基于多主体依互与在线学习的图神经网络"网-源-储-车"协同供能系统能量超前管控方法OA

Byintegratingnew energy systems and energy storage systems into the traction power supply system,a"grid-source-storage-train"coordinated power supply system has beenestablished.This system serves as a key measure to promote green,low-carbon development in rail transit.However,new energy sources exhibit fluctuating and unpredictable output characteristics.Electric locomotives show impactful power demands.These factors make the real-time energy management and control of the cooperative power supply system significantly more challenging.Therefore,this paper proposes a"grid-source-storage-train"collaborative energy supply system energy-overrun control method based on multi-subject dependencies and an online-learning-based graph neural network. First,K-Means clustering and kernel density estimation methods were used.Operating conditions were classified for multiple entities.These entities included new energy systems,energy storage systems,and electric locomotives.Second,an objective function was constructed within the optimization model to determine the optimal transition paths.For the new energy system,the energy management strategy maximized the utilization efficiency of the new energy.It also promoted local consumption.For the energy storage system,both the absorption efficiency of regenerative braking energy and the energy feedback efficiency were enhanced,reducing traction losses.An objective function was formulated to generate an optimal graph of transition relationships for operating conditions.Nodes represented the operating conditions of different entities,and edges denoted transition probabilities or conditions. Graph convolutional networks and multi-head attention mechanisms were employed to extracts patio temporal features.A multi-level feature extraction network architecture was established.Specifically,the graph convolutional layer aggregated information from local neighbors.Connectivity features among multiple entities were extracted.Subsequently,the multi-head attention mechanism extracted temporal features,enhancing the model's ability to capture non-local relationships.Next,local spatial features and non-local temporal features were fused.Two layers of graph convolutions were used to further extract deep global features.Finally,the energy management layer mapped the deep representations to corresponding operating states.To enhance the applicability and robustness of the graph neural network model,an online learning strategy was implemented.Real-time operational data were collected.The model was incrementally updated.New data were progressively incorporated without discarding existing information.This approach strengthened the model's ability to handle anomalies,thereby improving the graph neural network's energy management accuracy. Simulation results validate the effectiveness of the proposed method.For the photovoltaic system,the energy management strategy achieves a PV energy absorption of 10 556.59 kW·h.It also reaches a high energy utilization efficiency of 81.81%.For the energy storage system,the proposed energy management strategy achieves a regenerative braking energy absorption efficiency of 38.35%and a storage system feedback efficiency of 53.82%,saving a total of 13 539.41 kW·h of traction energy per day. In summary,a collaborative energy-supply system for a"grid-source-storage-train"with an overrun-control method was proposed.This method achieves efficient control of energy interaction.Future applications can extend this proposed graph neural network model to energy interaction management across multiple stations.This expansion of the model's applicability will provide stronger technical support for the green transformation of rail transit systems.