This paper proposes a multi-agent deep graph reinforcement learning-based EV on-route charging guidance strategy, aiming at minimizing the charging cost for EV drivers in an uncertain and complex environment. First, a real-time online EV charging guidance framework with bi-timescales coupled transportation electrification is proposed. On the slow timescale, the distribution locational marginal pricing of the node where the charging station is located is solved by the power purchase cost optimization of the power distribution network. On the fast timescale, multi-agent deep reinforcement learning is used to solve real-time EV requests. Second, charging stations are considered as agents, and the potential competition for future charging demand at charging stations is considered. A multi-agent actor critic algorithm with embedded graph attention networks is proposed to optimize charging decision-making of EV drivers, which exploits the interactions between agents' observations using graph attention networks. Case studies are carried out in a practical area of Xi'an, China. We analyze the necessity of model components. The effectiveness of the proposed approach in reducing charging cost and its applicability in PV-equipped scenarios is verified. The convergence performance and scalability are verified by the comparison with SAC and MADDPG algorithms.

With the increasing requirements of the railway sector for electrified railways and the development of society, the traction power supply system needs to become more flexible, economic and reliable. The traditional traction power supply system has many problems due to its own topology, such as negative sequence current, load fluctuation, low utilization rate of regenerative braking energy, and inability to use the distributed generation along the railway. In order to solve these problems and make the traction power supply system (TPSS) more flexible, efficient and reliable, a novel TPSS called flexible smart traction power supply system (FSTPSS) is proposed. This traction power supply system consists of AC-DC-AC traction substation and the microgrid along the railway. It can effectively mitigate the problems of negative sequence current, improve regenerative braking energy utilization and distributed generation utilization. A hierarchical energy management strategy is also proposed here, which can effectively improve the operating economy of FSTPSS and reduce the load fluctuation of FSTPSS. And based on the actual load measurement power of a traction substation of China’s Beijing-Shanghai high-speed railway, the proposed energy management strategy has been verified for the economic improvement and load fluctuation reduction of TPSS.

The flexible smart traction power supply system (FSTPSS) is a fully electronic traction power supply system (TPSS), which integrates ac-dc–ac traction substations, distributed generation, and hybrid energy storage system (HESS). Compared with the traditional TPSS, the power supply of FSTPSS becomes more diverse, but the randomness of power flow becomes greater. A suitable energy management strategy is needed for FSTPSS to control the power flow of the system. At present, the day-ahead energy management strategy (DAEMS) has been proposed to make an optimal plan for the whole-day operation of FSTPSS, but this energy management strategy does not consider the impact of real-time fluctuation caused by nondeterministic loads or sources on the reliable execution of DAEMS. To solve this problem, this article proposes an online energy management strategy (OEMS) based on long short-term memory (LSTM) network and deep deterministic policy gradient (DDPG) algorithm to counteract the effects of these real-time fluctuations. The proposed OEMS has the advantages of small tracking error, model-free control, and continuous action control. To verify these advantages, a series of comparative simulations is carried out, and the simulation results verify the advantages of the proposed OEMS.

In this article, a low computational burden model predictive control (MPC) strategy without weighting factor is proposed for the single-phase cascaded H-bridge CHB converters. To reduce the switching state candidates, a hierarchy control algorithm is proposed. The grid current is controlled by selecting a subregion from the designed 2-D control plane, instead of the entire area. Two vectors are chosen in one sampling period for more accurate tracking. Then, the voltage balancing is achieved by selecting the optimal switching state from the subregion candidates to form the above two vectors. The cost function can be constructed of one variable: load voltage. Therefore, the weighting factor can be eliminated. No tuning or retuning processes are required in the proposed method. To reduce the computational time further, the principle of eliminating the switching state candidates operating the same voltage balancing performance is proposed. Conventional and proposed MPC methods are verified by experimental tests via a laboratory setup of a three-cell connected CHB converter. Steady-state and transient operations demonstrate that the proposed method guarantees less distortion grid current and shorter execution time (reduced from 15 to 3 $\mu s$ ). Fast response speed to variations in voltage reference and load resistance can be achieved t.

This paper presents the four-quadrant operation modes of bidirectional chargers for electric vehicles (EVs) framed in smart car parks. A cascaded model predictive control (MPC) scheme for the bidirectional two-stage off-board chargers is proposed. The controller is constructed in two stages. The model predictive direct power control for the grid side is applied to track the active/reactive power references. The model predictive direct current control is proposed to achieve constant current charging/discharging for the EV load side. With this MPC strategy, EV chargers are able to transmit the active and reactive powers between the EV batteries and the power grid. Apart from exchanging the active power, the vehicle-for-grid (V4G) mode is proposed, where the chargers are used to deliver the reactive power to support the grid, simultaneously combined with grid-to-vehicle or vehicle-to-grid operation modes. In the V4G mode, the EV battery functions as the static var compensator. According to the simulation results, the system can operate effectively in the full control regions of the active and reactive power (PQ) plane under the aforementioned operation modes. Fast dynamic response and great steady-state system performances can be verified through various simulation and experimental results.

Due to the high degree of controllability of the flexible smart traction power supply system (FSTPSS), day-ahead energy management strategy (DAEMS) was developed to optimize the power flow of the FSTPSS. However, the use of DAEMS is not based on real-time information. For FSTPSS, without real-time information, it cannot solve the problem of planning deviation caused by the real-time fluctuation of uncertain loads or sources. Therefore, this paper proposes a real-time energy management strategy (REMS) which is based on the real-time information to address the problem of planning deviation. REMS is implemented by LSTM and deep Q learning algorithm, where LSTM predicts uncertain loads or sources, and the deep Q-learning controls the operation of FSTPSS based on real-time predicted state. The proposed strategy is validated with the power flow simulation model of TPSS and the real measured data. The simulation results verify the necessity and superiority of the proposed method.