parking-garage occupancy is critical for urban traffic management, yet individualized models often underperform at low-volume or irregular sites. To improve scalability and generalization, we propose a structured forecasting pipeline that combines behavioral clustering with random forest regression. We extract three interpretable feature sets: Summary weekday-average occupancies, Catch22 time-series descriptors, and Hyndman-style metrics (e.g., trend and seasonality). We apply k-means clustering to group garages with similar temporal patterns. Forecasting models are trained both per series and per cluster using a rolling-window evaluation. Models based on Summary features consistently outperform garage-specific baselines across key metrics, while Catch22 and Hyndman-based clusters show more modest improvements. A hybrid deployment strategy, which selects the better-performing model for each garage based on early validation folds, achieves the best overall performance and reduces RMSE by 2.26%. These results demonstrate that interpretable feature-based clustering, combined with selective deployment, supports accurate, scalable, and robust short-term forecasting in urban parking systems, especially for underperforming or volatile sites.