A path planning method for a wind and solar power plant substation inspection robot

Abstract

This invention provides a path planning method for a substation inspection robot in a wind and solar power plant. The method includes: S1, acquiring various data of the inspection scene collected by the monitoring system / robot sensors and the SCADA system of the wind and solar power plant; S2, processing the various data to obtain processed data; S3, inputting the processed data into a pre-trained multi-objective path optimization model to determine the optimal inspection route that satisfies physical constraints; S4, determining whether the inspection robot has failed to reach the preset target point or encountered new obstacles on the optimal inspection route; if so, executing steps S1 to S3 until the inspection robot reaches the preset target point and completes the inspection task. This method can effectively adapt to the dynamic and complex environment of substations, significantly improving inspection efficiency and equipment operation safety.

Description

Technical Field

[0001] This invention relates to the field of intelligent substation inspection, and more specifically, to a path planning method for a substation inspection robot in a wind and solar power plant. Background Technology

[0002] The equipment layout of wind and solar power plant substations is complex and the environment is dynamically changing. Efficient path planning of inspection robots is one of the core technologies to ensure the safe operation of equipment and improve inspection efficiency. Inspection robots need to cover all inspection points, avoid dynamic obstacles such as temporary workers and mobile equipment, meet kinematic constraints such as turning radius and speed limits, and ensure a safe distance from equipment. This places extremely high demands on the rationality and adaptability of the path.

[0003] However, substations are typical dynamic and unstructured environments with densely distributed equipment and diverse and variable obstacles, making it extremely difficult to establish a path planning model that can adapt to all scenarios. Summary of the Invention

[0004] In view of this, the present invention discloses a path planning method for a substation inspection robot in a wind and solar power plant, which can at least effectively adapt to the dynamic and complex environment of the substation and significantly improve inspection efficiency and equipment operation safety.

[0005] Specifically, the present invention is achieved through the following technical solutions:

[0006] This application proposes a path planning method for a substation inspection robot in a wind and solar power plant, the method comprising:

[0007] S1 acquires various data from inspection scenarios collected by monitoring systems / robot sensors and SCADA systems at wind and solar power plants;

[0008] S2, perform data processing on the various data to obtain processed data;

[0009] S3, input the processed data into a pre-trained multi-objective path optimization model to determine the optimal inspection route that satisfies the physical constraints;

[0010] S4, determine whether the target point has not been reached or whether new obstacles have appeared on the optimal inspection route of the inspection robot;

[0011] If so, proceed with steps S1 to S3 until the inspection robot reaches the preset target point and completes the inspection task.

[0012] Optionally, the multi-objective path optimization model performs the following steps:

[0013] Dynamically predict the speed and direction angle of obstacles using a GRU neural network;

[0014] A grid modeling algorithm is used to divide the inspection environment and update the grid status to determine a high-precision dynamic environment spatial model.

[0015] A global path search is performed using a heuristic search algorithm, with a safety distance weight introduced.

[0016] By embedding kinematics and safety constraints and combining physical information reinforcement learning, local dynamic planning and obstacle avoidance are completed, and a preliminary planned path that satisfies physical constraints is output.

[0017] The preliminary planned path is input into the multi-objective optimization and solution stage. Objective functions are established with path length, energy consumption, and smoothness as optimization objectives. The PSO algorithm is used to solve for minimizing the total loss function, thereby obtaining the optimal inspection route that satisfies the physical constraints.

[0018] Optionally, a grid modeling algorithm is used to divide the inspection environment and update the grid state to determine a high-precision dynamic environment spatial model, including:

[0019] First, initialize a static grid map, then combine it with dynamic obstacle trajectories, and obtain the predicted spatial state through grid space modeling and updating;

[0020] Obtain the spatial state values ​​measured by the sensor, and calculate the spatial state loss and obstacle position loss respectively. The total loss is obtained by weighted summation.

[0021] If the total loss does not converge, backpropagation is used to update the weights of the raster space modeling model until the loss converges, and finally a high-precision environmental space model is obtained.

[0022] The grid space modeling model includes a GRU dynamic prediction submodule, which performs the following steps:

[0023] The control GRU dynamic prediction submodule receives input features, processes them through the GRU hidden layer and fully connected layer, and outputs predicted motion parameters to obtain the predicted spatial state.

[0024] Optionally, the spatial state loss can be calculated using the following formula:

[0025] ;

[0026] in, This represents the loss of spatial state. This represents the number of sampling points; These are the predicted spatial state values; These are the spatial state values ​​measured by the sensor.

[0027] The obstacle position loss is calculated using the following formula:

[0028] ;

[0029] in, Predict loss based on obstacle location; The timing length; , These are the predicted obstacle coordinates; These are the measured coordinates of the obstacle;

[0030] The total loss is calculated using the following formula.

[0031]

[0032] Among them, among them, , These are the spatial state loss weights and obstacle position loss weights, respectively.

[0033] Optionally, the multi-objective path optimization model includes a global path planning module, which performs the following steps:

[0034] Initialize the start / end point and search list, then calculate and expand the minimum cost node using the node cost formula until the target point is reached, and determine the global reference path;

[0035] Initialize the robot state by inputting the state space information (position / velocity / environmental state / global path) into the deep reinforcement learning module. The state space information includes position, velocity, environmental state, and global reference path.

[0036] The Actor network in the deep reinforcement learning module is controlled to make decisions and generate actions. After the actions are executed, the state is updated. Then, the multi-objective reward is calculated by combining the verification results of kinematic constraints and safe distance constraints. After the Q value is evaluated by the Critic network, the network parameters are updated through policy gradient.

[0037] Determine whether the inspection robot has reached the target point or violated physical constraints;

[0038] If the inspection robot reaches the target point without violating physical constraints, the final optimized path is determined.

[0039] Optionally, the node cost formula is:

[0040] ;

[0041] in, For nodes The total cost; From the starting point to the node The actual cost; For nodes Estimated cost of Manhattan distance to target point; For heuristic function weights; For nodes Distance to the nearest device / obstacle; For safety weights, The smaller the penalty, the greater the punishment.

[0042] Optionally, the multi-objective optimization and solution process includes the following steps:

[0043] The path length, energy consumption, and smoothness targets are determined, and the physical constraint loss, path performance loss, and data fitting loss are determined. The physical constraint loss is used to ensure that the path conforms to the actual physical rules, the path performance loss is used to match the optimization targets of path length, energy consumption, and smoothness, and the data fitting loss is used to fit the data characteristics of the actual scenario.

[0044] Calculate the total loss function and use the PSO algorithm to optimize and minimize the total loss function to obtain the optimal inspection route that satisfies the physical constraints;

[0045] The PSO algorithm performs the following steps:

[0046] Initialize the particle swarm and encode its velocity and angular velocity;

[0047] Calculate the particle fitness, update the particle velocity and position, and determine whether the termination condition is met. The termination condition is that the number of iterations reaches the target number or the result converges.

[0048] If the termination condition is not met, the particle fitness is recalculated and the particle's velocity and position are updated until the termination condition is met.

[0049] If the termination condition is met, the optimal particle parameters are output, and the final optimal inspection path is generated based on the optimal particle parameters.

[0050] Alternatively, the path length optimization objective can be calculated using the following formula:

[0051] ;

[0052] in, The start time; The time to reach the target point. For speed, minimize That is, to achieve the shortest path;

[0053] The energy consumption optimization target is calculated using the following formula:

[0054] ;

[0055] in, , These are the energy consumption coefficients, corresponding to linear motion and steering energy consumption, respectively. For speed, Angular velocity, minimize That is, to achieve a low-energy consumption path;

[0056] The path smoothness optimization objective is calculated using the following formula:

[0057] ;

[0058] Where, minimize That is, a path that achieves smooth steering.

[0059] Alternatively, the physical constraint loss can be calculated using the following formula:

[0060] ;

[0061] in, Loss due to physical constraints; , This refers to the actual distance between the robot and the equipment and obstacles.

[0062] The path performance loss is calculated using the following formula:

[0063] ;

[0064] in, This is due to path performance loss; , , These are the weights for length, energy consumption, and smoothness loss, respectively.

[0065] The data fitting loss is calculated using the following formula:

[0066] ;

[0067] in, For data fitting loss; The number of samples; Value for Critic network prediction; The target Q value;

[0068] The total loss function is calculated using the following formula:

[0069] ;

[0070] in, These are the weighting coefficients for each loss term.

[0071] Optionally, kinematic and safety constraints include: action space constraints, kinematic constraints, and safety distance constraints;

[0072] The action space constraints are as follows:

[0073] ;

[0074] in, For speed control, Angular velocity control quantity, Maximum speed, This is the maximum angular velocity;

[0075] The kinematic constraint equations are as follows:

[0076] ;

[0077] ;

[0078] in, , These are the minimum and maximum permissible speeds, respectively. These are the minimum and maximum permissible angular velocities, respectively. This is the maximum acceleration; This is the maximum angular acceleration;

[0079] The safety distance constraint equation is as follows:

[0080] ;

[0081] ;

[0082] in, This refers to the minimum safe distance between the inspection robot and the equipment. This refers to the minimum safe distance between the inspection robot and the obstacle; For the inspection robot in Coordinates along the axis; For substation equipment at all times of Axis coordinate position; For the obstacle at time of Axis coordinate position; For the inspection robot in Coordinates along the axis; For substation equipment at all times of Axis coordinate position; For the obstacle at time of Axis coordinate position.

[0083] The path planning method for a wind and solar power substation inspection robot proposed in this application adopts a collaborative planning approach that combines physical constraint mechanisms, physical information reinforcement learning, and improved heuristic search algorithms. This aims to improve the safety, efficiency, and environmental adaptability of the inspection path. Addressing the problems of complex equipment layouts and dynamically changing environments in wind and solar power substations, existing traditional mechanism algorithms suffer from poor dynamic adaptability, pure data-driven algorithms lack physical consistency, and fusion algorithms have limited generalization capabilities. This method acquires environmental operation data such as equipment position coordinates and obstacle dynamics from the substation monitoring system and robot sensors. It first constructs a spatial model based on a "GRU neural network + grid modeling algorithm," and then uses spatial constraint equations and dynamic obstacle motion equations... The system employs core mechanism equations to dynamically predict obstacle motion states and output a high-precision environmental spatial model. A fusion architecture combining "physical information reinforcement learning + improved A* algorithm" is then used. A heuristic search algorithm incorporating safety distance weights completes global path search, while reinforcement learning embedding kinematics and safety constraints enables local dynamic obstacle avoidance, generating a preliminary planned path. Finally, with path length, energy consumption, and smoothness as optimization objectives, a total loss function is constructed, including physical constraint loss, path performance loss, and data fitting loss. The optimal path is then solved using a particle swarm optimization algorithm, ultimately generating a short, low-energy, and highly smooth optimal inspection path. This effectively adapts to the dynamic and complex environment of substations, significantly improving inspection efficiency and equipment operational safety. Attached Figure Description

[0084] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0085] Figure 1 A flowchart illustrating a path planning method for a wind and solar power plant substation inspection robot provided in this application embodiment;

[0086] Figure 2 A flowchart illustrating the path planning process for a wind and solar power plant substation inspection robot based on a physical information reinforcement learning heuristic search algorithm provided in this application.

[0087] Figure 3 The spatial construction model structure frame for the path planning of the photovoltaic power plant and substation inspection robot provided in this application. Figure 1 ;

[0088] Figure 4 The path planning structure framework for the wind and solar power plant substation inspection robot provided in this application Figure 2 ;

[0089] Figure 5 The path optimization framework for path planning of the wind and solar power plant substation inspection robot provided in this application Figure 3 ;

[0090] Figure 6 The accompanying diagrams show the effect comparison between this application and related technologies. Detailed Implementation

[0091] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of systems and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0092] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0093] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0094] This application relates to the field of intelligent substation inspection, and more specifically, to a path planning method for a substation inspection robot in a wind and solar power plant.

[0095] The equipment layout of wind and solar power plant substations is complex and the environment is dynamically changing. Efficient path planning of inspection robots is one of the core technologies to ensure the safe operation of equipment and improve inspection efficiency. Inspection robots need to cover all inspection points, avoid dynamic obstacles such as temporary workers and mobile equipment, meet kinematic constraints such as turning radius and speed limits, and ensure a safe distance from equipment. This places extremely high demands on the rationality and adaptability of the path.

[0096] However, substations are typical dynamic and unstructured environments with densely distributed equipment and diverse and variable obstacles, making it extremely difficult to establish a path planning model that can adapt to all scenarios.

[0097] Currently, existing path planning methods are mainly divided into three categories, but all have significant limitations. The first category is the traditional mechanism algorithm based on geometric programming. This type of method establishes a mathematical model based on spatial geometric relationships and kinematic constraints, and solves for the optimal path through a search algorithm. Its advantage lies in its clear physical meaning and the path satisfying basic motion constraints. However, this method requires high accuracy in environmental modeling, has poor adaptability to dynamic obstacles, and struggles to balance path length and inspection efficiency. When the environment changes, it needs to be recalculated, resulting in insufficient real-time performance and making it difficult to meet the dynamic inspection needs of complex substations.

[0098] The second category is data-driven algorithms based on reinforcement learning. These methods utilize environmental interaction data to learn the optimal inspection strategy through reinforcement learning algorithms. Their advantage lies in not requiring an accurate environmental model and having strong fitting capabilities for nonlinear and dynamic environments. However, this method is entirely dependent on the quality and quantity of data, requiring a large amount of training data covering the entire scenario. Data on special operating conditions in substations, such as temporary equipment maintenance and extreme weather, is scarce. More importantly, the training results lack physical consistency, potentially generating invalid paths that violate kinematic constraints or safety distance requirements, posing a collision risk, and drastically reducing reliability under unfamiliar operating conditions.

[0099] The third category is simple fusion algorithms, such as heuristic search combined with local obstacle avoidance algorithms. These methods generate a global path using traditional algorithms and then handle dynamic obstacles using local algorithms. However, this simple combination fails to achieve deep collaboration and cannot embed physical rules into the algorithm optimization process in a structured manner, resulting in limited improvement in the model's performance and overall generalization ability in dynamic environments or data-scarce regions.

[0100] In summary, existing technologies cannot effectively balance the safety, efficiency, and environmental adaptability of path planning for inspection robots in wind and solar power plants and substations. Developing a novel path planning method that deeply integrates environmental physical mechanisms, global search algorithms, and data-driven learning capabilities to fundamentally overcome these shortcomings has become a critical technical problem urgently needing to be solved in this field.

[0101] Based on this, this application proposes a path planning method for a substation inspection robot in a wind and solar power plant.

[0102] Please see Figure 1 This application discloses a flowchart illustrating a path planning method for a wind and solar power plant substation inspection robot. Specifically, the path planning method for the wind and solar power plant substation inspection robot proposed in this application includes:

[0103] S1 acquires various data from inspection scenarios collected by monitoring systems / robot sensors and SCADA systems at wind and solar power plants;

[0104] S2, perform data processing on the various data to obtain processed data;

[0105] S3, input the processed data into a pre-trained multi-objective path optimization model to determine the optimal inspection route that satisfies the physical constraints;

[0106] S4, determine whether the preset target point has not been reached or whether a new obstacle has appeared on the optimal inspection route of the inspection robot;

[0107] If so, proceed with steps S1 to S3 until the inspection robot reaches the preset target point and completes the inspection task.

[0108] Specifically, the multi-objective path optimization model performs the following steps:

[0109] Dynamically predict the speed and direction angle of obstacles using a GRU neural network;

[0110] A grid modeling algorithm is used to divide the inspection environment and update the grid status to determine a high-precision dynamic environment spatial model.

[0111] A global path search is performed using a heuristic search algorithm, with a safety distance weight introduced.

[0112] By embedding kinematics and safety constraints and combining physical information reinforcement learning, local dynamic planning and obstacle avoidance are completed, and a preliminary planned path that satisfies physical constraints is output.

[0113] The preliminary planned path is input into the multi-objective optimization and solution stage. Objective functions are established with path length, energy consumption, and smoothness as optimization objectives. The PSO algorithm is used to solve for minimizing the total loss function, thereby obtaining the optimal inspection route that satisfies the physical constraints.

[0114] Among them, a grid modeling algorithm is used to divide the inspection environment and update the grid state to determine a high-precision dynamic environment spatial model, including:

[0115] First, initialize a static grid map, then combine it with dynamic obstacle trajectories, and obtain the predicted spatial state through grid space modeling and updating;

[0116] Obtain the spatial state values ​​measured by the sensor, and calculate the spatial state loss and obstacle position loss respectively. The total loss is obtained by weighted summation.

[0117] If the total loss does not converge, backpropagation is used to update the weights of the raster space modeling model until the loss converges, and finally a high-precision environmental space model is obtained.

[0118] The grid space modeling model includes a GRU dynamic prediction submodule, which performs the following steps:

[0119] The control GRU dynamic prediction submodule receives input features, processes them through the GRU hidden layer and fully connected layer, and outputs predicted motion parameters to obtain the predicted spatial state.

[0120] Specifically, the spatial state loss can be calculated using the following formula:

[0121] ;

[0122] in, This represents the loss of spatial state. This represents the number of sampling points; These are the predicted spatial state values; These are the spatial state values ​​measured by the sensor.

[0123] The obstacle position loss is calculated using the following formula:

[0124] ;

[0125] in, Predict loss based on obstacle location; The timing length; , These are the predicted obstacle coordinates; These are the measured coordinates of the obstacle;

[0126] The total loss is calculated using the following formula.

[0127] ;

[0128] Among them, among them, , These are the spatial state loss weights and obstacle position loss weights, respectively.

[0129] Specifically, the multi-objective path optimization model includes a global path planning module, which performs the following steps:

[0130] Initialize the start / end point and search list, then calculate and expand the minimum cost node using the node cost formula until the target point is reached, and determine the global reference path;

[0131] Initialize the robot state by inputting the state space information (position / velocity / environmental state / global path) into the deep reinforcement learning module. The state space information includes position, velocity, environmental state, and global reference path.

[0132] The Actor network in the deep reinforcement learning module is controlled to make decisions and generate actions. After the actions are executed, the state is updated. Then, the multi-objective reward is calculated by combining the verification results of kinematic constraints and safe distance constraints. After the Q value is evaluated by the Critic network, the network parameters are updated through policy gradient.

[0133] Determine whether the inspection robot has reached the target point or violated physical constraints;

[0134] If the inspection robot reaches the target point without violating physical constraints, the final optimized path is determined.

[0135] The node cost formula is as follows:

[0136] ;

[0137] in, For nodes The total cost; From the starting point to the node The actual cost; For nodes Estimated cost of Manhattan distance to target point; For heuristic function weights; For nodes Distance to the nearest device / obstacle; For safety weights, The smaller the penalty, the greater the punishment.

[0138] Specifically, the multi-objective optimization and solution process involves the following steps:

[0139] The path length, energy consumption, and smoothness targets are determined, and the physical constraint loss, path performance loss, and data fitting loss are determined. The physical constraint loss is used to ensure that the path conforms to the actual physical rules, the path performance loss is used to match the optimization targets of path length, energy consumption, and smoothness, and the data fitting loss is used to fit the data characteristics of the actual scenario.

[0140] Calculate the total loss function and use the PSO algorithm to optimize and minimize the total loss function to obtain the optimal inspection route that satisfies the physical constraints;

[0141] The PSO algorithm performs the following steps:

[0142] Initialize the particle swarm and encode its velocity and angular velocity;

[0143] Calculate the particle fitness, update the particle velocity and position, and determine whether the termination condition is met. The termination condition is that the number of iterations reaches the target number or the result converges.

[0144] If the termination condition is not met, the particle fitness is recalculated and the particle's velocity and position are updated until the termination condition is met.

[0145] If the termination condition is met, the optimal particle parameters are output, and the final optimal inspection path is generated based on the optimal particle parameters.

[0146] The path length optimization objective is calculated using the following formula:

[0147] ;

[0148] in, The start time; The time to reach the target point. , That is, to achieve the shortest path;

[0149] The energy consumption optimization target is calculated using the following formula:

[0150] ;

[0151] in, , These are the energy consumption coefficients, corresponding to linear motion and steering energy consumption, respectively. , Angular velocity, minimize That is, to achieve a low-energy consumption path;

[0152] The path smoothness optimization objective is calculated using the following formula:

[0153] ;

[0154] Where, minimize That is, a path that achieves smooth steering.

[0155] The physical constraint loss is calculated using the following formula:

[0156] ;

[0157] in, Loss due to physical constraints; , This refers to the actual distance between the robot and the equipment and obstacles.

[0158] The path performance loss is calculated using the following formula:

[0159] ;

[0160] in, This is due to path performance loss; These are the weights for length, energy consumption, and smoothness loss, respectively.

[0161] The data fitting loss is calculated using the following formula:

[0162] ;

[0163] in, For data fitting loss; The number of samples; Value for Critic network prediction; The target Q value;

[0164] The total loss function is calculated using the following formula:

[0165] ;

[0166] in, These are the weighting coefficients for each loss term.

[0167] Specifically, kinematic and safety constraints include: action space constraints, kinematic constraints, and safety distance constraints.

[0168] The action space constraints are as follows:

[0169] ;

[0170] in, For speed control, Angular velocity control quantity, Maximum speed, This is the maximum angular velocity;

[0171] The kinematic constraint equations are as follows:

[0172] ;

[0173] ;

[0174] in, , These are the minimum and maximum permissible speeds, respectively. , These are the minimum and maximum permissible angular velocities, respectively. This is the maximum acceleration; This is the maximum angular acceleration;

[0175] The safety distance constraint equation is as follows:

[0176] ;

[0177] ;

[0178] Among them, among them, This refers to the minimum safe distance between the inspection robot and the equipment. This refers to the minimum safe distance between the inspection robot and the obstacle; For the inspection robot in Coordinates along the axis; For substation equipment at all times of Axis coordinate position; For the obstacle at time of Axis coordinate position; For the inspection robot in Coordinates along the axis; For substation equipment at all times of Axis coordinate position; For the obstacle at time of Axis coordinate position.

[0179] Here, this application acquires environmental operation data through a substation monitoring system and robot sensors, including parameter variables required for modeling (such as equipment position coordinates, obstacle dynamic information, robot motion parameters, etc.). The path planning process of this invention is divided into three parts: space construction, path planning, and path optimization.

[0180] Among them, the spatial construction model takes the geometric constraints and dynamic characteristics of the substation's physical space as its foundation and embeds them into the fusion framework of spatial representation network and geometric modeling algorithm. Through algorithmic collaborative learning and adaptation to changes in dynamic obstacles and equipment layout characteristics in the environment, a high-precision substation environmental spatial model is finally output.

[0181] Specifically, this includes the core mechanism equation:

[0182] Based on the spatial geometric relationship of the substation and the distribution pattern of obstacles, the spatial constraint equations are established as follows:

[0183] (1);

[0184] in, for Time coordinates The spatial state value at the location (0 indicates feasible, 1 indicates infeasible); The substation's two-dimensional plane coordinates; For time.

[0185] Equations of motion for dynamic obstacles:

[0186] (2);

[0187] in, for The coordinates of dynamically changing obstacles in real time; The speed at which the obstacle moves; The direction angle of the obstacle's movement; and To dynamically change parameters, this invention uses a gated recurrent unit (GRU) neural network for dynamic prediction.

[0188] Spatial state update equation:

[0189] (3);

[0190] This equation serves as the physical constraint for the spatial construction model. Combined with the grid modeling algorithm, it divides the substation space and achieves a refined representation of the environment.

[0191] It also includes: algorithm fusion and parameter prediction algorithms.

[0192] Specifically, it adopts a fusion architecture of "GRU neural network + grid modeling algorithm" to dynamically predict the speed of dynamic obstacles. and direction angle

[0193] Network input : at each time step The input features include the historical location sequence of obstacles. Current ambient light intensity, personnel density in the inspection area, and time information;

[0194] Network output , and That is, the predicted speed and direction angle of the obstacle at the current moment;

[0195] Grid modeling algorithm: The substation space is divided into a fine grid of 0.1m×0.1m, and the grid status is updated in real time based on the GRU prediction results.

[0196] Network structure: The number of nodes in the input layer equals the number of input features; there are 1-2 GRU layers, each with 64-128 neurons, capturing temporal motion features; the fully connected layer maps the GRU output to the target dimension; the output layer has 2 nodes, outputting... and Activate mapping to .

[0197] It also includes: loss function design.

[0198] Among them, the spatial construction loss function It consists of two parts:

[0199] (4);

[0200] in, This represents the loss of spatial state. This represents the number of sampling points; These are the predicted spatial state values; This represents the spatial state value measured by the sensor.

[0201] (5);

[0202] in, Predict loss based on obstacle location; The timing length; , These are the predicted obstacle coordinates; These are the measured coordinates of the obstacle.

[0203] (6);

[0204] in, , These are the spatial state loss weights and the obstacle position loss weights, respectively; minimized by gradient descent. The GRU network weights are updated, and a high-precision substation environmental spatial model is finally output.

[0205] In this invention, a core path planning model integrating "physical information reinforcement learning + improved A* algorithm" is constructed. The improved A* algorithm enables fast global path search, and the physical information reinforcement learning is combined to dynamically adapt to the dynamic environment and physical constraints to achieve preliminary path planning.

[0206] Among them, the heuristic function of the improved heuristic search algorithm introduces a safe distance weight to avoid paths approaching dangerous areas:

[0207] (7);

[0208] in, For nodes The total cost; From the starting point to the node The actual cost; For nodes Estimated cost of Manhattan distance to target point; The weights for the heuristic function (values ​​range from 1.2 to 1.5). For nodes Distance to the nearest device / obstacle; The safety weight (values ​​range from 0.8 to 1.0). The smaller the penalty, the greater the punishment.

[0209] The global path search process includes:

[0210] A grid map based on a spatial construction model is initialized with a starting point, a target point, and an open / closed list.

[0211] Calculate the total cost for each candidate node Choose the node with the lowest cost to expand;

[0212] Repeat the expansion until the target point is reached, generating a global initial path.

[0213] It also includes: Local dynamic adaptation steps in physical information reinforcement learning; State space of reinforcement learning agents. : Includes the robot's current position Attitude angle ,speed angular velocity Coordinates of the target inspection point Environmental space status Improve the global path node information of the heuristic search algorithm.

[0214] Among them, action space Robot speed control quantity With angular velocity control quantity ,satisfy

[0215] .

[0216] Physical constraint embedding includes:

[0217] Kinematic constraint equations:

[0218] (8);

[0219] (9);

[0220] in, , These are the minimum and maximum permissible speeds, respectively. , These are the minimum and maximum permissible angular velocities, respectively. This is the maximum acceleration; This is the maximum angular acceleration.

[0221] Safety distance constraint equation:

[0222] (10);

[0223] (11);

[0224] in, To maintain the minimum safe distance from the equipment; The minimum safe distance from obstacles.

[0225] reward function Design that integrates physical constraints and mission objectives:

[0226] (12);

[0227] in, Rewards for achieving the target: ; For safety rewards: a positive reward is given when constraints (10) and (11) are satisfied, and a negative reward is given when they are violated; For the motion constraint reward: a positive reward is obtained when the constraint formulas (8) and (9) are satisfied, and a negative reward is obtained when they are violated; For path smoothing rewards: ;

[0228] To align global path with rewards: For Robots and Improvements Global path deviation distance.

[0229] Network Structure: Employs the Deep Deterministic Policy Gradient (DDPG) framework, with the Actor network inputting the state... Output Action Critic network input Output action value .

[0230] The path optimization model is established by using a "physical information loss function + particle swarm optimization (PSO) algorithm" to achieve path optimization. At the same time, path smoothing and energy consumption optimization objectives are introduced to achieve the final optimal path output.

[0231] Specifically, this includes optimizing the objective function.

[0232] The optimization objective function includes: path length optimization objective:

[0233] (13);

[0234] in, The start time; To minimize the time to reach the target point Achieve the shortest path.

[0235] The optimization objective function includes: energy consumption optimization objective:

[0236] (14);

[0237] in, , Here are the energy consumption coefficients, corresponding to linear motion and steering energy consumption respectively, which are minimized. Achieving a low-energy consumption path.

[0238] Path smoothness optimization objective:

[0239] (15);

[0240] minimize Achieve a smooth turning path and reduce mechanical wear on the robot.

[0241] It also includes: collaborative optimization loss function design.

[0242] Specifically, the total loss function It consists of physical constraint loss, path performance loss, and data fitting loss:

[0243] (16);

[0244] in, Loss due to physical constraints; , This represents the actual distance between the robot and the equipment and obstacles.

[0245] (17);

[0246] in, This is due to path performance loss; , , The weights are for length, energy consumption, and smoothness loss, respectively.

[0247] (18);

[0248] in, For data fitting loss; The number of samples; Value for Critic network prediction; The target Q value.

[0249] (19);

[0250] in, These are the weighting coefficients for each loss term.

[0251] This application also employs the PSO algorithm for optimization, transforming the path optimization problem into a multi-objective optimization problem, and using the PSO algorithm to find the optimal solution, specifically including:

[0252] Particle encoding: Each particle corresponds to a set of path parameters (velocity sequence) Angular velocity sequence );

[0253] Fitness function: A higher fitness value indicates a better path;

[0254] Iterative update: The particle positions are continuously optimized using the velocity update formula and position update formula of the particle swarm until the iteration termination condition is met (the number of iterations ≥ 100 or the fitness value converges).

[0255] Output the path parameters corresponding to the optimal particle to generate the final optimal inspection path.

[0256] For example, please refer to Figure 2 , Figure 2 This paper demonstrates the overall architecture and implementation process of the "Path Planning for Wind and Solar Power Plant Substation Inspection Robot Based on Physical Information Reinforcement Learning Heuristic Search Algorithm" described in this invention. The flowchart covers the entire chain of data acquisition, environment modeling, path planning, algorithm solving, and execution feedback, reflecting the inspection path optimization logic that combines data-driven approaches with intelligent algorithms.

[0257] Specifically, the process begins with multi-source data acquisition, using monitoring systems / robot sensors, wind and solar power plant SCADA systems, and other equipment to obtain raw data from the inspection scenario. After data preprocessing and fusion, it simultaneously supports two core modeling tasks: First, it uses GRU neural networks to dynamically predict the speed of obstacles. With direction angle This provides a data foundation for dynamic environmental perception; on the other hand, a grid modeling algorithm is used to divide the inspection environment and update the grid status, and finally integrate and output a high-precision dynamic environmental space model.

[0258] Based on the dynamic environment space model, the process enters the fusion path planning stage: global path search is performed through a heuristic search algorithm, a safety distance weight is introduced to ensure path safety, kinematics and safety constraints are embedded, and local dynamic planning and obstacle avoidance are completed by combining physical information reinforcement learning, and finally a preliminary planned path that meets the physical constraints is output.

[0259] The initial path planning then moves into the multi-objective optimization and solution phase, focusing on path length. Energy consumption , smoothness To optimize the objective, an objective function is established, and the PSO algorithm is used to minimize the total loss function. This yields the optimal inspection path that balances "short distance, low energy consumption, and high smoothness".

[0260] The optimal path enters the execution and feedback phase, which determines in real time whether the target point has been reached or a new obstacle has appeared: if the termination condition is not met, the process returns to the data collection phase to re-trigger environmental modeling and path adjustment; if the condition is met, the process ends and the inspection task is completed.

[0261] This invention presents the entire technical path from data acquisition to path implementation, highlighting its dynamic environmental adaptability and multi-objective optimization characteristics, and providing technical support for efficient and safe inspection operations.

[0262] For example, please refer to Figure 3 , Figure 3 It is a spatial construction model structure frame for path planning of inspection robots for wind and solar power plants and substations. Figure 1 This model is a typical multi-module collaborative spatial state prediction architecture.

[0263] When the model is working, the data acquisition and preprocessing module first operates: environmental parameters (light intensity, personnel density) are collected through the sensor system, and after constructing the time series, the data is cleaned and aligned to obtain the model input features X(t).

[0264] Subsequently, the core of the model operation is launched: first, the static grid map is initialized, and combined with the dynamic obstacle trajectory, the predicted spatial state is obtained through grid space modeling and updating; at the same time, the GRU dynamic prediction submodule receives input features, processes them through GRU hidden layers, fully connected layers, etc., and outputs predicted motion parameters to provide historical position sequence support for spatial state updates, and finally obtains the predicted spatial state.

[0265] The training and optimization of the model are completed by the model validation and loss calculation module: after collecting the real values, the spatial state loss and obstacle position loss are calculated separately, and the total loss is obtained by weighted summation; if the loss does not converge, the model weights are updated by backpropagation until the loss converges, and finally a high-precision environmental space model is obtained, which has both dynamic adaptability and prediction accuracy.

[0266] For example, please refer to Figure 4 , Figure 4 It is a spatial construction model structure frame for path planning of inspection robots for wind and solar power plants and substations. Figure 2 The model consists of a global path planning module and a deep reinforcement learning local path optimization module connected in series.

[0267] The workflow is as follows: The environmental perception and map building module first outputs a raster map / dynamic environment information, which is then input into the global path planning module; this module first initializes the start / end point and search list, and then uses the node cost formula... The minimum cost nodes are calculated and expanded until the target point is reached, and the global reference path is output. Then, the robot state (position / attitude / velocity) is initialized, and the state space information (position / velocity / environment state / global path) is input into the deep reinforcement learning module. The Actor network makes decisions to generate actions, updates the state after the actions are executed, and calculates the multi-objective reward by combining the verification results of kinematic constraints and safe distance constraints. After the Q value is evaluated by the Critic network, the network parameters are updated through policy gradient.

[0268] The collaborative driving mechanism is reflected in the connection logic between global and local modules. The global module provides path guidance, while the local modules dynamically optimize actions based on real-time status. At the same time, through the feedback mechanism of "data reward (state compliance) + goal orientation", the path is made to conform to global guidance while meeting the robot's motion constraints and environmental safety requirements, ultimately achieving a unity between the rationality of global planning and the safety of local motion.

[0269] For example, please refer to Figure 5 , Figure 5 It is a spatial construction model structure frame for path planning of inspection robots for wind and solar power plants and substations. Figure 3 This process is a multi-objective collaborative optimization path planning architecture.

[0270] When starting the process, first clarify the three core objectives of path optimization: path length. Energy consumption , smoothness Based on these objectives, a collaborative optimization loss function is designed, which includes three types of losses: physical constraint loss. This is used to ensure that the path conforms to the actual physical rules; path performance loss This is used to match optimization objectives such as path length, energy consumption, and smoothness; data fitting loss. Data features used to fit real-world scenarios.

[0271] Then the total loss function is calculated. The PSO algorithm is used for optimization. First, the particle swarm is initialized, and the velocity is... With angular velocity Encode the particles; then calculate their fitness using the formula: Then update the particle's velocity and position, and then check whether the termination condition is met—the number of iterations reaches 100 or the result converges. If not, return to the fitness calculation step and repeat the operation.

[0272] Once the termination condition is met, the optimal particle parameters are output, and the final optimal inspection path is generated accordingly, thus ending the entire process. This process achieves efficient path planning that balances path length, energy consumption, and smoothness by combining a multi-objective loss function with the PSO algorithm.

[0273] Please see Figure 6 , Figure 6 By using 100 sets of time-series data, the performance differences between the physical information neural network algorithm and the traditional algorithm in path planning were intuitively compared.

[0274] Figure 6 The left image shows a comparison of path lengths. The solid black line represents the path length of the traditional algorithm, which is generally over 600 meters and fluctuates significantly; the gray dotted line represents the path length of the physical information neural network algorithm, which is generally around 500 meters, significantly shorter than the traditional algorithm.

[0275] Figure 6 The right-hand figure in the image shows a comparison of path smoothness. The solid black line represents the smoothness of the traditional algorithm, with large fluctuations and mostly below 0.6; the gray dotted line represents the smoothness of the physical information neural network algorithm, mostly close to 1, indicating superior smoothness. This comparison demonstrates that the physical information neural network algorithm, by integrating physical constraints and deep learning, effectively overcomes the shortcomings of traditional algorithms, such as long paths, high energy consumption, and poor smoothness, achieving more efficient path planning.

[0276] The path planning method for a wind and solar power substation inspection robot proposed in this application adopts a collaborative planning approach that combines physical constraint mechanisms, physical information reinforcement learning, and improved heuristic search algorithms. This aims to improve the safety, efficiency, and environmental adaptability of the inspection path. Addressing the problems of complex equipment layouts and dynamically changing environments in wind and solar power substations, existing traditional mechanism algorithms suffer from poor dynamic adaptability, pure data-driven algorithms lack physical consistency, and fusion algorithms have limited generalization capabilities. This method acquires environmental operation data such as equipment position coordinates and obstacle dynamics from the substation monitoring system and robot sensors. It first constructs a spatial construction model based on a "GRU neural network + grid modeling algorithm," and then, based on spatial constraint equations and dynamic... The system employs core mechanism equations such as the obstacle motion equation to dynamically predict obstacle motion states and output a high-precision environmental spatial model. Then, a fusion architecture of "physical information reinforcement learning + improved heuristic search algorithm" is adopted. A heuristic search algorithm incorporating safety distance weights completes global path search, while reinforcement learning embedding kinematics and safety constraints enables local dynamic obstacle avoidance, generating a preliminary planned path. Finally, with path length, energy consumption, and smoothness as optimization objectives, a total loss function is constructed, including physical constraint loss, path performance loss, and data fitting loss. The optimal path is solved using a particle swarm optimization algorithm, ultimately generating a short, low-energy, and highly smooth optimal inspection path, effectively adapting to the dynamic and complex environment of substations.

[0277] Finally, it should be noted that although this specification contains many specific implementation details, these should not be construed as limiting the scope of any invention or the scope of the claims, but rather are primarily used to describe the features of specific embodiments of a particular invention. Certain features described in the various embodiments of this specification may also be implemented in combination in a single embodiment. On the other hand, various features described in a single embodiment may also be implemented separately in various embodiments or in any suitable sub-combination. Furthermore, while features may function in certain combinations as described above and even initially claimed in this way, one or more features from a claimed combination may be removed from that combination in some cases, and a claimed combination may refer to a sub-combination or a variation of a sub-combination.

[0278] Similarly, although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0279] Thus, specific embodiments of the subject matter have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims may be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings are not necessarily shown in a specific order or sequence to achieve the desired result. In some implementations, multitasking and parallel processing may be advantageous.

[0280] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A path planning method for a substation inspection robot in a wind and solar power plant, characterized in that, The method includes: S1 acquires various data from inspection scenarios collected by monitoring systems / robot sensors and SCADA systems at wind and solar power plants; S2, perform data processing on the various data to obtain processed data; S3, input the processed data into a pre-trained multi-objective path optimization model to determine the optimal inspection route that satisfies the physical constraints; S4, determine whether the preset target point has not been reached or whether a new obstacle has appeared on the optimal inspection route of the inspection robot; If so, then execute steps S1 to S3 until the inspection robot reaches the preset target point and completes the inspection task; The multi-objective path optimization model performs the following steps: Dynamically predict the speed and direction angle of obstacles using a GRU neural network; A grid modeling algorithm is used to divide the inspection environment and update the grid status to determine a high-precision dynamic environment spatial model. A global path search is performed using a heuristic search algorithm, with a safety distance weight introduced. By embedding kinematics and safety constraints and combining physical information reinforcement learning, local dynamic planning and obstacle avoidance are completed, and a preliminary planned path that satisfies physical constraints is output. The preliminary planned path is input into the multi-objective optimization and solution stage. Objective functions are established with path length, energy consumption, and smoothness as optimization objectives. The PSO algorithm is used to solve for minimizing the total loss function to obtain the optimal inspection route that satisfies the physical constraints. Among them, a grid modeling algorithm is used to divide the inspection environment and update the grid state to determine a high-precision dynamic environment spatial model, including: First, initialize a static grid map, then combine it with dynamic obstacle trajectories, and obtain the predicted spatial state through grid space modeling and updating; Obtain the spatial state values ​​measured by the sensor, and calculate the spatial state loss and obstacle position loss respectively. The total loss is obtained by weighted summation. If the total loss does not converge, backpropagation is used to update the weights of the raster space modeling model until the loss converges, and finally a high-precision environmental space model is obtained. The grid space modeling model includes a GRU dynamic prediction submodule, which performs the following steps: The control GRU dynamic prediction submodule receives input features, processes them through the GRU hidden layer and fully connected layer, and outputs predicted motion parameters to obtain the predicted spatial state.

2. The method according to claim 1, characterized in that, The spatial state loss is calculated using the following formula: ； in, This represents the loss of spatial state. This represents the number of sampling points; These are the predicted spatial state values; These are the spatial state values ​​measured by the sensor. The obstacle position loss is calculated using the following formula: ； in, Predict loss based on obstacle location; The timing length; , These are the predicted obstacle coordinates; These are the measured coordinates of the obstacle; The total loss is calculated using the following formula. ； Among them, among them, , These are the spatial state loss weights and obstacle position loss weights, respectively.

3. The method according to claim 1, characterized in that, The multi-objective path optimization model includes a global path planning module, which performs the following steps: Initialize the start / end point and search list, then calculate and expand the minimum cost node using the node cost formula until the target point is reached, and determine the global reference path; The robot state is initialized by inputting the state space information into the deep reinforcement learning module. The state space information includes position, velocity, environmental state, and global reference path. The Actor network in the deep reinforcement learning module is controlled to make decisions and generate actions. After the actions are executed, the state is updated. Then, the multi-objective reward is calculated by combining the verification results of kinematic constraints and safe distance constraints. After the Q value is evaluated by the Critic network, the network parameters are updated through policy gradient. Determine whether the inspection robot has reached the target point or violated physical constraints; If the inspection robot reaches the target point without violating physical constraints, the final optimized path is determined.

4. The method according to claim 3, characterized in that, The node cost formula is as follows: ； in, For nodes The total cost; From the starting point to the node The actual cost; For nodes Estimating the cost of the Manhattan distance to the target point; For heuristic function weights; For nodes Distance to the nearest device / obstacle; For safety weights, The smaller the penalty, the greater the punishment.

5. The method according to claim 1, characterized in that, The multi-objective optimization and solution process involves the following steps: The path length, energy consumption, and smoothness targets are determined, and the physical constraint loss, path performance loss, and data fitting loss are determined. The physical constraint loss is used to ensure that the path conforms to the actual physical rules, the path performance loss is used to match the optimization targets of path length, energy consumption, and smoothness, and the data fitting loss is used to fit the data characteristics of the actual scenario. Calculate the total loss function and use the PSO algorithm to optimize and minimize the total loss function to obtain the optimal inspection route that satisfies the physical constraints; The PSO algorithm performs the following steps: Initialize the particle swarm and encode its velocity and angular velocity; Calculate the particle fitness, update the particle velocity and position, and determine whether the termination condition is met. The termination condition is that the number of iterations reaches the target number or the result converges. If the termination condition is not met, the particle fitness is recalculated and the particle's velocity and position are updated until the termination condition is met. If the termination condition is met, the optimal particle parameters are output, and the final optimal inspection path is generated based on the optimal particle parameters.

6. The method according to claim 5, characterized in that, The path length optimization objective is calculated using the following formula: in, The start time; The time to reach the target point. , That is, to achieve the shortest path; The energy consumption optimization target is calculated using the following formula: ； in, , These are the energy consumption coefficients, corresponding to linear motion and steering energy consumption, respectively. , Angular velocity, minimize That is, to achieve a low-energy consumption path; The path smoothness optimization objective is calculated using the following formula: ； Where, minimize That is, to achieve a smooth turning path.

7. The method according to claim 6, characterized in that, The physical constraint loss is calculated using the following formula: ； in, Loss due to physical constraints; , This refers to the actual distance between the robot and the equipment and obstacles. The path performance loss is calculated using the following formula: ； in, This is due to path performance loss; , , These are the weights for length, energy consumption, and smoothness loss, respectively. The data fitting loss is calculated using the following formula: ； in, For data fitting loss; The number of samples; Value for Critic network prediction; The target Q value; The total loss function is calculated using the following formula: ； in, , , These are the weighting coefficients for each loss term.

8. The method according to claim 7, characterized in that, Kinematic and safety constraints include: motion space constraints, kinematic constraints, and safety distance constraints; The action space constraints are as follows: ； in, For speed control, Angular velocity control quantity, Maximum speed, This is the maximum angular velocity; The kinematic constraint equations are as follows: ； ； in, , These are the minimum and maximum permissible speeds, respectively. , These are the minimum and maximum permissible angular velocities, respectively. This is the maximum acceleration; This is the maximum angular acceleration; The safety distance constraint equation is as follows: ； in, This refers to the minimum safe distance between the inspection robot and the equipment. This refers to the minimum safe distance between the inspection robot and the obstacle; For the inspection robot in Coordinates along the axis; For substation equipment at all times of Axis coordinate position; For the obstacle at time of Axis coordinate position; For the inspection robot in Coordinates along the axis; For substation equipment at all times of Axis coordinate position; For the obstacle at time of Axis coordinate position.

Images