Medium-term forecasts of electric vehicle charging-station utilization, from several months up to one year ahead, are essential for infrastructure planning and operation. They support decisions such as defining site power limits for dynamic load management and sizing local battery storage. However, station-level forecasts are challenging when only limited local observations are available and when charging behavior is noisy, site-specific, and strongly influenced by local context. This work proposes a multi-site transfer learning approach for forecasting the future power utilization distribution, energy, and maximum power of charging stations. The approach combines short observation windows from the target station with extended station metadata, including station type, geographic context, and proximity-based descriptors. We benchmark feedforward Neural Networks (NN), XGBoost, the tabular foundation model