As electric vehicle (EV) adoption increases, the demand for efficientand reliable charging infrastructure becomes increasingly important. Accurately forecasting charging station occupancy is useful for optimizing the use of existing infrastructure and improving the overalluser experience. This study utilizes data from ChargeFinder, focusing on Swedish charging stations, and applies various data processing techniques to manage missing and intermittent data, as well ascleaning the data to produce a robust input for the model. To predict occupancy, this thesis implements a Spatio-TemporalGraph Convolutional Network (ST-GCN) that incorporates a graphconvolutional network (GCN) as the spatial component and a Transformer model as the time series component. Additionally, the thesisintroduces the Bridged Attention Module (BAM), an adapter-basedtransfer learning method designed to facilitate parameter-efficientfine-tuning across different city networks, even with limited data. Experimental results, conducted on data provided by ChargeFinder, revealed that the ST-GCN with the incorporated Transformer component produced the best overall performance, highlighting the strengthof this architecture for predicting EV charging station occupancy. Furthermore, the results demonstrate the effectiveness of the BAM adapter, which not only achieved competitive Mean Squared Error (MSE) performance but also outperformed baseline models in several scenarioswhile using a small number of trainable parameters. These findings underscore the potential of integrating spatial and temporal components into the forecast model with adaptive transfer learning techniques to enhance the scalability and reliability of EV charging infrastructure management.