Deep Reinforcement Learning (DRL) techniques have shown promising results in developing proficient energy management strategies (EMSs) for hybrid electric vehicles (HEVs). However, the current DRL algorithms are often trained on a limited number of standard driving cycles, which do not necessarily represent the highly variable and diverse driving conditions encountered by HEVs in the real world. Therefore, a significant current research objective is to develop DRL methods that can adapt to new and arbitrary driving cycles. Recent studies have shown that DRL algorithms based on Recurrent Neural Networks (RNNs) can better address this challenge by taking into account the temporal dependencies of environmental state. Building upon this insight, we propose an EMS based on a recurrent policy that is trained on a large number of random driving cycles to enhance the generalization of energy management strategies to arbitrary driving cycles. We evaluate the proposed method on a classic HEV, namely the Prius, and demonstrate that the recurrent policy significantly outperforms traditional DRL methods in terms of fuel economy. Furthermore, the recurrent policy is shown to be more robust and generalizable to variable driving cycles.