As regenerative braking systems become more widespread in railways, rising attention is paid to collaborative train operations under optimized timetables to enhance regenerative braking efficiency. The effective usage of regenerative braking energy (RBE) is determined by the dynamic nature of the traction power supply network, driven by constant changes in train power and positions. Solving the power flow with multiple trains significantly, however, increases the computing time required to solve the optimization model. Most existing methods have to solve optimization problems neglecting the dynamic power flow analysis, which sacrifices the accuracy of regeneration efficiency. In order to address this challenge, we propose a data-driven model that emulates the power flow analysis and reduces the computational demands. Initially, data from both single and multitrain simulators are collected and stored in a database, from which critical information regarding train position, power, and substation power is extracted. A neural network is then used to develop a data-driven model that predicts the power of a substation in a power supply network based on train positions and powers. Case studies with Beijing Yizhuang Metro line data show that the calculation time of the data-driven model is 0.33% of the power flow simulation while keeping the accuracy above 99%. Based on this data-driven model, by optimizing train speed profile and dwell time, the energy supplied by substations can be reduced by up to 13% compared to traction optimization.