人工智能技术支撑的集群电动汽车实时优化调度策略

Real-time Optimization Scheduling Strategy for Aggregated Electric Vehicles Supported by Artificial Intelligence Technology

电动汽车(electric vehicle,EV)在近年来得到了广泛的应用与部署,针对入网EV的充放电优化已成为研究热点。然而,传统的基于优化模型的EV优化调度方法在实际应用上面临模型参数难以准确获得和计算压力大的挑战。为了解决该问题,基于K-means聚类算法与长短期记忆神经网络(long short-term memory neural networks,LSTM)提出了一种集群电动汽车实时自动优化调度策略,直接从电动汽车的基础数据和电价生成满足约束的最优充放电计划。该策略基于分布式EV调度架构,由离线模型训练阶段和实时优化调度阶段2部分构成。在离线阶段,首先由K-means算法对海量EV数据聚类,之后用LSTM网络学习不同类型数据下的优化调度模式,建立从EV基础数据到优化决策之间的映射,并针对LSTM的输出设计了策略增强环节提高LSTM的决策性能。在实时阶段,在对EV类型识别的基础上,LSTM网络能够快速生成优化调度方案。仿真结果表明,与传统优化算法相比,所提策略能够在不依赖于用户提供准确的出行时间的情况下,毫秒级地输出近似最优解,适用于规模化EV的实时优化调度。

With the popular adoption and deployment of electric vehicles(EV), the optimization of EV charging and discharging power has become a research hotspot. However, the traditional optimization model-based EV scheduling method faces the challenges that it is difficult to obtain the accurate model parameters and that it has great pressure in calculation and communication. To address this issue, we formulate the EV charging/discharging scheduling strategy based on K-means clustering algorithm and the long short-term memory neural networks(LSTM), which directly generates the optimal scheduling results satisfying the constraints based on the EV information and the electricity prices. On the distributed scheduling architecture, this strategy consists of two parts: the off-line model training phase and the real-time optimal scheduling phase. In the off-line phase, the K-means algorithm is used to cluster the massive EV data. Then the LSTM network is adopted to learn the optimization scheduling modes under different types of data, andthe mapping from EV basic data to optimal decision is established. To improve the decision-making performance of the LSTM, a policy enhancement for the output of the LSTM is designed. In the real-time phase, on the basis of EV type identification, the LSTM network can quickly generate the optimized scheduling results. The simulation results show that compared with the traditional optimization methods, the proposed strategy can output the near-optimal solution in milliseconds without knowing exact EV departure time, which is suitable for the real-time optimization scheduling of a large-scale EVs.