Accurate forecasting of the State of Health (SOH) in electric vehicle (EV) batteries is critical for ensuring operational safety and extending battery lifespan. Existing centralized machine learning methods pose significant limitations, including data privacy concerns, communication overhead, and poor adaptability to heterogeneous EV data. In this paper, we propose a privacy-preserving Federated Learning (FL) framework integrated with the Temporal Fusion Transformer (TFT) – which incorporates an encoder-decoder long short-term memory (LSTM) backbone – for SOH prediction using multivariate time-series data. The framework allows multiple EV clients to train local TFT models on standardized telemetry data – such as voltage, temperature, and state of charge – while sharing only encrypted model updates with a central server. The server performs weighted aggregation of parameters and redistributes the global model across 50 communication rounds. Our approach leverages attention mechanisms and gating structures within the TFT architecture to capture temporal dependencies and enhance interpretability. The model is evaluated using error metrics such as MSE, MAE, RMSE, and MAPE. Results demonstrate that the FL-TFT framework achieves superior predictive accuracy (MSE of 0.139, MAE of 0.301, RMSE of 0.373 and MAPE of 0.298), representing a 23.6 % improvement in MSE over conventional federated approaches, with efficient convergence in just 30 communication rounds compared to 50 rounds required by baseline models (SGD, RF, DT, SVR), – making it a scalable and privacy-aware solution for battery health diagnostics across distributed EV datasets.