A secure reinforcement learning method for power transmission line energy optimization embedded with physical information

Abstract

This invention discloses a safety reinforcement learning method for optimizing energy consumption in transmission lines by embedding physical information. The method includes: collecting line operation data, environmental data, and equipment parameter data; constructing a topology sensitivity matrix and a thermoelectric carbon state vector; using an improved TSLANT algorithm to establish candidate and baseline action image sequences, generating a thermoelectric carbon state vector and stitching the states together; inputting the stitched states into a reinforcement learning model, outputting a discrete target operation state sequence; constructing a topological homotopy mapping controller, searching for the farthest path segment and mapping it to an equipment control vector; dividing risk sub-regions and calculating regional recovery features to form a system-level safety certificate; executing equipment control actions and feeding back the results when conditions are met, updating the reinforcement learning strategy. This invention achieves interpretable optimization and safe and stable operation control of transmission line energy loss through safety reinforcement learning embedded with physical information and improved TSLANT modeling.

Description

Technical Field

[0001] This invention relates to the field of power system operation optimization and intelligent control technology, and in particular to a method for energy-saving optimization of transmission lines by embedding physical information through security reinforcement learning. Background Technology

[0002] With the continuous development of new power systems and the ever-expanding scale of transmission networks, the energy loss problem generated by transmission lines during long-distance, high-capacity operation is becoming increasingly prominent. To achieve the synergistic goal of energy conservation and safe, stable operation, existing technologies typically adjust the operating parameters of transmission lines through mechanistic model-based optimization scheduling methods or data-driven intelligent optimization methods. Mechanistic model methods rely on accurate modeling of the power grid topology, electrical parameters, and operating constraints, achieving energy-saving control through power flow calculations and constraint optimization. Data-driven methods, on the other hand, are mostly based on machine learning or reinforcement learning algorithms, training control strategies using historical operating data to achieve adaptive adjustment of the operating state.

[0003] However, existing technologies still have certain limitations in practical applications. On the one hand, optimization methods based on mechanistic models usually rely on static or quasi-static models, which are difficult to fully characterize the dynamic loss evolution process of transmission lines under complex operating conditions, especially under multi-source disturbances and time-varying load conditions, making it difficult to achieve refined attribution and dynamic control of energy loss. On the other hand, although pure data-driven reinforcement learning methods have strong adaptive capabilities, they generally lack effective embedding of the physical constraints of the power system, which can easily lead to control strategies that do not conform to the actual operating mechanism, posing certain safety hazards and making it difficult to directly apply to transmission systems with high safety requirements.

[0004] Existing methods for energy-saving optimization are mostly based on single-step or discrete control actions, lacking the characterization of the continuity and global consistency of control actions in the temporal evolution process. It is difficult to form a physically interpretable energy loss migration path. For local operational risks that may be caused during the optimization process, existing technologies usually use global constraints or simple threshold judgment methods for control, lacking refined modeling of risk sub-regions and hierarchical safety verification mechanisms. This makes it difficult to achieve the coordinated unity of energy-saving optimization and operational safety in complex operating environments.

[0005] Therefore, how to provide a secure reinforcement learning method for optimizing energy-saving transmission lines by embedding physical information is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a secure reinforcement learning-based energy-saving optimization method for transmission lines that embeds physical information. This invention introduces an improved TSLane network to jointly extract the thermal state evolution characteristics, energy loss correlation characteristics, and action impact characteristics during transmission line operation. It combines a reinforcement learning decision-making mechanism with power system physical constraints to construct an action-attribution thermal liability representation model, a continuous loss migration path generation mechanism, and a control decoding method based on topological homotopy mapping. This enables interpretable modeling and continuous optimization control of transmission line energy loss. Furthermore, by combining a hierarchical verification mechanism of risk sub-region local recovery certificates and system-level security certificates, it achieves coordinated and unified control of energy-saving optimization and operational safety. This method possesses advantages such as strong thermal loss representation capability, high physical consistency, strong interpretability of the optimization process, continuous and stable control strategy, and high system operational safety.

[0007] An energy-saving optimization method for power transmission lines embedding physical information according to an embodiment of the present invention includes: Collect operational data, environmental data, and equipment parameter data of transmission lines to construct a transmission topology sensitivity matrix and a line thermal-electrical-carbon state vector; By using an improved TSLane network, candidate motion image sequences and baseline motion image sequences are established. Synchronous simulation is performed within a preset time window to calculate the corresponding temperature trajectory difference, obtain the action-attribution thermal liability vector, and concatenate it with the line thermal-electrical-carbon state vector to form the reinforcement learning input state. The reinforcement learning input state is input into the reinforcement learning decision model, and the continuous loss migration path parameters from the current operating state to the target energy-saving operating state are output to generate several discrete target operating state sequences. A topology homotopy mapping controller is constructed. For each target operating state in the target operating state sequence, combined with the transmission topology sensitivity matrix, the feasible farthest path segment that satisfies the line thermal stability constraints and equipment operating boundaries is searched. The feasible farthest path segment is mapped to the equipment control vector through piecewise linear programming to obtain the corresponding equipment control action. The power transmission network is divided into several risk sub-regions. The regional recovery characteristic of each risk sub-region is calculated, the equipment control actions are verified, a local recovery certificate is generated, and the local recovery certificates are combined based on the boundary power exchange constraints between regions to obtain a system-level security certificate. When the system-level security certificate meets the preset execution conditions, the corresponding equipment control action is executed, and the execution result is fed back to the reinforcement learning decision model for policy update, so as to realize the energy-saving and optimized operation of the transmission line.

[0008] Optionally, the operating data includes node voltage, line current and power flow distribution; the environmental data includes wind speed, air temperature, humidity and irradiance; and the equipment parameter data includes conductor parameters and line structure parameters.

[0009] Optionally, the construction of the transmission topology sensitivity matrix and the line thermal-electrical-carbon state vector includes: The data collected includes node voltage, line current, active power flow, reactive power flow, conductor resistance, conductor reactance, conductor length, conductor cross-sectional area, ambient temperature, wind speed, air humidity, and solar irradiance of the transmission lines. Based on the node connection relationship of the transmission line, the active power flow of the line, the reactive power flow of the line, the conductor resistance and the conductor reactance, a transmission topology sensitivity matrix is ​​constructed. Line resistance loss is calculated based on line current, conductor resistance, conductor length, and conductor cross-sectional area; line corona loss is calculated based on line operating voltage, conductor structural parameters, and environmental meteorological parameters. The thermal state of the conductor is calculated based on the line current, conductor resistance, ambient temperature, wind speed, air humidity and solar irradiance, and the carbon state of the line is calculated based on the line resistance loss, line corona loss and preset power carbon emission factor. By splicing together node voltage, line current, line active power flow, line reactive power flow, conductor thermal state quantity, and line carbon state quantity, a line thermal-electrical-carbon state vector is constructed.

[0010] Optionally, obtaining the action-attribution thermal liability vector and concatenating it with the line thermal-electrical-carbon state vector to form the reinforcement learning input state includes: The system acquires information on line current, node voltage, active power flow, reactive power flow, conductor temperature, wind speed, air temperature, humidity, irradiance, and control actions of transmission lines over continuous historical periods. It then constructs a time-series input sample in chronological order, where the control action information includes candidate action information and baseline safety action information. An improved TSLnet network is constructed, which consists of a thermal-electric domain coding layer, an action mirroring embedding layer, a temporal interaction enhancement layer, and a thermal liability generation layer connected in sequence. Line current, node voltage, line active power flow, and line reactive power flow are input into the thermal-electric domain coding layer to form electrical state coding results. Conductor temperature, wind speed, air temperature, humidity, and irradiance are input into the thermal-electric domain coding layer to form thermal environment state coding results. The electrical state coding results are then spliced ​​with the thermal environment state coding results to obtain the thermal-electric joint coding results. The candidate action information and the baseline safety action information are respectively input into the action image embedding layer and combined with the thermo-electric joint coding results to obtain the candidate action image sequence and the baseline action image sequence. Then, the candidate action image sequence and the baseline action image sequence are input into the temporal interaction enhancement layer and synchronous inference is performed within the same preset time window to obtain the future temperature trajectory corresponding to the candidate action and the future temperature trajectory corresponding to the baseline safety action. The future temperature trajectory corresponding to the candidate action and the future temperature trajectory corresponding to the baseline safety action are input into the thermal liability generation layer. The temperature trajectory difference between the two is calculated at each time step and aggregated according to the line object to obtain the action-attributed thermal liability vector. The action-attribution thermal debt vector and the line thermal-electrical-carbon state vector are concatenated according to the feature dimension to form the reinforcement learning input state.

[0011] Optionally, generating a sequence of discrete target running states includes: A reinforcement learning decision model based on deep deterministic policy gradient is constructed. The reinforcement learning decision model includes a policy network and a value network. The policy network generates continuous action parameters based on the reinforcement learning input state, and the value network evaluates the value of the continuous action parameters in the current running state. The reinforcement learning input state is input into the policy network, and the continuous loss migration path parameters corresponding to the current operating state are output. The continuous loss migration path parameters represent the direction and magnitude of the migration from the current operating state to the target energy-saving operating state. The current operating state is progressively advanced based on the continuous loss migration path parameters. The migration process is discretized according to the preset step size to obtain several intermediate operating states, and each intermediate operating state is sequentially combined into a target operating state sequence. The target operating state sequence is input into the value network, the value of each target operating state is evaluated, and the target operating state that meets the energy-saving optimization target and thermal safety constraint is selected based on the evaluation results. Based on the value assessment results, the parameters of the policy network and the value network are updated to generate the target operating state sequence.

[0012] Optionally, obtaining the corresponding device control action includes: A topological homotopy mapping controller is constructed, which consists of a path projection unit, a boundary shrinkage unit, and an action mapping unit. The target operating state sequence is sequentially input into the path projection unit. Based on the node power change, line power flow change and equipment regulation response relationship represented by the transmission topology sensitivity matrix, the change between adjacent states in the target operating state sequence is projected and calculated to form a candidate migration path that extends continuously along the target operating state sequence. The candidate migration path is input into the boundary contraction unit. The candidate migration path is checked segment by segment according to the line thermal stability constraint, generator output adjustment range, transformer tap adjustment range, flexible transmission device adjustment range and reactive power compensation device switching range. Before the first constraint over-limit position appears, the continuous feasible interval is intercepted and the continuous feasible interval is determined as the feasible farthest path segment. The feasible farthest path segment is input into the action mapping unit. According to the state change of each discrete state in the feasible farthest path segment, the piecewise linear programming method is used to solve the equipment control quantity corresponding to each discrete state. The equipment control quantities corresponding to each discrete state are combined in time order to form an equipment control vector. The corresponding equipment control action is generated based on the equipment control vector.

[0013] Optionally, the generation of local recovery certificates, based on the combination of local recovery certificates according to the inter-region boundary power exchange constraints, yields a system-level security certificate, including: Based on the line connection relationship, power flow coupling relationship and regional boundary connection relationship of the transmission network, the transmission network is divided into several risk sub-regions, and the node set, line set, internal equipment set and boundary power exchange channel with adjacent risk sub-regions are determined for each risk sub-region. For each risk sub-region, extract the action-attribution thermal liability vector, the line thermal-electrical-carbon state vector, and the equipment control actions to construct regional recovery feature quantities. The regional recovery feature quantities include the line thermal liability state within the region, the node voltage stability state within the region, the line current-carrying recovery state within the region, and the regional boundary power exchange state. A region recovery orchestration unit is constructed, which consists of a disturbance implantation subunit, a path expansion subunit, and a certificate extraction subunit; The equipment control actions corresponding to each risk sub-region are input into the region recovery orchestration unit. The disturbance implantation sub-unit sequentially applies line disconnection disturbances, load fluctuation disturbances, and environmental change disturbances. The path unfolding sub-unit gradually unfolds the impact of the equipment control actions on the line temperature, node voltage, line power flow, and boundary power exchange within the region. The certificate extraction sub-unit determines whether the region recovery feature can enter the preset safe recovery range within the preset number of recovery steps, and generates a local recovery certificate for the corresponding risk sub-region. Extract the regional boundary power exchange status from each local recovery certificate. Based on the constraints that the input power and output power of adjacent risk sub-regions on the same boundary power exchange channel are matched, the boundary power flow direction is consistent, and the boundary exchange amount does not exceed the preset boundary exchange upper limit, perform boundary splicing on each local recovery certificate. The combination of local recovery certificates that satisfies all regional boundary power exchange constraints is retained to obtain the system-level security certificate.

[0014] Optionally, the execution of the corresponding device control action and the feedback of the execution result to the reinforcement learning decision model for policy update include: Obtain a system-level security certificate and determine whether the system-level security certificate meets preset execution conditions. The preset execution conditions include that the regional recovery feature quantities of each risk sub-region in the system-level security certificate are all within a preset security range and that the boundary power exchange between regions meets the consistency constraint. When the system-level security certificate meets the preset execution conditions, the generator active power regulation command, generator reactive power regulation command, flexible transmission device control command, transformer tap adjustment command and reactive power compensation device switching command are issued step by step according to the execution sequence of the equipment control action sequence to realize the execution of equipment control actions; During the execution of equipment control actions, line current, node voltage, line temperature, line active power flow, line reactive power flow, and boundary power exchange are collected in real time, and the line thermal-electrical-carbon state vector is updated based on the collection results. Based on the updated line thermal-electrical-carbon state vector and the corresponding equipment control actions, the action-attributed thermal liability vector is reconstructed, and the updated line thermal-electrical-carbon state vector and the action-attributed thermal liability vector are concatenated to form a new reinforcement learning input state. The new reinforcement learning input state is input into the reinforcement learning decision model, and the strategy of the reinforcement learning decision model is updated according to the actual operating effect after the device control action is executed.

[0015] The beneficial effects of this invention are: This invention introduces an improved TSLanet network into a reinforcement learning optimization framework to jointly model the thermal state evolution, energy loss distribution, and the impact of control actions during transmission line operation. This enables fine-grained characterization and interpretable attribution of energy loss. Compared to existing technologies that rely on static mechanism models or single data-driven methods, this invention can dynamically capture the evolution law of loss under complex operating conditions and multi-source disturbances. It transforms the originally difficult-to-quantify sources of loss into physically meaningful thermal liability representations attributable to actions, thus improving the accuracy and interpretability of the energy-saving optimization process.

[0016] This invention constructs a continuous loss migration path generation mechanism and combines it with a topological homotopy mapping controller to transform discrete control decisions into control paths with temporal continuity, achieving smooth regulation and globally consistent optimization of the transmission line's operating state. Compared to traditional single-step control or discrete strategy output methods, this invention effectively avoids operational fluctuations caused by control abrupt changes, improves the stability and adaptability of the control strategy in complex network structures, and enhances the consistency between the optimization process and the actual operating mechanism of the power system.

[0017] This invention introduces a hierarchical verification mechanism combining risk sub-region local recovery certificates and system-level security certificates, enabling refined identification and tiered control of potential operational risks during energy-saving optimization. Compared to traditional security control methods based on global constraints or threshold judgments, this invention can perform targeted risk repair within a local scope and complete security consistency verification at the global level. This effectively achieves energy-saving optimization goals while ensuring the safe and stable operation of the power transmission system, thereby improving the safety and reliability of system operation. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a security reinforcement learning method for optimizing power transmission line energy efficiency by embedding physical information, as proposed in this invention. Figure 2 This is a schematic diagram of the structure of the TSLanet network used to construct the action-attribution thermal liability, which is an improved method for energy-saving optimization of transmission lines by embedding physical information in the security reinforcement learning method proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 and Figure 2 A secure reinforcement learning method for optimizing power transmission line energy efficiency by embedding physical information includes: Collect operational data, environmental data, and equipment parameter data of transmission lines to construct a transmission topology sensitivity matrix and a line thermal-electrical-carbon state vector; By using an improved TSLane network, candidate motion image sequences and baseline motion image sequences are established. Synchronous simulation is performed within a preset time window to calculate the corresponding temperature trajectory difference, obtain the action-attribution thermal liability vector, and concatenate it with the line thermal-electrical-carbon state vector to form the reinforcement learning input state. The reinforcement learning input state is input into the reinforcement learning decision model, and the continuous loss migration path parameters from the current operating state to the target energy-saving operating state are output to generate several discrete target operating state sequences. A topology homotopy mapping controller is constructed. For each target operating state in the target operating state sequence, combined with the transmission topology sensitivity matrix, the feasible farthest path segment that satisfies the line thermal stability constraints and equipment operating boundaries is searched. The feasible farthest path segment is mapped to the equipment control vector through piecewise linear programming to obtain the corresponding equipment control action. The power transmission network is divided into several risk sub-regions. The regional recovery characteristic of each risk sub-region is calculated, the equipment control actions are verified, a local recovery certificate is generated, and the local recovery certificates are combined based on the boundary power exchange constraints between regions to obtain a system-level security certificate. When the system-level security certificate meets the preset execution conditions, the corresponding equipment control action is executed, and the execution result is fed back to the reinforcement learning decision model for policy update, so as to realize the energy-saving and optimized operation of the transmission line.

[0021] In this embodiment, the operating data includes node voltage, line current and power flow distribution; the environmental data includes wind speed, air temperature, humidity and irradiance; and the equipment parameter data includes conductor parameters and line structure parameters.

[0022] In this embodiment, the construction of the transmission topology sensitivity matrix and the line thermal-electrical-carbon state vector includes: The data collected includes node voltage, line current, active power flow, reactive power flow, conductor resistance, conductor reactance, conductor length, conductor cross-sectional area, ambient temperature, wind speed, air humidity, and solar irradiance of the transmission lines. Based on the node connection relationships, active power flow, reactive power flow, conductor resistance, and conductor reactance of the transmission lines, a transmission topology sensitivity matrix is ​​constructed. Specifically, the construction of the transmission topology sensitivity matrix is ​​as follows: Based on the node connection relationship of the transmission line, determine the connection direction and adjacency relationship between each node and each line, generate a node-line topology index table, and establish a basic topology matrix; Based on the active power flow, reactive power flow, conductor resistance and conductor reactance of each line, calculate the sensitivity values ​​of node-injected disturbances to active power transmission, node voltage disturbances to reactive power distribution, and changes in line parameters to changes in line heat loss. According to the node number and line number, each sensitivity value is filled into the corresponding position of the basic topology matrix, and the various sensitivity values ​​are normalized and combined to form a transmission topology sensitivity matrix that simultaneously represents the network topology, electrical transmission coupling relationship and heat loss conduction relationship. Line resistance loss is calculated based on line current, conductor resistance, conductor length, and conductor cross-sectional area. Corona loss is calculated based on line operating voltage, conductor structural parameters, and environmental meteorological parameters. Calculate line resistance loss: Based on the phase current value and conductor resistance parameters at the current moment, calculate the active power loss of the line conductor. Expand the conductor resistance equivalently according to the conductor length, and correct the resistance per unit length by combining the conductor cross-sectional area to obtain the equivalent resistance value of the line. According to the relationship of the three-phase transmission line, multiply the square of each phase current by the corresponding equivalent resistance and sum them to obtain the line resistance loss. When considering the influence of temperature on resistance, correct the resistance parameters according to the conductor temperature so that the resistance loss is adjusted with the conductor temperature change to obtain a line resistance loss value that matches the actual operating conditions. Calculate the corona loss of the line: Determine the electric field strength on the conductor surface based on the line operating voltage and conductor structural parameters, and calculate the critical corona initiation voltage by combining the conductor radius, the spacing between split conductors, and the air density correction factor; When the line operating voltage is higher than the critical corona initiation voltage, calculate the corona loss power per unit length based on the difference between the operating voltage and the critical voltage. The corona loss power increases with the increase of the difference between the operating voltage and the critical voltage, and the air density factor is corrected by combining the air temperature, air pressure, and humidity; The corrected corona loss per unit length is accumulated along the line length to obtain the corona loss value of the entire line; The thermal state of the conductors is calculated based on line current, conductor resistance, ambient temperature, wind speed, air humidity, and solar irradiance. The carbon state of the line is calculated based on line resistance loss, line corona loss, and a preset power carbon emission factor. Specifically, the calculation of the conductor thermal state involves: Based on the thermal balance relationship of the conductor, the conductor temperature is calculated. The Joule heating power corresponding to the line current and conductor resistance is used as the heat input, and the radiative heat absorption corresponding to the solar irradiance intensity is superimposed. At the same time, the convective heat dissipation and radiative heat dissipation processes between the conductor and the surrounding environment are considered. The convective heat dissipation is calculated based on the wind speed and the temperature difference between the conductor surface temperature and the ambient temperature. The radiative heat dissipation is calculated based on the difference between the conductor surface temperature and the ambient temperature and the radiation coefficient after correction for air humidity. The heat input and heat output per unit length of conductor are balanced to obtain the corresponding temperature value or temperature rise change of the conductor. The results are updated step by step in a continuous time step to form the conductor thermal state quantity that reflects the process of heat accumulation and heat dissipation of the conductor. The calculation of the line carbon state parameters is as follows: The total energy loss of the line is calculated based on the line resistance loss and the line corona loss. The two types of losses are integrated over time to obtain the total energy loss per unit time. Combined with the preset power carbon emission factor, the energy loss is converted into the corresponding carbon emission. The power carbon emission factor is used to characterize the carbon emission level corresponding to a unit of electricity. The carbon emission is normalized or weighted and accumulated according to the load level of the transmission line in different time periods to obtain the line carbon state quantity that reflects the carbon emission intensity and change trend during the line operation. The preset carbon emission factor for electricity is based on the power generation structure of the power grid in the area to which the transmission line belongs. The carbon emission factor for electricity is the carbon emission per unit of electricity, which is set to 0.65 kg CO2 / kWh. By splicing together node voltage, line current, line active power flow, line reactive power flow, conductor thermal state quantity, and line carbon state quantity, a line thermal-electrical-carbon state vector is constructed.

[0023] In this embodiment, obtaining the action-attribution thermal liability vector and concatenating it with the line thermal-electrical-carbon state vector to form the reinforcement learning input state includes: The system acquires information on line current, node voltage, active power flow, reactive power flow, conductor temperature, wind speed, air temperature, humidity, irradiance, and control actions of transmission lines over continuous historical periods. It then constructs a time-series input sample in chronological order, where the control action information includes candidate action information and baseline safety action information. An improved TSLnet network is constructed, which consists of a thermal-electric domain coding layer, an action mirroring embedding layer, a temporal interaction enhancement layer, and a thermal liability generation layer connected in sequence. Line current, node voltage, line active power flow, and line reactive power flow are input into the thermal-electric domain coding layer to form the electrical state coding result. Conductor temperature, wind speed, air temperature, humidity, and irradiance are input into the thermal-electric domain coding layer to form the thermal environment state coding result. The electrical state coding result and the thermal environment state coding result are then concatenated to obtain the thermal-electric joint coding result, in which: The electrical state coding result is formed as follows: the line current, node voltage, line active power flow and line reactive power flow are constructed in time order to form an electrical time series feature sequence. Each electrical feature is normalized to eliminate the difference in dimensions. The electrical time series feature sequence is slid segmented based on a preset time window and input into the electrical feature coding module in the thermal-electric domain coding layer. The current amplitude change, voltage fluctuation characteristics and power flow distribution characteristics are jointly extracted through multi-scale temporal convolution and feature mapping transformation to obtain the electrical state coding result that characterizes the line electrical operating state and its temporal change relationship. The thermal environment state coding results are generated as follows: the thermal environment time series feature sequence is constructed by arranging conductor temperature, wind speed, air temperature, humidity and irradiance in chronological order. Each thermal environment feature is normalized and combined with the coupling relationship between conductor temperature and environmental meteorological parameters. The results are then input into the thermal environment feature coding module in the thermal-electric domain coding layer within a preset time window. Through multi-scale temporal convolution and nonlinear mapping, features of temperature rise trend, environmental heat dissipation conditions and radiation heat absorption characteristics are extracted to obtain the thermal environment state coding results that characterize the evolution process of conductor thermal state and environmental influencing factors. Candidate action information and baseline safety action information are input into the action image embedding layer, and combined with the thermo-electric joint coding results to obtain candidate action image sequences and baseline action image sequences. These sequences are then input into the temporal interaction enhancement layer, where synchronous simulation is performed within the same preset time window to obtain the future temperature trajectories corresponding to the candidate actions and the baseline safety actions, respectively. Specifically, the candidate action image sequences and baseline action image sequences are obtained as follows: The candidate action information and the baseline safety action information are arranged in chronological order to construct an action time sequence feature sequence, and each action feature is normalized. The action time sequence feature sequence is transformed into the same feature space as the thermo-electric joint coding result through embedding mapping. At each time step, the action features are concatenated or weighted and fused with the thermo-electric joint coding result at the corresponding time step to form a joint feature representation containing action impact information. The above combination process is performed on the candidate action information and the baseline safety action information respectively to obtain the corresponding candidate action image sequence and the baseline action image sequence. The future temperature trajectory corresponding to the candidate action and the future temperature trajectory corresponding to the baseline safety action are obtained as follows: The candidate motion image sequence and the baseline motion image sequence are respectively input into the temporal interaction enhancement layer. Within a preset time window, the two types of image sequences are synchronously extrapolated through a temporal feature extraction structure with shared parameters. Based on the thermal state and motion impact characteristics at the current moment, the conductor temperature at each future time step is recursively predicted. The temperature prediction result at each time step depends on the predicted temperature at the previous time step and the motion image characteristics at the current time step. The temperature change is smoothed by combining thermal inertia constraints. By iteratively calculating each time step, the future temperature trajectory under the candidate action and the future temperature trajectory under the baseline safety action are obtained respectively. The future temperature trajectory corresponding to the candidate action and the future temperature trajectory corresponding to the baseline safety action are input into the thermal liability generation layer. The temperature trajectory difference between the two is calculated time-by-time, and then aggregated according to the line object to obtain the action-attributed thermal liability vector. Specifically, the calculation of the temperature trajectory difference between the two is as follows: The future temperature trajectory corresponding to the candidate action is aligned with the future temperature trajectory corresponding to the baseline safety action according to the unified time index. The candidate temperature value and the baseline temperature value of the same line object at each prediction time are extracted, and the temperature trajectory difference at the corresponding time is calculated by subtracting the baseline temperature value from the candidate temperature value. When the temperature trajectory difference is greater than zero, it indicates that the candidate action introduces additional thermal burden relative to the baseline safe action. When the temperature trajectory difference is less than or equal to zero, it indicates that the candidate action does not increase additional thermal burden or produce thermal mitigation effect. The temperature trajectory differences obtained at each prediction time are retained in chronological order to form a time-series difference sequence for the corresponding line object; The action-attribution thermal debt vector and the line thermal-electrical-carbon state vector are concatenated according to the feature dimension to form the reinforcement learning input state.

[0024] In this embodiment, generating a sequence of discrete target running states includes: A reinforcement learning decision model based on deep deterministic policy gradient is constructed. The reinforcement learning decision model includes a policy network and a value network. The policy network generates continuous action parameters based on the reinforcement learning input state, and the value network evaluates the value of the continuous action parameters in the current running state. The reinforcement learning input state is input into the policy network, and the continuous loss migration path parameters corresponding to the current operating state are output. The continuous loss migration path parameters represent the direction and magnitude of the migration from the current operating state to the target energy-saving operating state. Specifically, the continuous loss migration path parameters corresponding to the current operating state are output as follows: The reinforcement learning input state is input into the policy network, and intermediate feature representations are obtained through multi-layer linear transformation and nonlinear activation. The intermediate feature representations are mapped in the policy output layer to generate continuous action vectors. The continuous action vectors are divided into parameter components corresponding to each controlled object according to the control variable dimension. Each parameter component is represented as a migration direction component and a migration amplitude component. After numerical weighting of each parameter component based on the action attribution heat debt vector, the weighting result is normalized and truncated within a preset numerical range to obtain continuous loss migration path parameters. The current operating state is progressively advanced based on the continuous loss migration path parameters. The migration process is discretized according to a preset step size to obtain several intermediate operating states. These intermediate operating states are then sequentially combined to form the target operating state sequence. Specifically, the discretization of the migration process according to the preset step size is as follows: Based on the continuous loss migration path parameters, the variation range of each control variable during the migration process is determined, and the variation range is divided into equal intervals according to the preset step size to obtain multiple discrete segment points. Starting from the current operating state, interpolation calculations are performed on each control variable step by step according to the segmented points, and corresponding control variable value combinations are generated at each segment position. The control variable values ​​corresponding to each segment position are mapped to the corresponding operating states, and arranged in chronological order to form several intermediate operating states. The target operating state sequence is input into the value network, and the value of each target operating state is evaluated. Based on the evaluation results, target operating states that meet the energy-saving optimization objectives and thermal safety constraints are selected. Specifically, the value evaluation of each target operating state is as follows: Each target operating state in the target operating state sequence is input into the value network, and the corresponding state feature representation is obtained through multi-layer linear transformation and nonlinear mapping. Based on the state feature representation, the thermal state quantities corresponding to line resistance loss, line corona loss and conductor temperature are numerically calculated, and the corresponding state evaluation quantity is formed by combining the line carbon state quantity. The state evaluation quantity is input into the value output layer for mapping to obtain the state value function value corresponding to each target operating state. The energy-saving optimization objectives include: aiming to reduce the total energy loss during the operation of transmission lines, jointly optimizing line resistance loss and line corona loss, while also taking into account the reduction of line carbon state, so as to minimize the comprehensive energy consumption or equivalent carbon emissions per unit time. The energy-saving optimization objectives also include constraining the adjustment range of each control variable to ensure that the control process remains smooth within the continuous migration path and to avoid the increase of additional losses caused by control abrupt changes. Thermal safety constraints include: limiting the temperature of the conductor under any operating state to not exceed a preset thermal stability upper limit threshold, which is set according to the conductor material properties; constraining the rate of change of conductor temperature in adjacent time steps to not exceed a preset range to ensure the continuity of the temperature evolution process; and constraining the coupling relationship between line current and conductor temperature to keep the thermal equilibrium state within a safe range under given ambient temperature, wind speed and irradiance conditions, ensuring that the transmission line meets the thermal stability operation requirements throughout the entire operation process. Based on the value assessment results, the parameters of the policy network and the value network are updated to generate the target operating state sequence.

[0025] In this embodiment, obtaining the corresponding device control action includes: A topological homotopy mapping controller is constructed, which consists of a path projection unit, a boundary shrinkage unit, and an action mapping unit. The target operating state sequence is sequentially input into the path projection unit. Based on the node power change, line power flow change, and equipment regulation response relationship represented by the transmission topology sensitivity matrix, the changes between adjacent states in the target operating state sequence are projected and calculated to form candidate migration paths that extend continuously along the target operating state sequence. Specifically, the projection calculation of the changes between adjacent states in the target operating state sequence is as follows: Differential calculation is performed on any two adjacent operating states in the target operating state sequence to obtain the state change vector of the corresponding control variable. The state change vector is then multiplied with the transmission topology sensitivity matrix to obtain the projection change in node power, line power flow, and equipment regulation space. Based on the numerical distribution of each projection component in the corresponding physical quantity dimension, the direction of the state change vector is adjusted and the amplitude is redistributed. The changes calculated by projection are then connected in the order of the target operating state sequence to form a candidate migration path that extends continuously along the target operating state sequence. The candidate migration path is input into the boundary contraction unit. The candidate migration path is checked segment by segment according to the line thermal stability constraint, generator output adjustment range, transformer tap adjustment range, flexible transmission device adjustment range and reactive power compensation device switching range. Before the first constraint over-limit position appears, the continuous feasible interval is intercepted and the continuous feasible interval is determined as the feasible farthest path segment. The feasible farthest path segment is input into the action mapping unit. Based on the state changes of each discrete state within the feasible farthest path segment, piecewise linear programming is used to solve for the equipment control quantities corresponding to each discrete state. These equipment control quantities are then combined in chronological order to form an equipment control vector. This equipment control vector includes generator active power regulation, generator reactive power regulation, flexible transmission device setpoints, transformer tap positions, and reactive power compensation device switching states. Corresponding equipment control actions are generated based on the equipment control vector. Specifically, the piecewise linear programming method used to solve for the equipment control quantities corresponding to each discrete state is as follows: For each discrete state in the feasible longest path segment, the state change of the discrete state relative to the previous state is extracted, and the state change is used as a constraint to establish a system of linear equations. In the system of linear equations, the active power regulation of the generator, the reactive power regulation of the generator, the setpoint of the flexible transmission device, the tap position of the transformer, and the switching state of the reactive power compensation device are used as variables to be solved, and the node power balance relationship, the line power flow distribution relationship, and the equipment adjustment range constitute the linear constraints. According to the transmission topology sensitivity matrix, the correspondence between the state change and each variable to be solved is expressed in matrix form, and the changes of each variable to be solved are linearly combined to form the solution objective. By solving the system of linear equations segment by segment, the equipment control quantities corresponding to each discrete state are obtained, and they are combined according to the path order to form the equipment control vector corresponding to each segment.

[0026] In this embodiment, the generation of local recovery certificates, based on the combination of local recovery certificates according to the inter-region boundary power exchange constraints, yields a system-level security certificate, including: Based on the line connection relationship, power flow coupling relationship and regional boundary connection relationship of the transmission network, the transmission network is divided into several risk sub-regions, and the node set, line set, internal equipment set and boundary power exchange channel with adjacent risk sub-regions are determined for each risk sub-region. For each risk sub-region, extract the action-attribution thermal liability vector, the line thermal-electrical-carbon state vector, and the equipment control actions to construct regional recovery feature quantities. The regional recovery feature quantities include the line thermal liability state within the region, the node voltage stability state within the region, the line current-carrying recovery state within the region, and the regional boundary power exchange state. A region recovery orchestration unit is constructed, which consists of a disturbance implantation subunit, a path expansion subunit, and a certificate extraction subunit; The equipment control actions corresponding to each risk sub-region are input into the region recovery orchestration unit. The disturbance implantation sub-unit sequentially applies line disconnection disturbances, load fluctuation disturbances, and environmental change disturbances. The path unfolding sub-unit gradually unfolds the impact of the equipment control actions on line temperature, node voltage, line power flow, and boundary power exchange within the region. The certificate extraction sub-unit determines whether the region recovery feature can enter the preset safe recovery range within a preset number of recovery steps, and generates a local recovery certificate for the corresponding risk sub-region. Specifically, the determination by the certificate extraction sub-unit based on whether the region recovery feature can enter the preset safe recovery range within a preset number of recovery steps is as follows: The regional recovery feature quantities formed by each risk sub-region under the disturbance are extracted step by step. The feature values ​​corresponding to each recovery step are obtained in chronological order and compared with the preset safe recovery interval. When the regional recovery feature quantity enters the preset safe recovery interval at any time within the preset number of recovery steps, the corresponding risk sub-region is marked as meeting the recovery conditions; when it does not enter the preset safe recovery interval within the preset number of recovery steps, the corresponding risk sub-region is marked as not meeting the recovery conditions. Generate a local recovery certificate for the corresponding risk sub-region, specifically as follows: Based on the change process of the regional recovery characteristic quantity within the preset recovery steps and the comparison results between the regional recovery characteristic quantity and the preset safe recovery interval, the recovery judgment result of the risk sub-region is determined, and the corresponding recovery steps, regional recovery characteristic quantity sequence and boundary power exchange quantity are extracted as certificate basis data. The recovery judgment result is correlated with the constraint satisfaction of line temperature, node voltage, line power flow and boundary power exchange quantity in the risk sub-region to form a structured feature description reflecting the recovery capability and operational safety of the risk sub-region. The structured feature description is encapsulated according to the preset coding rules to generate a local recovery certificate containing recovery judgment identifier, key feature parameters and constraint satisfaction information. Extract the regional boundary power exchange status from each local recovery certificate. Based on the constraints that the input power and output power of adjacent risk sub-regions on the same boundary power exchange channel are matched, the boundary power flow direction is consistent, and the boundary exchange amount does not exceed the preset boundary exchange upper limit, perform boundary splicing on each local recovery certificate. The combination of local recovery certificates that satisfies all regional boundary power exchange constraints is retained to obtain the system-level security certificate.

[0027] In this embodiment, the execution of the corresponding device control action and the feedback of the execution result to the reinforcement learning decision model for policy update include: Obtain a system-level security certificate and determine whether the system-level security certificate meets preset execution conditions. The preset execution conditions include that the regional recovery features of each risk sub-region in the system-level security certificate are all within a preset security range and that the boundary power exchange between regions meets consistency constraints. The consistency constraints refer to: Constraints are imposed on the exchange of active and reactive power on the boundary connecting lines of adjacent risk sub-regions to ensure that the power exchange amount of the same boundary line in two adjacent regions remains consistent in value and opposite in direction. At the same time, the power injection and outflow of each boundary node are constrained to meet the power balance relationship, so that the power transfer between regions is continuous and without abrupt changes. Allowable deviation ranges are set for the power change of the boundary lines to ensure that the power flow distribution of each risk sub-region in the overall network after combination is consistent with the node power relationship. When the system-level security certificate meets the preset execution conditions, the generator active power regulation command, generator reactive power regulation command, flexible transmission device control command, transformer tap adjustment command and reactive power compensation device switching command are issued step by step according to the execution sequence of the equipment control action sequence to realize the execution of equipment control actions; During the execution of equipment control actions, line current, node voltage, line temperature, line active power flow, line reactive power flow, and boundary power exchange are collected in real time, and the line thermal-electrical-carbon state vector is updated based on the collection results. Based on the updated line thermal-electrical-carbon state vector and the corresponding equipment control actions, the action-attributed thermal liability vector is reconstructed, and the updated line thermal-electrical-carbon state vector and the action-attributed thermal liability vector are concatenated to form a new reinforcement learning input state. The new reinforcement learning input state is input into the reinforcement learning decision model, and the strategy of the reinforcement learning decision model is updated according to the actual operating effect after the device control action is executed.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a 220kV transmission line, approximately 72km long, connecting two regional main substations and multiple industrial load access points, undertaking the daily load transmission task of approximately 180MW to 230MW in the region. The line operates in a complex environment, with high temperatures and frequent load fluctuations in summer, resulting in significant conductor temperature rise, increased line resistance, and consequently, increased active power loss. Existing operating methods mainly rely on static optimization strategies based on power flow calculations, with scheduling cycles typically ranging from 15 to 30 minutes, making it difficult to respond promptly to load changes. Furthermore, they lack a detailed characterization of loss sources, failing to clearly define the specific impact of different control actions on line losses. During high-load periods, some sections experience temperatures approaching safe thresholds. Traditional methods primarily rely on overall limit control for constraint, lacking detailed identification and targeted repair of local risk areas, resulting in a difficulty in balancing energy-saving optimization and operational safety.

[0029] In this scenario, the proposed security reinforcement learning-based power transmission line energy-saving optimization method embedding physical information is deployed in a dispatch control system. First, continuous data collection of transmission line operation data generates multi-dimensional operational status data including node voltage, current, line power, ambient temperature, and historical control actions, constructing a corresponding thermal-electrical-carbon state vector. Then, this state vector is input into an improved TSLaneet network. Leveraging its multi-scale temporal feature extraction capabilities, the network models the trend of line thermal state changes and energy loss distribution, generating action-attributed thermal liability results and quantifying the contribution of various control actions to losses.

[0030] The thermal load characterization results and the current operating status are input into the safety reinforcement learning strategy network to generate continuous loss migration path parameters, ensuring the continuity and smoothness of the control strategy in the time dimension. A topological homotopy mapping controller maps these continuous path parameters to specific reactive power compensation adjustments and transformer tap changer commands. Before the control strategy is executed, the line is divided into risk sub-regions, and local recovery certificates are generated for sections prone to temperature anomalies. These certificates are further combined to form a system-level safety certificate, performing overall safety verification of the control strategy to ensure that all adjustments meet grid operation constraints. Verified control commands are then issued to each execution unit, and the model is updated through real-time feedback to achieve closed-loop optimization control.

[0031] In a comparative test that ran continuously for a week, the method of the present invention was compared and analyzed with the original scheduling method.

[0032] Table 1 Comparison of Energy-Saving Optimization Effects of Transmission Lines

[0033] As shown in Table 1, during the seven-day comparative operation, under similar load levels, the method of this invention demonstrated stable and continuous optimization effects in energy loss control. With the average load fluctuating between 180MW and 210MW, the line loss under the traditional method remained between 4.7MWh and 5.2MWh, while the corresponding losses were reduced to between 4.3MWh and 4.7MWh after adopting the method of this invention, representing an overall reduction of approximately 8% to 10%. Furthermore, no significant fluctuations or instability were observed under different load levels, indicating that this invention, by improving the action-attribution thermal liability and continuous loss migration path mechanism in the TSLaneet network construction, can achieve stable and effective energy-saving optimization under multiple operating conditions.

[0034] Regarding conductor temperature control, a significant downward trend in daily peak temperatures was observed. Under traditional operating conditions, peak temperatures were concentrated between 86℃ and 92℃, while the method of this invention reduced them to the range of 82℃ to 88℃, an average decrease of approximately 4℃. Particularly during the high-load days of the 3rd and 5th, temperatures decreased from 91℃ and 92℃ to 87℃ and 88℃ respectively, effectively avoiding operation near the thermal stability limit. This demonstrates that by introducing risk sub-region local recovery certificates and system-level safety certificate mechanisms, this invention can proactively suppress potential overheating risks while optimizing losses, achieving synergistic control of energy saving and safety.

[0035] Regarding control response capability, the table shows that the scheduling response time of traditional methods is approximately 20 seconds, while the method of this invention is stable at around 6 seconds, representing a response speed improvement of over 60%. This improvement in response capability mainly stems from the combination of reinforcement learning policy networks and continuous control path generation mechanisms, enabling the control strategy to be updated in real time according to the operating state, without relying on fixed-period discrete optimization calculations. This not only improves the system's adaptability to load fluctuations but also reduces the accumulation of energy losses and local operational risks caused by control lag, further validating the advantages of this invention in terms of real-time performance and engineering application feasibility.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for energy-saving optimization of transmission lines using security reinforcement learning embedded with physical information, characterized in that, include: Collect operational data, environmental data, and equipment parameter data of transmission lines to construct a transmission topology sensitivity matrix and a line thermal-electrical-carbon state vector; By using an improved TSLane network, candidate motion image sequences and baseline motion image sequences are established. Synchronous simulation is performed within a preset time window to calculate the corresponding temperature trajectory difference, obtain the action-attribution thermal liability vector, and concatenate it with the line thermal-electrical-carbon state vector to form the reinforcement learning input state. The reinforcement learning input state is input into the reinforcement learning decision model, and the continuous loss migration path parameters from the current operating state to the target energy-saving operating state are output to generate several discrete target operating state sequences. A topology homotopy mapping controller is constructed. For each target operating state in the target operating state sequence, combined with the transmission topology sensitivity matrix, the feasible farthest path segment that satisfies the line thermal stability constraints and equipment operating boundaries is searched. The feasible farthest path segment is mapped to the equipment control vector through piecewise linear programming to obtain the corresponding equipment control action. The power transmission network is divided into several risk sub-regions. The regional recovery characteristic of each risk sub-region is calculated, the equipment control actions are verified, a local recovery certificate is generated, and the local recovery certificates are combined based on the boundary power exchange constraints between regions to obtain a system-level security certificate. When the system-level security certificate meets the preset execution conditions, the corresponding equipment control action is executed, and the execution result is fed back to the reinforcement learning decision model for policy update, so as to realize the energy-saving and optimized operation of the transmission line.

2. The energy-saving optimization method for power transmission lines with embedded physical information according to claim 1, characterized in that, The operational data includes node voltage, line current, and power flow distribution; the environmental data includes wind speed, air temperature, humidity, and irradiance; and the equipment parameter data includes conductor parameters and line structure parameters.

3. The energy-saving optimization method for power transmission lines with embedded physical information according to claim 1, characterized in that, The construction of the transmission topology sensitivity matrix and the line thermal-electrical-carbon state vector includes: The data collected includes node voltage, line current, active power flow, reactive power flow, conductor resistance, conductor reactance, conductor length, conductor cross-sectional area, ambient temperature, wind speed, air humidity, and solar irradiance of the transmission lines. Based on the node connection relationship of the transmission line, the active power flow of the line, the reactive power flow of the line, the conductor resistance and the conductor reactance, a transmission topology sensitivity matrix is ​​constructed. Line resistance loss is calculated based on line current, conductor resistance, conductor length, and conductor cross-sectional area; line corona loss is calculated based on line operating voltage, conductor structural parameters, and environmental meteorological parameters. The thermal state of the conductor is calculated based on the line current, conductor resistance, ambient temperature, wind speed, air humidity and solar irradiance, and the carbon state of the line is calculated based on the line resistance loss, line corona loss and preset power carbon emission factor. By splicing together node voltage, line current, line active power flow, line reactive power flow, conductor thermal state quantity, and line carbon state quantity, a line thermal-electrical-carbon state vector is constructed.

4. The energy-saving optimization method for transmission lines with embedded physical information according to claim 1, characterized in that, The process of obtaining the action-attribution thermal liability vector and concatenating it with the line thermal-electrical-carbon state vector to form the reinforcement learning input state includes: The system acquires information on line current, node voltage, active power flow, reactive power flow, conductor temperature, wind speed, air temperature, humidity, irradiance, and control actions of transmission lines over continuous historical periods. It then constructs a time-series input sample in chronological order, where the control action information includes candidate action information and baseline safety action information. An improved TSLnet network is constructed, which consists of a thermal-electric domain coding layer, an action mirroring embedding layer, a temporal interaction enhancement layer, and a thermal liability generation layer connected in sequence. Line current, node voltage, line active power flow, and line reactive power flow are input into the thermal-electric domain coding layer to form electrical state coding results. Conductor temperature, wind speed, air temperature, humidity, and irradiance are input into the thermal-electric domain coding layer to form thermal environment state coding results. The electrical state coding results are then spliced ​​with the thermal environment state coding results to obtain the thermal-electric joint coding results. The candidate action information and the baseline safety action information are respectively input into the action image embedding layer and combined with the thermo-electric joint coding results to obtain the candidate action image sequence and the baseline action image sequence. Then, the candidate action image sequence and the baseline action image sequence are input into the temporal interaction enhancement layer and synchronous inference is performed within the same preset time window to obtain the future temperature trajectory corresponding to the candidate action and the future temperature trajectory corresponding to the baseline safety action. The future temperature trajectory corresponding to the candidate action and the future temperature trajectory corresponding to the baseline safety action are input into the thermal liability generation layer. The temperature trajectory difference between the two is calculated at each time step and aggregated according to the line object to obtain the action-attributed thermal liability vector. The action-attribution thermal debt vector and the line thermal-electrical-carbon state vector are concatenated according to the feature dimension to form the reinforcement learning input state.

5. The method for energy-saving optimization of transmission lines using security reinforcement learning embedded with physical information according to claim 1, characterized in that, The generation of several discrete target running state sequences includes: A reinforcement learning decision model based on deep deterministic policy gradient is constructed. The reinforcement learning decision model includes a policy network and a value network. The policy network generates continuous action parameters based on the reinforcement learning input state, and the value network evaluates the value of the continuous action parameters in the current running state. The reinforcement learning input state is input into the policy network, and the continuous loss migration path parameters corresponding to the current operating state are output. The continuous loss migration path parameters represent the direction and magnitude of the migration from the current operating state to the target energy-saving operating state. The current operating state is progressively advanced based on the continuous loss migration path parameters. The migration process is discretized according to the preset step size to obtain several intermediate operating states, and each intermediate operating state is sequentially combined into a target operating state sequence. The target operating state sequence is input into the value network, the value of each target operating state is evaluated, and the target operating state that meets the energy-saving optimization target and thermal safety constraint is selected based on the evaluation results. Based on the value assessment results, the parameters of the policy network and the value network are updated to generate the target operating state sequence.

6. The energy-saving optimization method for transmission lines with embedded physical information according to claim 1, characterized in that, The obtained corresponding device control actions include: A topological homotopy mapping controller is constructed, which consists of a path projection unit, a boundary shrinkage unit, and an action mapping unit. The target operating state sequence is sequentially input into the path projection unit. Based on the node power change, line power flow change and equipment regulation response relationship represented by the transmission topology sensitivity matrix, the change between adjacent states in the target operating state sequence is projected and calculated to form a candidate migration path that extends continuously along the target operating state sequence. The candidate migration path is input into the boundary contraction unit. The candidate migration path is checked segment by segment according to the line thermal stability constraint, generator output adjustment range, transformer tap adjustment range, flexible transmission device adjustment range and reactive power compensation device switching range. Before the first constraint over-limit position appears, the continuous feasible interval is intercepted and the continuous feasible interval is determined as the feasible farthest path segment. The feasible farthest path segment is input into the action mapping unit. According to the state change of each discrete state in the feasible farthest path segment, the piecewise linear programming method is used to solve the equipment control quantity corresponding to each discrete state. The equipment control quantities corresponding to each discrete state are combined in time order to form an equipment control vector. The corresponding equipment control action is generated based on the equipment control vector.

7. The method for energy-saving optimization of transmission lines using security reinforcement learning embedded with physical information according to claim 1, characterized in that, The generation of local recovery certificates, based on the combination of inter-region boundary power exchange constraints, yields a system-level security certificate, including: Based on the line connection relationship, power flow coupling relationship and regional boundary connection relationship of the transmission network, the transmission network is divided into several risk sub-regions, and the node set, line set, internal equipment set and boundary power exchange channel with adjacent risk sub-regions are determined for each risk sub-region. For each risk sub-region, extract the action-attribution thermal liability vector, the line thermal-electrical-carbon state vector, and the equipment control actions to construct regional recovery feature quantities. The regional recovery feature quantities include the line thermal liability state within the region, the node voltage stability state within the region, the line current-carrying recovery state within the region, and the regional boundary power exchange state. A region recovery orchestration unit is constructed, which consists of a disturbance implantation subunit, a path expansion subunit, and a certificate extraction subunit; The equipment control actions corresponding to each risk sub-region are input into the region recovery orchestration unit. The disturbance implantation sub-unit sequentially applies line disconnection disturbances, load fluctuation disturbances, and environmental change disturbances. The path unfolding sub-unit gradually unfolds the impact of the equipment control actions on the line temperature, node voltage, line power flow, and boundary power exchange within the region. The certificate extraction sub-unit determines whether the region recovery feature can enter the preset safe recovery range within the preset number of recovery steps, and generates a local recovery certificate for the corresponding risk sub-region. Extract the regional boundary power exchange status from each local recovery certificate. Based on the constraints that the input power and output power of adjacent risk sub-regions on the same boundary power exchange channel are matched, the boundary power flow direction is consistent, and the boundary exchange amount does not exceed the preset boundary exchange upper limit, perform boundary splicing on each local recovery certificate. The combination of local recovery certificates that satisfies all regional boundary power exchange constraints is retained to obtain the system-level security certificate.

8. The method for energy-saving optimization of transmission lines using security reinforcement learning embedded with physical information according to claim 1, characterized in that, The execution of the corresponding device control action, and the feedback of the execution result to the reinforcement learning decision model for policy update, includes: Obtain a system-level security certificate and determine whether the system-level security certificate meets preset execution conditions. The preset execution conditions include that the regional recovery feature quantities of each risk sub-region in the system-level security certificate are all within a preset security range and that the boundary power exchange between regions meets the consistency constraint. When the system-level security certificate meets the preset execution conditions, the generator active power regulation command, generator reactive power regulation command, flexible transmission device control command, transformer tap adjustment command and reactive power compensation device switching command are issued step by step according to the execution sequence of the equipment control action sequence to realize the execution of equipment control actions; During the execution of equipment control actions, line current, node voltage, line temperature, line active power flow, line reactive power flow, and boundary power exchange are collected in real time, and the line thermal-electrical-carbon state vector is updated based on the collection results. Based on the updated line thermal-electrical-carbon state vector and the corresponding equipment control actions, the action-attributed thermal liability vector is reconstructed, and the updated line thermal-electrical-carbon state vector and the action-attributed thermal liability vector are concatenated to form a new reinforcement learning input state. The new reinforcement learning input state is input into the reinforcement learning decision model, and the strategy of the reinforcement learning decision model is updated according to the actual operating effect after the device control action is executed.

Images