Traffic prediction enhances the quality of network services, yet existing methods typically require large-scale traffic data transmission, raising concerns about user privacy leakage. To address this, federated learning is adopted to preserve user privacy and reduce computational workload. However, in practical applications, traffic data from different base stations are often heterogeneous and limited in volume, which hinders the generalization ability of the resulting global model. To overcome these issues, a deep learning model, termed CALS, based on an attention mechanism within the framework of federated learning is proposed. The proposed model is trained on three publicly available datasets with non-identical distributions, enabling the model to better capture the dynamic characteristics of base station traffic. Compared to standard deep learning models such as GRU and LSTM, the proposed model reduces the mean absolute error by 9.42% and 11.1%, respectively.