Smart railway is in the process of vigorous development based on the advancement of numerous emerging and advanced techniques, such as artificial intelligence and big data. However, a smart railway network is exceedingly time-varying and complicated, which presents great challenges to power allocation. Deep reinforcement learning is capable of effectually improving the intelligence as well as cognition of high-speed railway network, which helps to optimize the power allocation problems with both time-varying and complicated characteristics. In this article, we first provide an overview of the existing power allocation methods, including advantages, disadvantages, and complexity, as well as characteristics. Then we investigate an innovative power allocation algorithm based on multi-agent deep recurrent deterministic policy gradient (MADRDPG), which is capable of learning power decisions from past experience instead of an accurate mathematical model. Finally, numerical results indicate that the performance of the MADRDPG-based method significantly outperforms existing state-of-the-art methods in terms of spectrum efficiency and execution latency.