Solving high-speed rolling stock planning with an improved LSTM-PPO reinforcement learning algorithm

Rolling stock planning (RSP) is a key component of railway operations, directly affecting transport efficiency and operating cost. To address the high-dimensional decision space, strong temporal dependencies and multiple operational constraints in RSP, this paper proposes a hybrid optimisation approach combining deep reinforcement learning (DRL) with heuristic search. The problem is formulated as a Markov decision process (MDP), where states describe remaining train services, circulation status and depot resources, and actions represent train assignment and circulation construction. A long short-term memory (LSTM)-enhanced proximal policy optimisation (PPO) algorithm is developed to capture temporal dependencies, improve exploration and stabilise convergence. In addition, simulated annealing (SA) is introduced as a local search module to refine policy-generated solutions and help escape local optima. Experiments on real-world instances from the China high-speed rail network show that the proposed method outperforms conventional DRL methods and a commercial solver in rolling stock usage and computational efficiency.