|
| 基于深度图神经网络的电动货车集群调度与路径优化方法 |
| Method for Electric Truck Cluster Scheduling and Path Optimization Based on Deep Graph Neural Networks |
| 投稿时间:2025-03-27 |
| DOI:10.16018/j.cnki.cn32-1650/n.202502007 |
| 中文关键词: 图注意力机制 深度强化学习 电动货车 取货和送货问题 最小化能耗 |
| 英文关键词: graph attention mechanism deep reinforcement learning electric truck the problem of pick-up and delivery minimize energy consumption |
| 基金项目:国家自然科学基金面上项目(52172383);安徽省高校协同创新项目(GXTT-20230042) |
|
| 摘要点击次数: 50 |
| 全文下载次数: 25 |
| 中文摘要: |
| 针对电动货车集群取送货问题中减少电动货车能耗和降低运营成本的关键指标,提出了一种结合图注意力网络和深度强化学习的模型。首先建立真实车辆能耗模型,在考虑电动车续航限制、载重变化以及阻力对能耗影响的情况下,以强化学习算法为基础求解能耗最优路径;结合图注意力机制有效学习图结构信息,以更好地捕捉节点之间的关系和重要性,并自适应地选择重要节点进行信息聚合,从而提高计算效率;最后,通过策略梯度方法进行训练,保证路线的可行性。验证与分析结果表明,与不加入图注意力网络的模型相比,该方法能够有效降低车队能耗;不同解码策略节能效果不同,其中DRL(sample128)和DRL(sample1280)收获的目标能耗均较小,但随着计算规模的扩大,在求解时间上DRL(sample1280)比DRL(sample128)大很多。 |
| 英文摘要: |
| To address the key objectives of reducing energy consumption of electric trucks and lowering operational costs in the pickup and delivery problem for electric truck clusters, a model combining Graph Attention Network (GAT) and Deep Reinforcement Learning (DRL) is proposed. First, a real vehicle energy consumption model is established. Considering the range limitations of electric vehicles, changes in load, and the impact of resistance on energy consumption, a reinforcement learning algorithm is used as the foundation to solve for the energy-optimal path. The graph attention mechanism is integrated to effectively learn graph structure information, better capture the relationships and importance between nodes, and adaptively select important nodes for information aggregation, thereby improving computational efficiency. Finally, training is conducted through policy gradient methods to ensure the feasibility of the routes. Verification and analysis results show that compared with the model without the graph attention network, this method can effectively reduce the energy consumption of the fleet. Different decoding strategies have different energy-saving effects, where both DRL (sample128) and DRL (sample1280) achieve relatively low target energy consumption. However, as the calculation scale expands, the solution time of DRL (sample1280) is significantly longer than that of DRL(sample128). |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|