A power grid space-time dynamic path planning method fusing gnn-utvd-cbf
By integrating the GNN-UTVD-CBF approach, the path planning problem under dynamic obstacles and complex constraints in the power grid was solved, achieving fast, safe, and efficient path generation and repair, and improving the intelligent operation and maintenance level of power grid inspection equipment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TIANJIN UNIV
- Filing Date
- 2026-02-26
- Publication Date
- 2026-05-26
AI Technical Summary
Traditional path planning methods struggle to handle dynamic obstacles and complex constraints in power grids, resulting in low computational efficiency, insufficient security, and difficulty in meeting real-time requirements. Existing learning-based methods also lack the ability to handle dynamic environments.
A spatiotemporal dynamic path planning method for power grids, which integrates GNN, UTVD, and CBF, is adopted. This method generates multiple candidate paths by calculating the safety time interval, constructing a dynamic connection view, using graph neural networks to predict expansion priorities, and combining linear quadratic regulators and control obstacle functions for path search. Local repair is also performed during the trajectory execution process.
It achieves rapid response in dynamic environments, provides mathematically provable safety guarantees, takes into account equipment dynamics, reduces redundant sampling and collision detection, and improves planning efficiency and robustness.
Smart Images
Figure CN122083976A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power grid inspection technology, and in particular to a spatiotemporal dynamic path planning method for power grids that integrates GNN-UTVD-CBF. Background Technology
[0002] In the process of building smart grids, drone inspections, robotic distribution path planning, and intelligent dispatching systems have become important technical means to improve the efficiency and safety of grid operation and maintenance. With the continuous expansion of the grid's scale and the increasing complexity, traditional manual inspections and static path planning methods face numerous problems such as high cost, high risk, low efficiency, and poor environmental adaptability. Especially in complex scenarios such as mountain lines, high-voltage corridors, and substations, how to achieve fast, safe, and efficient dynamic path planning has become a key technical challenge that urgently needs to be solved.
[0003] Traditional path planning methods mainly include sampling-based methods (such as RRT and PRM) and search-based methods (such as A* and Dijkstra). These methods perform well in static environments, but in scenarios such as power grids, which contain dynamic obstacles (such as mobile work vehicles, temporary construction areas, and bird activity zones), complex constraints (such as electromagnetic field strength limits, safety distance requirements, and equipment dynamics constraints), and high real-time requirements, they often fail to meet practical application needs. Existing path planning technologies mainly face the following challenges:
[0004] 1. Limitations of static programming: Most traditional methods assume that the environment is static or changes slowly. When encountering sudden obstacles or rapid changes in the environment, frequent replanning is required, resulting in wasted computing resources and response delays.
[0005] 2. Insufficient handling of dynamic constraints: Power grid inspection equipment (drones, robots) have complex kinematic and dynamic characteristics. Existing methods often simplify them into a point mass model, ignoring dynamic constraints such as velocity and acceleration, which makes the generated paths difficult to execute in practice.
[0006] 3. Difficulty in ensuring safety: The power grid environment contains dangerous factors such as high-voltage equipment and electromagnetic fields. Existing methods lack formal guarantees for safety, which can easily lead to potential safety hazards.
[0007] 4. Low computational efficiency: In complex three-dimensional power grid environments, traditional sampling and search methods require a large number of collision detection and state expansion operations, resulting in high computational complexity and making it difficult to meet the real-time requirements of online planning.
[0008] In recent years, with the development of deep learning and intelligent optimization technologies, learning-based path planning methods have begun to attract attention. However, most existing learning-based methods are designed for static environments, lack the ability to handle dynamic obstacles and complex constraints, and have limited generalization capabilities. Summary of the Invention
[0009] To overcome the above shortcomings, this invention provides a spatiotemporal dynamic path planning method for power grids that integrates GNN-UTVD-CBF, aiming to improve the planning efficiency, security and robustness of power grid inspection, distribution path planning and intelligent dispatching systems in complex dynamic environments.
[0010] This invention provides the following technical solution: a spatiotemporal dynamic path planning method for power grids integrating GNN-UTVD-CBF, comprising:
[0011] S1. Based on the three-dimensional model of the power grid environment and dynamic obstacle information, combined with equipment dynamic constraints, calculate the safe time interval of potential path edges;
[0012] S2. Based on the security time interval, sampling and connection are performed in free space to construct a dynamic connection view;
[0013] S3. Utilize graph neural networks to predict the expansion priority of edges in the dynamically connected visibility view, and combine this with a unified temporal visibility deformation criterion to perform path search and generate multiple candidate paths.
[0014] S4. During the path search process, the safety of the extended edge is verified by combining the linear quadratic regulator and the control barrier function.
[0015] S5. Generate a spatial-temporal corridor for the candidate path based on the security time interval and perform an expansion process;
[0016] S6. Within the space-time corridor, B-spline curves are used to parameterize the path, and the final trajectory is solved through multi-objective optimization.
[0017] S7 monitors environmental changes during trajectory execution and performs local cropping and incremental repair on the dynamic connection view.
[0018] Preferably, in step S1, the step of calculating the safe time interval of potential path edges includes:
[0019] Static obstacles in the power grid environment are mapped as a 3D occupancy grid map, and dynamic obstacles are modeled as bounded ellipsoids with trajectory prediction information.
[0020] Establish a set of safety constraints that include minimum safe distance, electromagnetic field exposure intensity, airspace restrictions, and equipment dynamic parameters;
[0021] Based on the device dynamic parameters, the minimum passage time of potential path edges is determined, the collision time interval between the potential path edges and dynamic obstacles is calculated, and the collision time interval is removed from the planning cycle to determine the safe time interval that meets the passage requirements.
[0022] Preferably, in step S2, the step of constructing the dynamically linked view includes:
[0023] Heuristic sampling is performed in free space to obtain sampling points, and the guard points visible to the sampling points are identified;
[0024] Obtain the safe time interval of the connecting edge between the sampling point and the guard point, and determine whether the safe time intervals of the connecting edge overlap.
[0025] If the security time intervals overlap, a connection relationship is established between the sampling point and the guard point, and the sampling point is added to the dynamic connection view.
[0026] Preferably, the step of constructing a dynamically linked view further includes:
[0027] Using the UTVD criterion, we evaluate whether the paths formed by newly added connection points and neighboring nodes are topologically equivalent in the spatiotemporal domain.
[0028] If the new path and the existing path belong to the same UTVD class, only the nodes with lower costs are retained and updated to the dynamic connection view; if they belong to different UTVD classes, the connection point is added as a new node.
[0029] Preferably, in step S3, the step of generating multiple candidate paths includes:
[0030] The nodes and edges of the dynamically connected visual graph are embedded and encoded using a graph neural network, and message passing is performed. The spatiotemporal constraint information of the local neighborhood is aggregated, and the expansion priority score of each edge is output.
[0031] Based on the expansion priority score, a depth-first search strategy is adopted to prioritize the expansion of edges that satisfy the safety time interval constraint until they are connected to the starting point or an existing subtree to generate an initial path.
[0032] Based on the UTVD criteria, the topology category of the initial path is identified, and paths of different topology categories are retained to form a candidate path set.
[0033] Preferably, in step S4, the step of performing security verification on the extended edge includes:
[0034] The system is locally linearized near the current state, the algebraic Riccati equation is solved to obtain the feedback gain matrix, and the state trajectory is generated based on the linear quadratic regulator control law.
[0035] For each time step of the state trajectory, calculate the value of the higher-order control barrier function;
[0036] If the higher-order control barrier function value satisfies the safety condition at all time steps, the extended edge is determined to be feasible; otherwise, the extension is terminated and marked as infeasible.
[0037] Preferably, in step S5, the step of generating a spatial-temporal corridor for the candidate path based on the security time interval and performing dilation processing includes:
[0038] The candidate path is divided into several road segments, and an initial spatial corridor that avoids static obstacles is constructed with the road segments as the axis. The effective time window of the initial spatial corridor is determined based on the safe time interval.
[0039] Select a seed point and expand the boundary of the initial spatial corridor in all directions. During the expansion process, detect whether the new area collides with static or dynamic obstacles.
[0040] If a collision is detected, the expansion in the corresponding direction is stopped, and the effective time window is updated synchronously, ultimately forming a series of expanded space-time corridors.
[0041] Preferably, in step S6, the step of solving the final trajectory through multi-objective optimization includes:
[0042] Construct a multi-objective optimization function that includes control cost, dynamic feasibility cost, corridor constraint cost, and dynamic obstacle collision cost;
[0043] The control point positions of the B-spline curve are iteratively adjusted using the gradient descent method to minimize the multi-objective optimization function;
[0044] The quick smoother is used to perform a sliding window search on the optimized trajectory. If there is a collision-free path between non-adjacent control points, the intermediate redundant control points are deleted and the trajectory is updated.
[0045] Preferably, in step S7, the step of performing local cropping and incremental repair on the dynamic connection view includes:
[0046] Calculate the radius of influence of dynamic obstacles affected by environmental changes, remove only nodes and their associated edges within the radius of influence, and retain the unaffected graph structure;
[0047] Identify disconnected connected components in the dynamic connection visualization after pruning, wherein the connected components include the starting subtree, the ending subtree, and other isolated subtrees;
[0048] Based on the expansion priority score, attempts are made to connect unexplored nodes or the starting subtree starting from the endpoint subtree. If the connection fails, additional sampling points are added in the affected area or the path is switched to another path in the candidate path set.
[0049] The present invention has the following beneficial effects:
[0050] 1. By introducing a dynamic connectivity visualization construction method based on a safe time interval and a spatiotemporal topological equivalence evaluation mechanism, the system can maximize the reuse of existing exploration information when the environment changes dynamically, and only make incremental adjustments to the affected local areas, avoiding complete replanning, significantly reducing computational overhead, and achieving rapid response.
[0051] 2. By integrating higher-order control barrier function theory, the system provides mathematically provable safety guarantees for power grid inspection equipment. The system can verify in real time during the planning process whether the path meets multiple safety constraints (such as safe distance from high-voltage equipment, electromagnetic field exposure limits, and flight airspace restrictions), ensuring that the generated path remains safe throughout the entire execution process and avoiding potential safety hazards that may arise from post-event inspections in traditional methods.
[0052] 3. By applying linear quadratic regulators and adaptive dynamic programming techniques to the path planning process, the system can consider the dynamic characteristics (constraints such as velocity, acceleration, and angular velocity) of the UAV or robot during the planning stage. A tangent decomposition algorithm simplifies complex planning problems by breaking them down into multiple simple steps that consider a single key obstacle. Each step verifies the dynamic feasibility, ensuring that the planned path can be directly executed without significant subsequent adjustments.
[0053] 4. By employing a graph neural network to learn edge expansion priorities, the path search process is guided, reducing redundant sampling and collision detection. Combining a top-priority search strategy and adjustable termination conditions, the system can flexibly balance computation time and path quality according to task requirements. For time-sensitive tasks, feasible solutions can be obtained quickly, while for tasks with high quality requirements, further optimization is possible.
[0054] 5. Based on UTVD-based spatiotemporal topological equivalence judgment, the system can identify and maintain multiple candidate paths with different topologies. When the main path fails due to dynamic obstacles, it can quickly switch to an alternative path without having to plan from scratch. Simultaneously, the system can perform parallel optimization of paths of different topological categories, selecting the path with the lowest cost for execution, thus improving overall robustness.
[0055] 6. Design a graph-based path repair algorithm. When local environmental changes cause partial failure of the original path, the system only prunes the smallest affected nodes and edges, and then quickly restores the path integrity by reconnecting the broken subtrees. Combined with spatial-temporal corridor expansion technology, this provides greater flexibility for path optimization, improving path smoothness and execution efficiency while ensuring safety.
[0056] By achieving the above technical objectives, this invention will provide an efficient, safe, and robust dynamic path planning solution for application scenarios such as power grid inspection drones, distribution line inspection robots, and substation intelligent inspection systems, significantly improving the level of intelligent operation and maintenance of power grids. Attached Figure Description
[0057] Figure 1 This is a flowchart of a spatiotemporal dynamic path planning method for power grids that integrates GNN-UTVD-CBF proposed in this invention. Detailed Implementation
[0058] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0059] In one embodiment of the present invention, a spatiotemporal dynamic path planning method for power grids integrating GNN-UTVD-CBF is provided, such as... Figure 1 As shown, it includes the following steps:
[0060] S1. Based on the three-dimensional model of the power grid environment and dynamic obstacle information, combined with equipment dynamic constraints, calculate the safe time interval of potential path edges;
[0061] Preferably, in step S1, the step of calculating the safe time interval of potential path edges includes:
[0062] Static obstacles in the power grid environment are mapped as a 3D occupancy grid map, and dynamic obstacles are modeled as bounded ellipsoids with trajectory prediction information.
[0063] Establish a set of safety constraints that include minimum safe distance, electromagnetic field exposure intensity, airspace restrictions, and equipment dynamic parameters;
[0064] Based on the device dynamic parameters, the minimum passage time of potential path edges is determined, the collision time interval between the potential path edges and dynamic obstacles is calculated, and the collision time interval is removed from the planning cycle to determine the safe time interval that meets the passage requirements.
[0065] Specifically, for dynamic obstacles with time-varying characteristics, such as ground work vehicles, flocks of birds in the air, temporary construction areas, and mobile hoisting equipment, the system models them as bounded ellipsoids with trajectory prediction information. Specifically, considering the uncertainty of the dynamic obstacle's motion, a time-varying probability distribution or worst-case boundary is used to determine the geometric dimensions of the bounded ellipsoid, and its center point trajectory within the planning period is described by a polynomial function. For example, the... A dynamic obstacle at any time The position can be represented as Polynomial equations of order:
[0066] ;
[0067] in, These are the polynomial coefficients.
[0068] To ensure the safe operation of inspection equipment in complex power grid environments, this embodiment establishes a multi-dimensional set of safety constraints. This set explicitly includes the following parameter indicators:
[0069] Minimum safe distance constraint: The minimum Euclidean distance that must be maintained between equipment and live high-voltage equipment to prevent discharge accidents;
[0070] Electromagnetic field exposure intensity constraint: Limit the time that equipment stays in a strong electromagnetic field area or prohibit it from entering a specific high field strength area to prevent electronic components from failing;
[0071] Airspace and right-of-way restrictions: No-fly zones or restricted areas defined according to power grid operation specifications;
[0072] Equipment dynamic parameters: including the maximum permissible linear velocity of the equipment. Maximum acceleration And the maximum angular velocity.
[0073] Based on the above modeling, for any two nodes in the environment (denoted as nodes 1, 2, and 3), and nodes Potential path edges formed between ) Calculate its safe time interval The process consists of the following three sub-steps:
[0074] (1) Determining the minimum passage time: Based on the dynamic parameters of the equipment, a trapezoidal velocity profile model is used to approximate the calculation of the equipment's passage time through the edge. The minimum time required. Assuming the device starts from rest and accelerates to its maximum speed. Accelerate to maximum speed It then moves at a constant speed, and finally accelerates to its maximum speed. Decelerate to a stop. Define the sides. The length is Then the minimum passage time The calculation is as follows:
[0075] ;
[0076] This formula provides the necessary physical lower bound for subsequent time window selection.
[0077] (2) Calculate the collision time interval to determine the edge To determine when dynamic obstacles will block the space, this embodiment employs the Minkowski method to construct the spatial bounding volume. Specifically, it considers the edges... The geometry of the device and the safety envelope sphere are subjected to Minkowski sum operations to obtain the edges. The space encloses the cuboid. Then, the boundary equations of this enclosed cuboid are combined with the polynomial trajectory equations of the dynamic obstacle. By solving the system of equations, the first... The moment a dynamic obstacle enters the space surrounding the cuboid. and the moment of leaving the space surrounding the cuboid Define the time period. For the opposite side of the obstacle Collision time interval Traverse all environments Given a dynamic obstacle, calculate the union of all collision time intervals to obtain the edge. Total collision time interval set
[0078] ;
[0079] (3) The total planning period for the safety time interval setting system is determined to be... Remove the aforementioned set of total collision time intervals from the overall planning cycle. This yields a series of discontinuous time periods. These time periods are then filtered, retaining only those with a duration greater than or equal to the minimum travel time. The time period. The final set of time periods is the edge. safe time interval .
[0080] Through the above steps, this embodiment transforms the geometric path in space into a spatiotemporal path with a clear passage time window, ensuring that the subsequently planned path can avoid all dynamic obstacles in the time dimension and conforms to the physical movement capabilities of the device.
[0081] S2. Based on the security time interval, sampling and connection are performed in free space to construct a dynamic connection view;
[0082] Preferably, in step S2, the step of constructing the dynamically linked view includes:
[0083] Heuristic sampling is performed in free space to obtain sampling points, and the guard points visible to the sampling points are identified;
[0084] Obtain the safe time interval of the connecting edge between the sampling point and the guard point, and determine whether the safe time intervals of the connecting edge overlap.
[0085] If the security time intervals overlap, a connection relationship is established between the sampling point and the guard point, and the sampling point is added to the dynamic connection view.
[0086] Preferably, the step of constructing a dynamically linked view further includes:
[0087] Using the UTVD criterion, we evaluate whether the paths formed by newly added connection points and neighboring nodes are topologically equivalent in the spatiotemporal domain.
[0088] If the new path and the existing path belong to the same UTVD class, only the nodes with lower costs are retained and updated to the dynamic connection view; if they belong to different UTVD classes, the connection point is added as a new node.
[0089] Specifically, to efficiently explore feasible regions in complex power grid environments, the system first performs heuristic sampling in three-dimensional free space. Unlike completely random sampling, this embodiment employs a guided sampling strategy, prioritizing the generation of sampling points in sparse regions with a low "connector-guardian" ratio. Specifically, the system identifies "guard points" (typically known critical nodes or obstacle boundary points) and "connector points" (intermediate nodes connecting different guard points) in the current dynamic connectivity visualization. The system calculates the node density within local regions and increases the sampling probability in regions with density below a preset threshold. For each generated sampling point... The system uses a line-of-sight detection algorithm to identify the set of all guard points within its visible range.
[0090] Determining sampling points and its visible guard points Subsequently, the system not only needs to determine whether the two are connected in geometric space, but also must verify their feasibility in the time dimension. The system acquires sampling points. With the guard point Connecting edges safe time interval This safe time interval The calculations in step S1 determine one or more time windows that allow passage. The system further examines the sampling points. When used as a connecting bridge, is there any overlap in the safety time interval between its incoming and outgoing edges? If the connecting edge Safety time interval and sampling point If the safety time intervals of other associated edges have a non-empty intersection, then the connection is determined to be valid in time and space. At this point, the system establishes sampling points. With the guard point The connection relationship, and the sampling points It is added as a valid node to the dynamic connection view.
[0091] To prevent the graph size from growing exponentially with the increase of sampling points, this embodiment introduces the Unified Temporal Visibility Deformation (UTVD) criterion to evaluate and prune newly added nodes. For newly generated paths and existing neighboring paths in the graph, the system first parameterizes them as time-dependent trajectory functions. If these two trajectories can be transformed into each other through continuous time deformation, and during the deformation process, the line segment connecting the corresponding points of the two trajectories does not collide with any dynamic obstacles throughout the entire planning period, then the two paths are determined to be topologically equivalent, i.e., belonging to the same UTVD class. In specific verification, the system discretizes the two paths into a timestamp sequence. For each corresponding time point, a line segment connecting the positions of the two paths is constructed, and it is detected whether the line segment experiences a sweep collision under the influence of dynamic obstacle movement. If the connecting line segments at all time points do not collide, then the UTVD equivalence condition is satisfied.
[0092] Based on the above evaluation results, the system executes the following update strategy: If the path formed by the newly added connection point is determined to be topologically equivalent to an existing path, it means that both bypass dynamic obstacles in the same way. In this case, the system compares the costs of the two paths, retaining only the node with the lower cost and its connection relationships, and discarding redundant nodes with higher costs. If the new path and the existing path do not belong to the same UTVD class, it means that a completely new feasible topological passage has been discovered. In this case, the system retains the connection point as a new independent node in the dynamic connection view.
[0093] In this way, the system preserves all candidate paths with unique topological features while ensuring the sparsity of the graph structure, providing a wealth of alternative solutions for responding to dynamic environmental changes.
[0094] S3. Utilize graph neural networks to predict the expansion priority of edges in the dynamically connected visibility view, and combine this with a unified temporal visibility deformation criterion to perform path search and generate multiple candidate paths.
[0095] Preferably, in step S3, the step of generating multiple candidate paths includes:
[0096] The nodes and edges of the dynamically connected visual graph are embedded and encoded using a graph neural network, and message passing is performed. The spatiotemporal constraint information of the local neighborhood is aggregated, and the expansion priority score of each edge is output.
[0097] Based on the expansion priority score, a depth-first search strategy is adopted to prioritize the expansion of edges that satisfy the safety time interval constraint until they are connected to the starting point or an existing subtree to generate an initial path.
[0098] Based on the UTVD criteria, the topology category of the initial path is identified, and paths of different topology categories are retained to form a candidate path set.
[0099] Specifically, the system performs initial embedding encoding on nodes and edges in the dynamically connected view. For nodes, their 3D position coordinates are mapped to high-dimensional feature vectors; for edges, their length, the duration of the safety time interval, and the relative positions of the connecting endpoints are encoded into edge feature vectors. Subsequently, multiple rounds of message passing are executed. In each iteration, each node aggregates the feature information of its neighboring nodes and associated edges, and each edge also aggregates the feature information of its connected nodes. After rounds of iteration, the network output layer uses the final edge embedding features to calculate the expansion priority score for each edge through a fully connected layer.
[0100] After obtaining the priority score for each edge, the system uses an improved depth-first search strategy to generate an initial path. The search process starts from the target node and expands backwards towards the starting node. In each search step, the system sorts all adjacent edges of the current node according to the priority scores and prioritizes trying the edge with the highest score. When attempting to expand an edge, the system must re-verify the edge's safe time interval constraint. Specifically, it checks whether the time it takes for the device to arrive at the edge falls within the safe time interval allowed for passage along that edge. If the time constraint is met, the edge is added to the current path and the recursive search continues; if the constraint is not met or a dead end is encountered, the process backtracks to the previous node and selects the next highest priority edge to continue trying. This process continues until a successful connection to the starting point is achieved or the edge is merged with an existing explored subtree, thus forming a complete initial path.
[0101] To construct a robust set of candidate paths, the system maintains a path database in real time during the search process. Whenever a new initial path is generated, the system applies the UTVD criterion to identify its topology class. The system compares the new path with existing paths in the database. If the new path and a path in the database are determined to belong to the same UTVD class, their cumulative costs are compared, and only the path with the lower cost is added to the database. If the new path and none of the existing paths in the database belong to the same UTVD class, the path is considered to represent a novel obstacle avoidance strategy. In this case, the system adds the path as a new independent candidate path to the set. By repeating the above search and filtering process, a set of candidate paths with diverse topologies, all satisfying dynamic and safety constraints, is finally output, providing diverse initialization schemes for subsequent trajectory optimization.
[0102] Through the above steps, a deep learning model is used to quickly identify potential optimal paths from a complex dynamic connectivity view, and to ensure that the path set has sufficient topological diversity to cope with environmental uncertainties.
[0103] S4. During the path search process, the safety of the extended edge is verified by combining the linear quadratic regulator and the control barrier function.
[0104] Preferably, in step S4, the step of performing security verification on the extended edge includes:
[0105] The system is locally linearized near the current state, the algebraic Riccati equation is solved to obtain the feedback gain matrix, and the state trajectory is generated based on the linear quadratic regulator control law.
[0106] For each time step of the state trajectory, calculate the value of the higher-order control barrier function;
[0107] If the higher-order control barrier function value satisfies the safety condition at all time steps, the extended edge is determined to be feasible; otherwise, the extension is terminated and marked as infeasible.
[0108] Specifically, for the nonlinear dynamic system of power grid inspection equipment, its state equation is generally expressed as: To quickly assess control feasibility during edge expansion, the system first checks the current state. The nonlinear model is then subjected to a first-order Taylor expansion, and local linearization is performed. Specifically, the system calculates the nonlinear functions respectively. Regarding the state vector and control input vector The Jacobian matrix is used to obtain the linearized state matrix. and input matrix
[0109] ;
[0110] Based on the linear model, the system constructs a linear quadratic regulator problem and defines the state deviation weight matrix. and control input weight matrix By solving the algebraic Riccati equation:
[0111] ;
[0112] The system obtains a symmetric positive definite solution matrix. Then the optimal feedback gain matrix is calculated. To further improve computational efficiency, the system employs a caching mechanism in its engineering implementation. The calculated feedback gain matrix is then used... An index is created for the corresponding local state point and stored in a hash table. When a similar state is encountered again in a subsequent search, the cached gain matrix is directly called to avoid repeatedly solving the Riccati equation.
[0113] While generating the predicted state trajectory, the system introduces higher-order control obstacle function theory to verify the safety of the trajectory point by point. For safety constraints in power grid scenarios, such as safety distance constraints with high-voltage equipment, the obstacle function is defined as follows:
[0114] ;
[0115] in For equipment location, Location of the obstacle. For an additional safety margin, It is the radius of the equipment. It is the radius of the obstacle.
[0116] Construct higher-order control barrier functions:
[0117] ;
[0118] in , For class K functions, It is a higher-order control barrier function of order r. It is the first The time derivative of the higher-order control barrier function. It is relative.
[0119] During the edge expansion process, the system simulates the state trajectory x(t) under LQR control at each time step. calculate:
[0120] ;
[0121] It is the first higher-order control barrier function at time ,state The value at that location, It is the state at the i-th time step. It is the i-th discrete time step. It is a barrier function Along the system dynamics drift term The Lie derivative of order, It is along the system dynamics control term The Lie derivative operator.
[0122] like If the condition holds for all i, then the edge satisfies the safety constraint; otherwise, the edge extension terminates at the first violation of the constraint, and the edge is marked as infeasible.
[0123] Through the above steps, this embodiment transforms complex dynamic security constraint checks into efficient algebraic operations, achieving millisecond-level online security verification.
[0124] S5. Generate a spatial-temporal corridor for the candidate path based on the security time interval and perform an expansion process;
[0125] Preferably, in step S5, the step of generating a spatial-temporal corridor for the candidate path based on the security time interval and performing dilation processing includes:
[0126] The candidate path is divided into several road segments, and an initial spatial corridor that avoids static obstacles is constructed with the road segments as the axis. The effective time window of the initial spatial corridor is determined based on the safe time interval.
[0127] Select a seed point and expand the boundary of the initial spatial corridor in all directions. During the expansion process, detect whether the new area collides with static or dynamic obstacles.
[0128] If a collision is detected, the expansion in the corresponding direction is stopped, and the effective time window is updated synchronously, ultimately forming a series of expanded space-time corridors.
[0129] Specifically, the system initializes a cuboid or polyhedron aligned along the coordinate axes as an initial spatial corridor. During initialization, the system ensures that the geometric boundaries of the cuboid tightly enclose the corridor segment and that it contains no static obstacles. Simultaneously, the system uses the safety time interval calculated in step S1 as a basis. Determine the effective time window for the initial spatial corridor. This time window defines the time range during which devices are allowed to enter and leave the spatial region, thus expanding the three-dimensional spatial corridor into a four-dimensional spatiotemporal corridor.
[0130] To provide greater smoothness for subsequent B-spline trajectory optimization, the system iteratively expands the initial corridor. The system selects several seed points on the road segment and extends the boundary of a cuboid in all directions of 3D space from these seed points. During each boundary extension, the system performs dual collision detection: static detection checks whether the extended new spatial region overlaps with static obstacles in the 3D occupied grid map; dynamic detection checks whether the extended new spatial region overlaps with obstacles within the current effective time window. Within, whether it intersects spatiotemporally with the predicted trajectory envelope of dynamic obstacles.
[0131] If a collision risk is detected during the extended detection in a certain direction, the boundary extension in that direction is immediately stopped, and the boundary position is fixed to ensure that the corridor remains within free space. If the collision is caused by a dynamic obstacle, the system reassesses the safe passage time within the current spatial range. As the volume of the spatial corridor increases, the potential interaction area between equipment and dynamic obstacles also increases, which may lead to a narrowing of the original safe time interval. The system recalculates and shrinks the effective time window based on the new spatial boundary. This ensures that safety constraints are met at any location and time within the corridor.
[0132] Through the above steps, until no further expansion in any direction is possible or the preset maximum size is reached, a series of interconnected, expanded space-time corridors are ultimately formed. These corridors constitute a safety conduit, strictly constraining the range of the final trajectory generation.
[0133] S6. Within the space-time corridor, B-spline curves are used to parameterize the path, and the final trajectory is solved through multi-objective optimization.
[0134] Preferably, in step S6, the step of solving the final trajectory through multi-objective optimization includes:
[0135] Construct a multi-objective optimization function that includes control cost, dynamic feasibility cost, corridor constraint cost, and dynamic obstacle collision cost;
[0136] The control point positions of the B-spline curve are iteratively adjusted using the gradient descent method to minimize the multi-objective optimization function;
[0137] The quick smoother is used to perform a sliding window search on the optimized trajectory. If there is a collision-free path between non-adjacent control points, the intermediate redundant control points are deleted and the trajectory is updated.
[0138] Specifically, to ensure the continuity and local controllability of the trajectory, this embodiment uses a uniform B-spline curve to parameterize the path within the generated time-space corridor. Let the trajectory... Depend on Control points sum order Definition. Utilizing the convex hull property of B-spline curves, where the curve always lies inside the convex hull formed by its control points, the generated trajectory naturally satisfies static obstacle avoidance constraints as long as the control points are constrained within the space-time corridor generated in step S5.
[0139] To obtain high-quality trajectories, the system constructs a comprehensive multi-objective optimization function. The function consists of a weighted sum of the following four sub-cost terms:
[0140] ;
[0141] in, These are the weighting coefficients for each item.
[0142] The specific definitions of each sub-cost term are as follows:
[0143] Control costs ( To ensure trajectory smoothness and reduce mechanical wear, the square integral of the jerk is used as the control cost.
[0144] The cost of dynamic feasibility ( To ensure the trajectory conforms to the physical limits of the device, for speeds exceeding the maximum speed... and maximum acceleration A penalty is imposed on the portion of the velocity and acceleration that exceeds the limit. This cost term is constructed as the L2 norm of the portion of the velocity and acceleration that exceeds the limit.
[0145] Corridor constraint cost ( Even with hard constraints, to guide the trajectory as far away from obstacles as possible, a distance potential function from the control point to the spatial-temporal corridor boundary is defined. This cost term increases sharply as the control point approaches the corridor boundary, thus driving the trajectory towards the corridor center.
[0146] Dynamic obstacle collision cost ( For dynamic obstacles, the Euclidean distance between the trajectory point and the predicted trajectory of the obstacle is calculated. If this distance is less than a safety threshold, a high collision penalty is incurred, forcing the optimizer to adjust the control point position or time allocation to avoid the dynamic collision zone.
[0147] After constructing the objective function, the system uses a gradient-based numerical optimization algorithm to iteratively update the control point positions. Post-processing of the gradient-optimized trajectory may still result in local redundancy. To further improve execution efficiency, the system introduces a fast smoother for trajectory post-processing. The system uses a sliding window mechanism to traverse the optimized trajectory control point sequence. For two non-adjacent control points in the sequence... and (in The system performs visibility and feasibility checks: establishing connections. and Find the straight line segment; verify whether the straight line segment is completely located inside the space-time corridor; verify whether the straight line segment satisfies the dynamic constraints.
[0148] If all the above checks are passed, it means that you can proceed directly from... Exercise to Without going through an intermediary At this point, the system deletes these intermediate redundant control points and... and The paths are directly connected, and then the local trajectories are resampled and interpolated for updates. This process is repeated until no new shortcuts can be found, ultimately outputting a concise and smooth execution trajectory.
[0149] Through the above steps, the coarse discrete path is transformed into a smooth and continuous executable trajectory that satisfies dynamic constraints.
[0150] S7 monitors environmental changes during trajectory execution and performs local cropping and incremental repair on the dynamic connection view.
[0151] Preferably, in step S7, the step of performing local cropping and incremental repair on the dynamic connection view includes:
[0152] Calculate the radius of influence of dynamic obstacles affected by environmental changes, remove only nodes and their associated edges within the radius of influence, and retain the unaffected graph structure;
[0153] Identify disconnected connected components in the dynamic connection visualization after pruning, wherein the connected components include the starting subtree, the ending subtree, and other isolated subtrees;
[0154] Based on the expansion priority score, attempts are made to connect unexplored nodes or the starting subtree starting from the endpoint subtree. If the connection fails, additional sampling points are added in the affected area or the path is switched to another path in the candidate path set.
[0155] Specifically, the system continuously monitors environmental changes during trajectory execution. When the sensing device detects an environmental change, the system first identifies the dynamic obstacle that has changed. To determine the affected graph node range, the system calculates the radius of influence of the dynamic obstacle. This radius of influence is defined as the sum of the obstacle's own geometric dimensions, the device's safe avoidance distance, and the device's maximum displacement per unit response time. Mathematically, this region constitutes a three-dimensional sphere or ellipsoidal space. The system traverses the current dynamic connectivity view, identifying all nodes whose spatial locations fall within this radius of influence. Subsequently, the system performs a pruning operation, permanently removing these affected nodes and the edges directly connected to them from the graph. Simultaneously, the system preserves the remaining graph structure outside the radius of influence, thereby maximizing the reuse of existing environmental exploration information and avoiding redundant calculations.
[0156] After partial pruning, the original complete graph structure may break into several disconnected parts. The system uses a graph traversal algorithm to perform connectivity analysis on the remaining graph structure, classifying it into three categories of connected components:
[0157] Starting point subtree: contains the connected components of the planned starting node or the node where the current device is located;
[0158] Endpoint subtree: contains the connected components of the planned target node;
[0159] Isolated subtree: Other free tiles that do not contain a start point or an end point.
[0160] To reconnect the starting and ending subtrees with minimal computational cost, the system executes an incremental repair strategy based on graph neural network priorities. The repair process employs a reverse growth logic, starting from the boundary nodes of the ending subtree and attempting to extend towards the starting subtree or unexplored regions. When selecting extension edges, the system still prioritizes connecting potential edges with higher scores and those satisfying the current safe time interval constraints of the new environment, based on the edge extension priority scores output by the GNN in step S3. If the ending and starting subtrees can be successfully merged through existing isolated subtrees or direct connections, the path repair is complete, and the system immediately outputs a new feasible path. If the path cannot be established using only existing residual nodes, the system initiates a supplementary sampling mechanism. The system performs localized high-density sampling only within the calculated obstacle radius and its surrounding area, generating new connection points and attempting to integrate them into existing subtrees.
[0161] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A spatiotemporal dynamic path planning method for power grids integrating GNN-UTVD-CBF, characterized in that, include: S1. Based on the three-dimensional model of the power grid environment and dynamic obstacle information, combined with equipment dynamic constraints, calculate the safe time interval of potential path edges; S2. Based on the security time interval, sampling and connection are performed in free space to construct a dynamic connection view; S3. Utilize graph neural networks to predict the expansion priority of edges in the dynamically connected visibility view, and combine this with a unified temporal visibility deformation criterion to perform path search and generate multiple candidate paths. S4. During the path search process, the safety of the extended edge is verified by combining the linear quadratic regulator and the control barrier function. S5. Generate a spatial-temporal corridor for the candidate path based on the security time interval and perform an expansion process; S6. Within the space-time corridor, B-spline curves are used to parameterize the path, and the final trajectory is solved through multi-objective optimization. S7 monitors environmental changes during trajectory execution and performs local cropping and incremental repair on the dynamic connection view.
2. The spatiotemporal dynamic path planning method for power grids integrating GNN-UTVD-CBF as described in claim 1, characterized in that, In step S1, the step of calculating the safe time interval for potential path edges includes: Static obstacles in the power grid environment are mapped as a 3D occupancy grid map, and dynamic obstacles are modeled as bounded ellipsoids with trajectory prediction information. Establish a set of safety constraints that include minimum safe distance, electromagnetic field exposure intensity, airspace restrictions, and equipment dynamic parameters; Based on the device dynamic parameters, the minimum passage time of potential path edges is determined, the collision time interval between the potential path edges and dynamic obstacles is calculated, and the collision time interval is removed from the planning cycle to determine the safe time interval that meets the passage requirements.
3. The spatiotemporal dynamic path planning method for power grids integrating GNN-UTVD-CBF as described in claim 1, characterized in that, In step S2, the steps of constructing the dynamically linked view include: Heuristic sampling is performed in free space to obtain sampling points, and the guard points visible to the sampling points are identified; Obtain the safe time interval of the connecting edge between the sampling point and the guard point, and determine whether the safe time intervals of the connecting edge overlap. If the security time intervals overlap, a connection relationship is established between the sampling point and the guard point, and the sampling point is added to the dynamic connection view.
4. The spatiotemporal dynamic path planning method for power grids integrating GNN-UTVD-CBF as described in claim 3, characterized in that, The steps for constructing a dynamically linked view also include: Using the UTVD criterion, we evaluate whether the paths formed by newly added connection points and neighboring nodes are topologically equivalent in the spatiotemporal domain. If the new path and the existing path belong to the same UTVD class, only the nodes with lower costs are retained and updated to the dynamic connection view; if they belong to different UTVD classes, the connection point is added as a new node.
5. The spatiotemporal dynamic path planning method for power grids integrating GNN-UTVD-CBF as described in claim 1, characterized in that, In step S3, the steps for generating multiple candidate paths include: The nodes and edges of the dynamically connected visual graph are embedded and encoded using a graph neural network, and message passing is performed. The spatiotemporal constraint information of the local neighborhood is aggregated, and the expansion priority score of each edge is output. Based on the expansion priority score, a depth-first search strategy is adopted to prioritize the expansion of edges that satisfy the safety time interval constraint until they are connected to the starting point or an existing subtree to generate an initial path. Based on the UTVD criteria, the topology category of the initial path is identified, and paths of different topology categories are retained to form a candidate path set.
6. The spatiotemporal dynamic path planning method for power grids integrating GNN-UTVD-CBF as described in claim 1, characterized in that, In step S4, the steps for performing security verification on the extended edge include: The system is locally linearized near the current state, the algebraic Riccati equation is solved to obtain the feedback gain matrix, and the state trajectory is generated based on the linear quadratic regulator control law. For each time step of the state trajectory, calculate the value of the higher-order control barrier function; If the higher-order control barrier function value satisfies the safety condition at all time steps, the extended edge is determined to be feasible; otherwise, the extension is terminated and marked as infeasible.
7. The spatiotemporal dynamic path planning method for power grids integrating GNN-UTVD-CBF as described in claim 1, characterized in that, In step S5, the step of generating a spatial-temporal corridor for the candidate path based on the security time interval and performing dilation processing includes: The candidate path is divided into several road segments, and an initial spatial corridor that avoids static obstacles is constructed with the road segments as the axis. The effective time window of the initial spatial corridor is determined based on the safe time interval. Select a seed point and expand the boundary of the initial spatial corridor in all directions. During the expansion process, detect whether the new area collides with static or dynamic obstacles. If a collision is detected, the expansion in the corresponding direction is stopped, and the effective time window is updated synchronously, ultimately forming a series of expanded space-time corridors.
8. The spatiotemporal dynamic path planning method for power grids integrating GNN-UTVD-CBF as described in claim 1, characterized in that, In step S6, the steps for solving the final trajectory through multi-objective optimization include: Construct a multi-objective optimization function that includes control cost, dynamic feasibility cost, corridor constraint cost, and dynamic obstacle collision cost; The control point positions of the B-spline curve are iteratively adjusted using the gradient descent method to minimize the multi-objective optimization function; The quick smoother is used to perform a sliding window search on the optimized trajectory. If there is a collision-free path between non-adjacent control points, the intermediate redundant control points are deleted and the trajectory is updated.
9. The spatiotemporal dynamic path planning method for power grids integrating GNN-UTVD-CBF as described in claim 1, characterized in that, In step S7, the steps of performing local pruning and incremental repair on the dynamically connected view include: Calculate the radius of influence of dynamic obstacles affected by environmental changes, remove only nodes and their associated edges within the radius of influence, and retain the unaffected graph structure; Identify disconnected connected components in the dynamic connection visualization after pruning, wherein the connected components include the starting subtree, the ending subtree, and other isolated subtrees; Based on the expansion priority score, attempts are made to connect unexplored nodes or the starting subtree starting from the endpoint subtree. If the connection fails, additional sampling points are added in the affected area or the path is switched to another path in the candidate path set.