A Reinforcement Learning and Unsupervised Clustering-Based Method for Automated Driving Cycle Construction for Fuel Cell Light-Duty Trucks

Addressing the lack of high-fidelity test cycles for fuel cell light-duty trucks, this paper proposes an automated driving cycle construction method that integrates unsupervised clustering and reinforcement learning. Firstly, based on large-sample real-world driving data, four libraries of typical driving pattern segments are extracted through dimensionality reduction via Principal Component Analysis (PCA) and K-means clustering. Subsequently, the cycle construction process is formulated as a sequential decision-making problem, and a framework based on the Proximal Policy Optimization (PPO) algorithm, incorporating an action masking mechanism, is designed. This framework innovatively injects macro-level time budget allocation as a hard constraint into the agent’s policy space via action masking, while utilizing micro-level Markov transition probabilities as a soft guide. This dual approach drives the agent to learn an optimal segment concatenation strategy, thereby simultaneously ensuring both the macro-level statistical representativeness and the micro-level driving logic coherence of the synthesized cycle. Validation results demonstrate that the cycle constructed by the proposed method achieves an average relative error of only 7.53% in key characteristic parameters, and its joint speed-acceleration distribution exhibits a similarity as high as 0.9886 with the original data, significantly outperforming traditional methods such as the clustering method, the Markov chain method, and standard driving cycles. This study provides an effective tool for generating high-fidelity driving cycles and testing energy management strategies for fuel cell commercial vehicles.