Intelligent scheduling method, device and equipment for battery swap station based on virtual power plant cooperation

Abstract

This invention discloses a method, device, and equipment for intelligent scheduling of battery swapping stations based on virtual power plant collaboration, relating to the field of intelligent scheduling of battery swapping stations. The method includes: first, constructing a spatiotemporal correlation model for a group of battery swapping stations based on a graph attention network; second, achieving joint prediction of grid load and battery swapping station status based on an LSTM-Attention dual-stream fusion model; and third, constructing a distributed optimization decision-making architecture based on multi-agent deep reinforcement learning. This invention boasts strong spatiotemporal collaboration capabilities: constructing a graph attention network to achieve spatiotemporal correlation modeling of a group of battery swapping stations accurately captures the coupling relationships and dynamic characteristics between stations, overcoming the limitations of weak collaboration in traditional centralized scheduling, and achieving high efficiency in inter-station power mutual assistance; it also features high prediction accuracy and reliability: employing an LSTM-Attention dual-stream fusion model, integrating multi-source data such as battery health, meteorology, and new energy sources, it achieves high long-term prediction accuracy and effectively reduces grid fluctuation response errors.

Description

Technical Field

[0001] This invention relates to the field of intelligent scheduling technology for battery swapping stations, and specifically to intelligent scheduling methods, devices, and equipment for battery swapping stations based on virtual power plant collaboration. Background Technology

[0002] With the large-scale popularization of new energy vehicles and the high proportion of new energy connected to the grid, battery swapping stations, as key facilities with both load and energy storage attributes, play an important role in grid peak shaving, new energy consumption, and load smoothing in the virtual power plant system. Existing battery swapping station dispatching technologies mostly adopt centralized optimization strategies, which have significant drawbacks: First, weak inter-station coordination, ignoring the spatiotemporal correlation and power mutual assistance potential between battery swapping stations, resulting in high communication overhead and poor robustness; second, insufficient prediction accuracy, failing to fully integrate multi-source heterogeneous information such as grid dynamics, battery health status, meteorological data, and new energy fluctuations, leading to large errors in long-series load and capacity predictions; third, difficulty in meeting real-time requirements, relying on centralized cloud computing, resulting in high response latency and inability to adapt to the millisecond-level dispatching needs of the grid; and fourth, poor scenario adaptability, failing to design collaborative optimization and fault self-healing mechanisms for the characteristics of large peak-valley differences and prominent peak-shaving pressure in wind and solar-rich areas, resulting in insufficient dispatching reliability and economy.

[0003] In the existing technology, therefore, we provide a method, device and equipment for intelligent scheduling of battery swapping stations based on virtual power plant collaboration to solve the above problems. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method, device, and equipment for intelligent scheduling of battery swapping stations based on virtual power plant collaboration, which takes into account spatiotemporal collaboration, high-precision prediction, millisecond-level response, and scenario adaptation for intelligent scheduling of battery swapping stations.

[0005] To achieve the above objectives, the present invention employs an intelligent scheduling method for battery swapping stations based on virtual power plant collaboration, comprising: The first step is to construct a spatiotemporal correlation model of the battery swapping station group based on graph attention network, collect and preprocess data from the power grid side and the battery swapping station side, construct a topology graph with battery swapping stations as nodes and power grid connection and communication links as edges, and extract the spatiotemporal correlation features of battery swapping stations through a two-layer GAT network. The second step is to jointly predict the grid load and the status of the swapping station based on the LSTM-Attention dual-stream fusion model. This involves fusing grid time series, swapping station time series and meteorological data to output the grid load curve and the time series trajectory of the adjustable capacity of each swapping station for the next 24 hours. The third step involves constructing a distributed optimization decision-making architecture based on multi-agent deep reinforcement learning. This architecture consists of one cloud-based evaluation agent (Critic) and multiple edge-based execution agents (Actor), with offline centralized training and online distributed execution to achieve decentralized collaborative scheduling and millisecond-level adaptive control of the battery swapping station group.

[0006] As a further optimization of the above scheme, in the first step, the topology graph is a dynamic topology graph, which collects grid switch status, line on / off status and impedance data in real time, with a communication delay of ≤20ms. Dynamic edges are established when the line impedance is <0.5Ω, and the attention weights are reconstructed and updated in real time when the topology change rate is >5%, with a weight update cycle of ≤100ms.

[0007] As a further optimization of the above scheme, in the second step, the LSTM-Attention model integrates battery health status (SOH) data, adds battery cycle count, internal resistance, and temperature features, reduces the scheduling weight for batteries with SOH < 80%, and outputs the time-series trajectory of healthy and usable capacity.

[0008] As a further optimization of the above scheme, the third step adopts a layered lightweight deployment at the edge and the cloud. The edge deploys lightweight GAT and LSTM models with INT8 quantization, and the inference time of a single model is ≤10ms. The cloud is responsible for global training and anomaly handling, and the cloud takeover switching time is ≤200ms when the edge is offline.

[0009] As a further optimization of the above scheme, the third step integrates wind and solar power output with time-of-use pricing for coordinated scheduling. During off-peak hours, battery swapping stations are given priority for full-power charging, while during peak hours, they are given priority for discharging to absorb wind and solar power output. The system has a fault self-healing mechanism with a fault identification time of ≤50ms. After a fault, the topology is reconstructed and the scheduling tasks are smoothly shared, with a power adjustment rate of ≤10kW / s.

[0010] As a further optimization of the above scheme, in the third step, the Actor's action space is the continuous charging and discharging power, ranging from [- , The step size is 10kW; the Critic agent reward function includes three parts: load matching reward, economic benefit reward, and safety constraint penalty.

[0011] This invention also discloses an intelligent scheduling device for battery swapping stations based on virtual power plant collaboration, used in conjunction with an intelligent scheduling method for battery swapping stations based on virtual power plant collaboration, and further includes: The spatiotemporal correlation modeling module is used to construct a dynamic topology graph attention network and extract spatiotemporal correlation features of the battery swapping station group; The joint prediction module is used to construct an LSTM-Attention dual-stream fusion model and output the prediction results of grid load and adjustable capacity of battery swapping stations. The distributed scheduling module is used to deploy a multi-agent deep reinforcement learning framework to achieve decentralized collaborative scheduling and millisecond-level control of battery swapping stations.

[0012] The present invention also discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the program to implement the steps of a smart scheduling method for battery swapping stations based on virtual power plant collaboration.

[0013] The intelligent scheduling method, apparatus, and equipment for battery swapping stations based on virtual power plant collaboration of the present invention have the following beneficial effects: Strong spatiotemporal coordination capability: Constructing a graph attention network to realize spatiotemporal correlation modeling of battery swapping station groups, accurately capturing the coupling relationship and dynamic characteristics between stations, breaking through the limitations of weak coordination in traditional centralized scheduling, and achieving high efficiency of power mutual assistance between stations; High prediction accuracy and good reliability: The LSTM-Attention dual-stream fusion model is adopted to integrate multi-source data such as battery health, meteorology, and new energy, resulting in high long-term prediction accuracy and effectively reducing grid fluctuation response error. Fast real-time response and strong engineering feasibility: It adopts a layered lightweight deployment of edge and cloud, with millisecond-level autonomous scheduling at the edge and a response latency of ≤30ms. The cloud computing power consumption is reduced, and it can be deployed without large-scale hardware modification. Excellent scenario adaptability and significant economic benefits: It integrates time-of-use pricing with wind and solar new energy for synergistic optimization, improves the new energy consumption rate, and reduces the peak-valley difference of the power grid; it has the ability to self-heal faults, with short fault identification time and high self-healing success rate, ensuring stable dispatch operation; Comprehensive safety constraints and long service life: By introducing battery health constraints, the intensity of aging battery scheduling is limited, the battery failure rate is reduced, the service life of equipment is extended, and the safety of system operation is improved.

[0014] Specific embodiments of the present invention are disclosed in detail with reference to the following description and accompanying drawings, indicating how the principles of the present invention can be adopted. It should be understood that the embodiments of the present invention are not limited in scope as a result, and that the embodiments of the present invention include many changes, modifications and equivalents within the spirit and scope of the appended claims. Attached Figure Description

[0015] Figure 1 This is a diagram of the intelligent scheduling method for battery swapping stations based on virtual power plant collaboration according to the present invention. Detailed Implementation

[0016] Example 1 refer to Figure 1As shown in the illustration, this embodiment provides an intelligent scheduling method for battery swapping stations based on virtual power plant collaboration. It integrates deep temporal prediction, graph neural networks, and multi-agent deep reinforcement learning to achieve intelligent collaboration and millisecond-level adaptive control between the battery swapping station group and the power grid. The system is deployed on a virtual power plant cloud platform and multiple distributed battery swapping stations. Each battery swapping station is equipped with: a battery management system (BMS), an energy storage converter (PCS), an edge acquisition terminal, and a communication module (4G / 5G). The virtual power plant platform deploys a server cluster to complete modeling, prediction, training, and scheduling decisions. The specific steps are as follows: Step 1: Spatiotemporal correlation modeling of battery swapping station groups based on graph attention network (GAT); 1. Data acquisition and preprocessing; The virtual power plant platform collects basic data every 10 seconds. Grid side: Regional grid bus voltage, line power flow, topology connections, and operating status of adjacent substations; On the battery swapping station side: geographical coordinates of the battery swapping station, inter-station communication delay, line impedance, maximum charging and discharging power, and current operating status; Data cleaning: Outliers (voltage deviation from rated ±10%, delay >100ms) are removed. Missing values ​​are interpolated using the average of adjacent stations. The sampling frequency is uniformly 1 minute.

[0017] 2. Construct the battery swapping station-grid topology; Each battery swapping station is treated as a node in the graph. The node feature vector has 16 dimensions, including: station number, geographical location, maximum charging power, maximum discharging power, communication latency, line impedance, number of currently available batteries, average battery SOC, charging efficiency, discharging efficiency, historical 1-hour load, historical 1-hour output, electricity price, response delay, adjustable capacity, and equipment health status. The grid physical connection and communication link are used as edges: a directed edge is established when there is a direct connection between two stations or the communication latency is ≤50ms. The edge features include: line impedance, communication bandwidth, and transmission loss. Finally, a static topology graph G=(V,E) is formed, with N≥3 nodes and M≥N-1 edges.

[0018] 3. Construction and Training of Graph Attention Network (GAT) Model Network structure: A 2-layer GAT is used, with the first layer having 8 attention heads and an output dimension of 64; the second layer has 4 attention heads and an output dimension of 32; the activation function is ELU; Attention calculation: For each node i, calculate the attention weights for its neighboring nodes j. ; in For node features, Let be the original feature vector of the adjacent node j, and W be the weight matrix. Let be the vector transformed from the feature path of node i, and let a be the attention vector. for The set of adjacent nodes; j,k are: node / battery swapping station number subscripts ( (where j and k are adjacent nodes), and exp is the natural exponential function; ( ) represents the leaky ReLU activation function; This involves concatenating vectors. Attention vector Transpose of; For nodes Concatenate with the transformed eigenvector of j.

[0019] Feature aggregation: Output features of node i ; The sigmoid activation function is used. For nodes The output feature vector updated by GAT; To sum the values ​​of all adjacent nodes j of node i; Weighted contribution of features to adjacent node j.

[0020] Training data: Historical data of 3 months of battery swapping station operation and power grid topology data were used. The loss function was node feature reconstruction error. Adam optimizer was used with a learning rate of 0.001 and 100 iterations. Output: After training, output a 32-dimensional spatiotemporal correlation feature vector per station, which will be used as the input feature for subsequent prediction models.

[0021] The second step is to jointly predict the grid load and the state of the battery swapping station based on LSTM-Attention. 1. Multi-source heterogeneous data fusion; Input data was sampled at 15-minute intervals per step, with a time series length of 7 historical days (672 steps): Power grid time-series data: historical load, real-time load, voltage, frequency, and weather (temperature, humidity, irradiance, wind speed); Battery swapping station time-series data: battery SOC, number of available battery packs, charging power, discharging power, charging and discharging efficiency, and number of battery swapping orders for each station; Feature normalization: All time-series features are normalized to [-1,1]. Samples are generated using a sliding window with a window length of 672 and a prediction step size of 96 (24 hours).

[0022] 2. LSTM-Attention two-stream fusion model structure; A parallel dual-stream + cross-attention fusion architecture is adopted: Flow 1 (Grid Load Flow): 2-layer LSTM, hidden cells 128→64, extracting long-range temporal dependencies of grid load; Flow 2 (Swapping Station State Flow): 2-layer LSTM, hidden units 128→64, extracting the SOC and capacity time-series features of each swapping station; Attention layer: Employs multi-head attention (4 heads) to calculate the cross-attention weight between the two flows, dynamically focusing on key time steps (such as peak hours) and key battery swapping stations (large capacity, fast response). Output layer: Fully connected layer + linear activation, outputting the following for the next 24 hours (96 points): ① Regional power grid load forecast curve; ② Adjustable capacity time-series trajectory for each battery swapping station (adjustable capacity = number of available batteries × single battery capacity × SOC × efficiency).

[0023] 3. Model training and inference; Training set: 6 months of historical data, validation set: 1 month, test set: 1 month; Loss function: MSE (load) + MAE (adjustable capacity), weighted 1:1; Training parameters: batch_size=32, learning rate 0.0005, 200 iterations, early stopping (validation set loss does not decrease for 10 consecutive iterations). Inference: The prediction results are automatically updated every 15 minutes, and the output results are stored in the time series database as the feedforward input for multi-agent scheduling.

[0024] The third step is to develop a distributed optimization decision-making architecture based on multi-agent collaboration. 1. Multi-agent system initialization; Intelligent agent division: 1 centralized evaluation intelligent agent Critic (deployed in the cloud) + N distributed execution intelligent agents Actor (1 per station, deployed at the edge); Actor State Space (local per station): Battery SOC (0~1), Current Charging Power, Current Discharging Power, Time-of-Use Electricity Price (Peak / Side / Valley), Adjustable Capacity, Equipment Status (Normal / Alarm); Actor action space: continuous action, charging and discharging power. Step size 10kW; P is power (unit: kW); This represents the maximum discharge power of a single station. This represents the maximum charging power of a single station. Critic state space (global): Real-time grid load, predicted grid load, SOC of all battery swapping stations, adjustable capacity of all stations, electricity price, total output; Critic Value Space: Global Cumulative Reward Value Where R is the real number space and Q is the value function.

[0025] 2. Offline centralized training (completed in the cloud) Training framework: MAPPO (Multi-agent Proximal Policy Optimization); Reward function (Critic): ; in: Total global reward; The negative penalty for the deviation between the actual load of the power grid and the total output of the battery swapping station; the smaller the deviation, the higher the reward. Peak discharge revenue - Off-peak charging cost; Negative penalty when SOC < 20% or > 90%, and power penalty for exceeding the limit; Training process: Randomly generate power grid load fluctuation scenarios and initial SOC scenarios for battery swapping stations, iterate and train for 5000 rounds, and after convergence, save the parameters of the Actor policy network and Critic value network and distribute them to the edge terminals of each battery swapping station.

[0026] 3. Online distributed execution (edge ​​autonomy + cloud evaluation) Real-time decision cycle: 100ms; Edge Actor: Collects local status every 100ms, inputs it into the policy network, outputs the current optimal charge / discharge power command, and sends it to the PCS for execution; Cloud-based Critic: Receives all station actions and global status every 100ms, evaluates the value of joint actions, and if the global reward is lower than the threshold, sends a small correction signal (±5% power) for rapid adjustment at the edge. Communication mechanism: There is no direct communication between Actors, only one-way communication between Critic and Actor, with communication overhead ≤100kbps / station, avoiding centralized communication bottlenecks; Control effect: In the pilot project at 5 battery swapping stations, the dispatch response delay was ≤50ms, the grid load tracking error was ≤3%, the inter-station power mutual assistance efficiency was ≥90%, and decentralized autonomous collaboration was achieved.

[0027] Example 2 Based on Example 1, the following steps are implemented to address the scheduling inaccuracies caused by dynamic changes in power grid topology and battery aging: Dynamic topology adaptive GAT spatiotemporal modeling: Data Acquisition: The virtual power plant platform collects the status of regional power grid switches, line connectivity, voltage levels, and line impedance every 50ms; it also collects the geographical location and communication latency of the battery swapping stations simultaneously.

[0028] Dynamic graph construction: Dynamic edges are established when the communication delay between two stations is ≤20ms and the line impedance is <0.5Ω; the graph structure is reconstructed in real time when the topology change rate is >5%.

[0029] Adaptive attention update: GAT adds a topology-aware gating unit with a weight update period of ≤100ms; outputs a spatiotemporal correlation vector containing topological dynamic features.

[0030] Battery health fusion LSTM-Attention prediction Data expansion: Added battery cycle count, internal resistance, temperature, and aging degradation coefficient to construct the SOH time series with a sampling interval of 15 minutes.

[0031] Model Improvement: The dual-flow structure is divided into grid load flow and swapping station status + SOH flow; an SOH weight module is added to the attention layer, and the weight is reduced by 30%–50% when SOH < 80%.

[0032] Output: Grid load, theoretical adjustable capacity, and healthy available capacity for the next 24 hours (=theoretical capacity × SOH coefficient).

[0033] Health-constrained multi-agent scheduling Actor Input: Add SOH and aging rate; discharge power ≤ 60% of rated power when SOH is 60%–80%; discharge is prohibited when SOH < 60%.

[0034] Critic value function: Incorporate aging loss cost term and prioritize battery swapping stations with high health.

[0035] Results: In the pilot station, the topology response was ≤80ms, the failure rate of aging batteries decreased by 40%, and the scheduling accuracy reached 92%.

[0036] Example 3 To address the bottlenecks of high latency in the cloud and insufficient edge computing power, the implementation steps are as follows: Layered deployment from edge to cloud: Edge layer: Each station deploys embedded units with a computing power of ≥1 TOPS (such as JetsonOrin); GAT pruning sparsification, LSTM depthwise separable convolution replacing fully connected layers, INT8 quantization, and model volume compressed to 1 / 4.

[0037] Cloud layer: Deploys the complete model and is responsible for global training, parameter distribution, and anomaly decision-making; dual-link redundancy, switching ≤200ms.

[0038] Lightweight reasoning: GAT: Only calculates attention to adjacent stations within 10km, inference ≤5ms; LSTM-Attention: 24h prediction ≤10ms.

[0039] The prediction results are cached locally at the edge and updated every 15 minutes.

[0040] Millisecond-level closed-loop control: Status acquisition: SOC, power, and voltage are acquired every 10ms at the edge and preprocessed locally.

[0041] Action execution: The Actor outputs a command to the PCS, with a response time of ≤20ms.

[0042] Global correction: Cloud-based Critic evaluates every 100ms, and if the deviation is >5%, correction is issued. Closed-loop correction is ≤100ms.

[0043] Results: In the pilot station, the response time was ≤30ms, cloud computing power was reduced by 60%, and it was compatible with existing hardware.

[0044] Example 4 Taking into account the region's abundant scenic spots and significant peak-to-valley differences, the implementation steps are as follows: Wind-solar-battery swapping station synergistic prediction: Data fusion: Access wind power output, photovoltaic irradiance, and inverter status, with a 15-minute resolution.

[0045] Multimodal model: A new renewable energy branch has been added, which integrates load, battery swapping station, and wind and solar data to output a net load curve (load - wind and solar).

[0046] Peak-valley identification: Time periods are marked according to local time-of-use electricity prices (peak hours 8:00–22:00, off-peak hours 22:00–8:00).

[0047] Time-of-use pricing-driven dispatch: Off-peak hours: Full-power charging to store low-cost electricity and smooth out off-peak loads.

[0048] Peak periods: Prioritize power generation, absorb wind and solar power, and fill peak demand gaps.

[0049] For flat periods: maintain SOC at 50%–70%, leaving room for future adjustments.

[0050] Fault self-healing scheduling: Fault detection: Edge monitoring communication interruption, PCS failure, battery short circuit, identification ≤50ms.

[0051] Topology reconstruction: GAT deletes faulty nodes and reconstructs the graph; health stations are assigned tasks according to capacity.

[0052] Smooth control: Power ramp adjustment ≤10kW / s to avoid impact.

[0053] Results: In the pilot stations, the wind and solar energy absorption rate was 95%, the peak-valley difference decreased by 25%, and the self-healing rate was 99%.

Claims

1. A method for intelligent scheduling of battery swapping stations based on virtual power plant collaboration, characterized in that, include: The first step is to construct a spatiotemporal correlation model of the battery swapping station group based on graph attention network, collect and preprocess data from the power grid side and the battery swapping station side, construct a topology graph with battery swapping stations as nodes and power grid connection and communication links as edges, and extract the spatiotemporal correlation features of battery swapping stations through a two-layer GAT network. The second step is to jointly predict the grid load and the status of the swapping station based on the LSTM-Attention dual-stream fusion model. This involves fusing grid time series, swapping station time series and meteorological data to output the grid load curve and the time series trajectory of the adjustable capacity of each swapping station for the next 24 hours. The third step involves constructing a distributed optimization decision-making architecture based on multi-agent deep reinforcement learning. This architecture consists of one cloud-based evaluation agent (Critic) and multiple edge-based execution agents (Actor), with offline centralized training and online distributed execution to achieve decentralized collaborative scheduling and millisecond-level adaptive control of the battery swapping station group.

2. The intelligent scheduling method for battery swapping stations based on virtual power plant collaboration according to claim 1, characterized in that: In the first step, the topology graph is a dynamic topology graph, which collects grid switch status, line on / off status and impedance data in real time. The communication delay is ≤20ms. Dynamic edges are established when the line impedance is <0.5Ω. When the topology change rate is >5%, the attention weights are reconstructed and updated in real time, and the weight update cycle is ≤100ms.

3. The intelligent scheduling method for battery swapping stations based on virtual power plant collaboration according to claim 1, characterized in that: In the second step, the LSTM-Attention model integrates battery health status (SOH) data, adds battery cycle count, internal resistance, and temperature features, reduces the scheduling weight for batteries with SOH < 80%, and outputs a time-series trajectory of healthy and usable capacity.

4. The intelligent scheduling method for battery swapping stations based on virtual power plant collaboration according to claim 1, characterized in that: In the third step, a layered lightweight deployment is adopted at the edge and cloud. Lightweight GAT and LSTM models with INT8 quantization are deployed at the edge, with a single model inference time of ≤10ms. The cloud is responsible for global training and anomaly handling, and the cloud takeover switching time is ≤200ms when the edge is offline.

5. The intelligent scheduling method for battery swapping stations based on virtual power plant collaboration according to claim 1, characterized in that: In the third step, wind and solar power output is integrated with time-of-use pricing for coordinated scheduling. During off-peak hours, battery swapping stations are prioritized for full-power charging, while during peak hours, they are prioritized for discharging to absorb wind and solar power output. The system has a fault self-healing mechanism with a fault identification time of ≤50ms. After a fault, the topology is reconstructed and the scheduling tasks are smoothly shared, with a power adjustment rate of ≤10kW / s.

6. The intelligent scheduling method for battery swapping stations based on virtual power plant collaboration according to claim 1, characterized in that: In the third step, the Actor's action space is the continuous charge and discharge power, ranging from [- , The step size is 10kW; the Critic agent reward function includes three parts: load matching reward, economic benefit reward, and safety constraint penalty.

7. A smart dispatching device for battery swapping stations based on virtual power plant collaboration, characterized in that: The method for intelligent scheduling of battery swapping stations based on virtual power plant collaboration, in conjunction with any one of claims 1-6, further includes: The spatiotemporal correlation modeling module is used to construct a dynamic topology graph attention network and extract spatiotemporal correlation features of the battery swapping station group; The joint prediction module is used to construct an LSTM-Attention dual-stream fusion model and output the prediction results of grid load and adjustable capacity of battery swapping stations. The distributed scheduling module is used to deploy a multi-agent deep reinforcement learning framework to achieve decentralized collaborative scheduling and millisecond-level control of battery swapping stations.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the intelligent scheduling method for battery swapping stations based on virtual power plant collaboration as described in any one of claims 1 to 6.

Images