Accurate forecasting of electric vehicle (EV) charging demand is essential for smart charging and efficient infrastructure use - particularly in hybrid environments with public access and shift-based internal users. This study introduces a dual-level forecasting framework: we propose a Hybrid Gated Recurrent Units (GRU) model for day-ahead station-level occupancy prediction and evaluate real-time EV-level behavior forecasting using LightGBM (LGBM) and a session similarity (SIMs) approach, enabling both strategic planning and realtime decision-making for efficient EV charging management. At the station level, our model predicts multistep charging station occupancy using behavioral and environmental features. At the EV-level, we forecast plug-in duration and energy consumption based on user-specific trends. Forecasting performance is assessed using both standard and behavior-sensitive metrics to evaluate reliability under real-world conditions. Results show that average metrics often mask behavioral variability and edge-case errors. Our analysis highlights that forecasting accuracy depends more on contextual data quality than on model complexity.