This paper demonstrates the feasibility of using 5G Synchronization Signal Block (SSB) for tram localization in the context of Autonomous Tram (AT) systems. We develop a comprehensive MATLAB-based Urban Macrocell (UMa) simulation framework that generates synthetic SSB signals under realistic urban channel conditions, enabling systematic evaluation of power-feature-based positioning algorithms. Through extensive experiments, we show that a deep learning approach based on Long Short-Term Memory (LSTM) networks, augmented with cross-attention mechanisms and Kalman Filter (KF), significantly outperforms traditional Round Trip Time (RTT)-based Least Squares (LS) approach, reducing the mean localization error by approximately 87%, from 59.79 m to 7.83 m. The combination of simulation-driven evaluation, comparative analysis against conventional methods, and incorporation of temporal and statistical modeling provides a solid methodological and performance foundation for subsequent experimental and practical deployments.