A flexible hvdc traction power supply double-layer planning optimization method combining safety checking and operation simulation

Abstract

This application relates to a two-layer planning optimization method for flexible DC traction power supply that integrates safety verification and operational simulation. The method includes: spatiotemporally and uniformly adjusting multi-source planning data and operational baseline data of the flexible DC traction power supply to obtain two-layer planning variable data and power supply operation scenario data; performing differentiable coupled simulation on the power supply based on the two-layer planning variable data and power supply operation scenario data to obtain operational trajectory data and sensitivity information; performing formal safety verification on the operational trajectory data to obtain constraint enhancement information and risk index data; conducting adversarial scenario analysis on the power supply based on the risk index data to obtain worst-case adversarial scenario data; and performing iterative two-layer planning on the power supply based on the sensitivity information, constraint enhancement information, and worst-case adversarial scenario data to obtain power supply safety output planning data. This method can improve the overall spatiotemporal redundancy of the flexible DC traction power supply.

Description

Technical Field

[0001] This application relates to the field of intelligent power supply technology, and in particular to a flexible DC traction power supply two-layer planning and optimization method that integrates safety verification and operation simulation. Background Technology

[0002] In traditional technologies, optimization of flexible DC traction power supply for rail transit typically involves using flexible DC converters in substations in conjunction with DC bus voltage droop control. This allows multiple substations to automatically share the load and suppress voltage fluctuations. Energy storage devices such as supercapacitors / batteries are installed on the line or station side to store the energy from regenerative braking of trains before using it for re-acceleration or peak shaving, improving energy recovery rate and reducing peak power. Simultaneously, simple power limiting and segmented switching strategies coordinate the output of converters and energy storage during concentrated train acceleration or braking. However, traditional technologies for flexible DC traction power supply for rail transit lack sufficient spatial and temporal redundancy in overall power supply. Summary of the Invention

[0003] Based on this, it is necessary to provide a method, apparatus, and computer equipment for dual-layer planning optimization of flexible DC traction power supply that integrates safety verification and operation simulation to improve the overall spatiotemporal redundancy of flexible DC traction power supply.

[0004] Firstly, this application provides a two-layer planning and optimization method for flexible DC traction power supply that integrates safety verification and operational simulation, including:

[0005] Spatiotemporal unified adjustment of multi-source planning data and basic operational data of flexible DC traction power supply is performed to obtain dual-layer planning variable data and power supply operation scenario data;

[0006] Based on the bi-level planning variable data and the power supply operation scenario data, a differentiable coupling simulation is performed on the flexible DC traction power supply to obtain the operation trajectory data and sensitivity information.

[0007] The operational trajectory data is subjected to formal security verification to obtain constraint enhancement information and risk indicator data;

[0008] Based on the risk indicator data, the flexible DC traction power supply is subjected to a counter-scenario analysis to obtain the worst-case counter-scenario data;

[0009] Based on the sensitivity information, the constraint enhancement information, and the worst-case scenario data, a two-level planning iterative solution is performed on the flexible DC traction power supply to obtain power supply safety output planning data.

[0010] Secondly, this application also provides a flexible DC traction power supply dual-layer planning and optimization device that integrates safety verification and operation simulation, comprising:

[0011] The data adjustment module is used to perform spatiotemporal unified adjustment of the multi-source planning data and operational basic data of flexible DC traction power supply to obtain dual-level planning variable data and power supply operation scenario data.

[0012] The power supply simulation module is used to perform differentiable coupling simulation of the flexible DC traction power supply based on the dual-level planning variable data and the power supply operation scenario data, so as to obtain the operation trajectory data and sensitivity information.

[0013] The data verification module is used to perform formal security verification on the running trajectory data to obtain constraint enhancement information and risk indicator data;

[0014] The adversarial analysis module is used to perform adversarial scenario analysis on the flexible DC traction power supply based on the risk index data, and obtain the worst-case adversarial scenario data.

[0015] The planning and solving module is used to perform a two-level planning and iterative solution for the flexible DC traction power supply based on the sensitivity information, the constraint enhancement information, and the worst-case adversarial scenario data, to obtain power supply safety output planning data.

[0016] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any step of a flexible DC traction power supply dual-layer planning optimization method that integrates safety verification and operation simulation.

[0017] The aforementioned flexible DC traction power supply dual-layer planning optimization method, device, and computer equipment, which integrates safety verification and operational simulation, achieves consistent expression of planning variables and operational scenarios in a unified time axis and topology space by uniformly adjusting multi-source planning data and operational basic data in a spatiotemporal manner. This significantly reduces simulation deviations and planning rework caused by inconsistent data calibers. Furthermore, it unifies the modeling of the traction power supply network, power electronic control, energy storage SOC, and topology events using a differentiable coupled simulation method, enabling the direct quantification of the impact of planning variables on key indicators such as voltage margin, thermal margin, losses, and over-limit duration in terms of sensitivity. Simultaneously, it introduces formal safety verification, upgrading safety requirements such as voltage, temperature rise, and protection action sequence from post-acceptance to a provable constraint generation mechanism, automatically inputting... Constraint enhancement information is used to dynamically shrink the safe and feasible domain and suppress hidden risks. Then, combined with adversarial scenario analysis targeting risk indicators, the worst-case operating conditions are automatically searched to avoid insufficient coverage caused by relying on limited typical scenario enumeration. This improves the safety reliability under extreme disturbances such as regeneration peaks, fault isolation and recovery, and parameter uncertainty. Finally, driven by sensitivity information, constraint enhancement information, and worst-case adversarial scenario data, a two-level planning iterative solution is completed. The output includes power supply safety planning results including equipment configuration, energy storage configuration, switch topology, and protection settings, as well as their corresponding safety margins / proof evidence. Under the premise of meeting formal safety conditions, it can achieve higher power supply resilience, smaller safety redundancy, and better life cycle cost, and improve the overall spatiotemporal redundancy of flexible DC traction power supply. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is an application environment diagram of a flexible DC traction power supply two-layer planning and optimization method that integrates safety verification and operation simulation in one embodiment.

[0020] Figure 2 This is a flowchart illustrating a two-layer planning and optimization method for flexible DC traction power supply that integrates safety verification and operational simulation in one embodiment.

[0021] Figure 3 This is a structural block diagram of a flexible DC traction power supply dual-layer planning and optimization device that integrates safety verification and operation simulation in one embodiment.

[0022] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0024] This application provides a flexible DC traction power supply dual-layer planning optimization method that integrates safety verification and operation simulation, which can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on other network servers. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0025] In one exemplary embodiment, such as Figure 2 As shown, a two-level planning and optimization method for flexible DC traction power supply that integrates safety verification and operation simulation is provided. This method is applied to... Figure 1 Taking the server in the example, the explanation includes the following steps 202 to 210. Wherein:

[0026] Step 202: Perform spatiotemporal unified adjustment on the multi-source planning data and operational basic data of flexible DC traction power supply to obtain dual-layer planning variable data and power supply operation scenario data.

[0027] Step 204: Based on the bi-level planning variable data and power supply operation scenario data, perform differentiable coupling simulation on the flexible DC traction power supply to obtain operation trajectory data and sensitivity information.

[0028] Step 206: Perform formal security verification on the trajectory data to obtain constraint enhancement information and risk indicator data.

[0029] Step 208: Based on the risk indicator data, conduct a counter-scenario analysis of the flexible DC traction power supply to obtain the worst-case counter-scenario data.

[0030] Step 210: Based on the sensitivity information, constraint enhancement information, and worst-case adversarial scenario data, perform a two-level planning iterative solution for the flexible DC traction power supply to obtain power supply safety output planning data.

[0031] Among them, flexible DC traction power supply is a power supply system that introduces controllable power electronic conversion devices, energy storage devices and reconfigurable switching topologies into the DC traction power supply system, so that the voltage, current and power flow of the traction network can be bidirectionally adjusted according to the control strategy and support fault isolation and power supply reconfiguration.

[0032] Among them, multi-source planning data is a collection of data from multiple sources used in the planning and design phase, including at least data such as line and station topology, traction network parameters, power supply and energy storage equipment ledgers and selection candidate sets, segmentation / connection schemes, cost parameters, and protection adjustable range.

[0033] Among them, the basic operational data is the set of basic input data used for operational simulation and safety assessment, which includes at least train timetables / arrival and departure times, traction and braking characteristic curves, passenger flow or load fluctuations, regenerative energy characteristics, environmental conditions, and fault statistics or typical fault modes.

[0034] Among them, the spatiotemporal unification adjustment is to align, map and unify the data from multiple sources under a unified time reference and a unified spatial topological coordinate system, so as to ensure that the data can be used together in the same simulation step size and the same node-branch model.

[0035] Among them, the bi-level programming variable data is a set of parameterized data used as decision variables in the bi-level programming model, which includes at least the capacity / control capability parameters of power supply devices and energy storage, line configuration parameters, switch topology parameters, and protection setting parameters.

[0036] Among them, power supply operation scenario data is a scenario-based data set describing the combination of operational uncertainties and disturbances, including at least operation graph disturbances, regeneration spikes, topology status, fault location and type, parameter uncertainties, and SOC initial values / boundaries.

[0037] Among them, differentiable coupled simulation is a simulation calculation process that constructs coupled models such as train power injection, traction power supply network equations, converter control and energy storage dynamics into a simulation calculation process that can output gradients or associated sensitivities, so as to support the differentiable mapping of planning variables to operating indicators.

[0038] Among them, the running trajectory data is the system state sequence data that evolves over time by simulation under a given planning scheme and scenario conditions, including at least the time series of node voltage, branch current, power flow, SOC and temperature.

[0039] Sensitivity information refers to the derivative, gradient, or equivalent accompanying sensitivity of the operating index or safety margin relative to the changes in the planning variables.

[0040] Formal safety verification is a process that uses formal methods such as reachable domains / invariant sets, barrier certificates, or timing logic to provably determine whether the operating trajectory meets safety specifications such as voltage, temperature rise, and protection action timing.

[0041] Among them, the constraint enhancement information is new constraint information extracted from the violation evidence or security boundary of formal security verification and can be directly incorporated into the optimization model, including at least cutting plane constraints, barrier certificate boundary constraints, invariant set boundary constraints, or temporal logic constraints.

[0042] Among them, risk indicator data is a set of data used to measure the level of operational risk, including at least aggregated measures such as voltage / temperature rise over-limit amplitude, over-limit duration, protection violation severity, and its worst value or CVaR.

[0043] Among them, adversarial scenario analysis is an analytical process that aims to maximize risk indicators or minimize safety margins by searching for the most unfavorable combination of disturbances within the constraint domain of scenario variables to approximate the worst-case scenario.

[0044] Among them, the worst-case adversarial scenario data is the combination of scenario variables obtained from adversarial scenario analysis that maximizes the risk index or minimizes the safety margin.

[0045] Among them, the bi-level planning iterative solution is a solution process that combines sensitivity information, constraint enhancement information and worst-case adversarial scenarios under the condition of mutual coupling between upper-level planning and lower-level operation simulation, and iteratively updates planning variables until the target and safety conditions are met.

[0046] Among them, the power supply safety output planning data is the planning result data finally output by the two-level planning iteration solution, which includes at least equipment configuration and capacity, energy storage configuration, switch topology scheme, protection setting parameters and corresponding safety margin or safety proof evidence.

[0047] Specifically, it gathers multi-source planning data and basic operational data, including the topology of each line and station of flexible DC traction power supply, traction network parameters, converter and energy storage equipment parameters, candidate configurations of sectional / connection switches, train timetables / arrival and departure times, traction braking characteristics, and protection adjustment range. It performs unified time axis alignment on data with different sampling periods and time bases (e.g., aligning to a unified simulation step size and filling in missing time slices), and performs node mapping on data with different spatial identifiers (e.g., mapping sections, power supply arms, and station equipment to a unified node-branch topology and equipment connection relationship). It also performs consistency and constraint coding on parameter calibers (e.g., per-unit, upper and lower limits, discrete candidate set coding), forming a two-layer planning variable data that includes equipment capacity / control capability, energy storage power and SOC boundary, line configuration, switch topology and protection settings, as well as power supply operation scenario data composed of timetable disturbances, regenerative fluctuations, topology status, environmental conditions, etc.

[0048] Given that the power supply operation scenario data limits the modeling scope, and using bi-level planning variable data as parameters, the traction / braking power injection and regenerative feedback of the train group corresponding to flexible DC traction power supply are modeled as time-varying injection sources. The traction power supply network is modeled as a node voltage-branch current equation. Converter current and voltage limiting control, energy storage charging and discharging and SOC dynamics, and topology events such as switching / fault / recovery are differentiable through continuous relaxation or smooth approximation and incorporated into a unified state equation. Operating trajectory data such as voltage, current, power, SOC, and temperature are obtained through multi-physics coupled simulation. Simultaneously, based on automatic differentiation or adjoint sensitivity framework, the mapping between "planning variables and operating indicators" is differentiated, outputting the gradient or equivalent sensitivity information of indicators such as voltage margin, thermal margin, loss, and over-limit duration relative to the planning variables.

[0049] The system extracts safety-related data from the operational trajectory data, including voltage over-limit amplitude and duration, conductor / device temperature rise and current carrying capacity, and timing sequences of protection action triggering / clearing / reconstruction. It then performs formal judgment according to preset safety specifications (e.g., using invariant set / reachable domain constraints or barrier certificate constraints to ensure the state always remains within the safe set, and using timing logic constraints to verify the selectivity and timing consistency of protection actions). When the verification passes, it outputs safety proof evidence and safety margin; when the verification fails, it outputs the violation trajectory and violation boundary. The violation evidence is then structured and extracted into constraint enhancement information that can be directly used by higher layers (such as cut plane constraints, barrier certificate boundary constraints, reachable domain invariant set boundary constraints, or timing logic constraints). Simultaneously, risk indicator data is calculated based on over-limit severity, over-limit probability, or CVaR.

[0050] Using risk indicator data as the objective or evaluation function, the system first performs inverse analysis on the over-limit amplitude, over-limit duration, and severity of protection violations in the risk indicator data to obtain adversarial seed parameters (representing the direction or combination of the most sensitive scenario variables). Then, within the scenario variable constraint domain, adversarial search is performed on variables such as timetable perturbations, regeneration spikes, fault location and type, parameter uncertainty, initial SOC value, and boundaries to generate candidate adversarial scenarios. A risk proxy model and confidence bounds under uncertainty are used to filter and retain a high-risk subset, ultimately obtaining the worst-case adversarial scenario data that increases the risk indicators. If necessary, Latin hypercube sampling is used to perform perturbation expansion verification on the worst-case candidate scenarios to confirm that they still maintain worst-case characteristics under neighborhood perturbations.

[0051] Constraint enhancement information is incorporated into the bi-level programming model to shrink and form a safe feasible region. Sensitivity information is used to perform directional pre-updates of planning variables within the safe feasible region to obtain better initial values. Worst-case adversarial scenario data is solidified into equivalent worst-case constraints, and voltage / thermal / temporal margins are reconstructed into a safety margin objective function. Furthermore, differentiable equivalence is used to map the lower-level simulation to optimizable equivalent constraints / equivalent objectives. Evidence-driven backtracking updates are performed in conjunction with violation boundaries or barrier certificate boundaries to correct initial values. Finally, iterative optimization solvers (such as sequential quadratic programming, interior-point method, alternating iteration, or decomposition coordination) are used to iteratively solve the equivalent optimization problem until convergence and formal safety verification conditions are met. The output includes power supply safety output planning data, including equipment configuration, energy storage configuration, switch topology, and protection settings, along with corresponding provable safety evidence or safety margin indicators.

[0052] In the aforementioned flexible DC traction power supply dual-layer planning optimization method that integrates safety verification and operational simulation, the planning variables and operational scenarios are consistently expressed in a unified time axis and topology space by adjusting the multi-source planning data and operational basic data in a unified spatiotemporal manner. This significantly reduces simulation deviations and planning rework caused by inconsistent data calibers. Furthermore, the traction power supply network, power electronic control, energy storage SOC, and topology events are modeled in a unified manner using a differentiable coupling simulation method, enabling the impact of planning variables on key indicators such as voltage margin, thermal margin, loss, and over-limit duration to be directly quantified in the form of sensitivity. At the same time, a formal safety verification is introduced, upgrading safety requirements such as voltage, temperature rise, and protection action sequence from post-acceptance to a provable constraint generation mechanism, automatically outputting constraint increments. Strong information is used to dynamically shrink the safe and feasible domain and suppress hidden risks. Combined with adversarial scenario analysis targeting risk indicators, the worst-case operating conditions are automatically searched to avoid insufficient coverage caused by relying on limited typical scenario enumeration. This improves the safety reliability under extreme disturbances such as regeneration peaks, fault isolation and recovery, and parameter uncertainty. Finally, driven by sensitivity information, constraint enhancement information and worst-case adversarial scenario data, a two-level planning iterative solution is completed. The output includes power supply safety planning results including equipment configuration, energy storage configuration, switch topology and protection settings, as well as their corresponding safety margins / proof evidence. Under the premise of meeting formal safety conditions, it can achieve higher power supply resilience, smaller safety redundancy and better life cycle cost, and improve the overall spatiotemporal redundancy of flexible DC traction power supply.

[0053] In an exemplary embodiment, based on risk index data, a worst-case scenario analysis is performed on the flexible DC traction power supply to obtain worst-case scenario data, including steps 302 to 306. Wherein:

[0054] Step 302: Invert and locate the over-limit amplitude data, over-limit duration data, and protection violation severity data in the risk indicator data to obtain the adversarial seed parameter data.

[0055] Step 304: Based on the adversarial seed parameter data and the scenario variable constraint data, perform adversarial search on the power supply operation scenario data to obtain the initial adversarial scenario data.

[0056] Step 306: Perform perturbation expansion of the initial adversarial scenario data using Latin hypercube sampling to obtain the worst-case adversarial scenario data.

[0057] Among them, the over-limit amplitude data is the numerical data of the maximum deviation of key traction power supply state quantities (such as pantograph voltage, bus voltage, branch current or temperature) from their allowable upper or lower limits.

[0058] Among them, the over-limit duration data is the numerical data of the cumulative duration or the maximum continuous duration during which the critical state quantity of traction power supply exceeds the allowable threshold and remains in the over-limit state.

[0059] Among them, the severity data of protection violations is the quantitative result data of the severity when the protection action is inconsistent with the preset protection specification, which includes at least quantitative indicators of false action / refusal to act, action delay, selective mismatch or reconstruction timing violation.

[0060] Inverse positioning is an analytical process that uses observed risk indicators to infer the scenario variables and their value directions or combinations that lead to increased risk.

[0061] Among them, the adversarial seed parameter data is parameterized information data obtained from inversion localization and used to initialize or guide adversarial search. It represents the priority set, value tendency and variable coupling relationship of variables in high-risk scenarios.

[0062] Among them, the scenario variable constraint data is a set of rule data that constrains the range and combination relationship of scenario variables, including at least boundary constraints, discrete value constraints, and coupling constraints between variables.

[0063] Among them, adversarial search is a computational process that optimizes the combination of scenario variables within the constraint domain of scenario variables with the goal of maximizing risk indicators or minimizing safety margins in order to approximate the most unfavorable scenario.

[0064] Among them, the initial adversarial scenario data is the first set or batch of scenario variable combinations that meet the scenario constraints and significantly increase the risk indicators obtained by adversarial search, along with their corresponding risk assessment results.

[0065] Among them, the perturbation expansion of Latin hypercube sampling is an expansion process that takes the initial adversarial scenario as the center and uses Latin hypercube sampling to generate a more comprehensive neighborhood sample set within the perturbation range of each scenario variable for worst-case verification.

[0066] Specifically, the data on the magnitude of exceeding limits, the duration of exceeding limits, and the severity of protection violations in the risk indicator data corresponding to each scenario are uniformly normalized and a comprehensive risk data is constructed. Based on the available correlation information between the comprehensive risk data and scenario variables, an inversion localization is performed. Specifically, the gradient direction, sensitivity ranking, or contribution decomposition of the comprehensive risk data on the scenario sample set are solved to identify the subset of scenario variables that most significantly drive the risk and their value tendencies. The subset of variables, the variable value direction, and the coupling relationship between variables are parameterized and output as adversarial seed parameter data.

[0067] Based on scenario variable constraint data, a feasible domain for scenario variables (including boundary constraints, coupling constraints, and discrete value constraints) is constructed. Using adversarial seed parameter data as the starting point or direction of the search, adversarial optimization searches are performed on variables such as timetable disturbances, regeneration spikes, fault locations and types, parameter uncertainties, and initial / boundary values ​​of State of Charge (SOC) in the power supply operation scenario data within the feasible domain. During the search process, the comprehensive risk score or its approximate proxy is used as the objective function, and a "combined generation—proxy evaluation—local search" approach is employed to gradually approach high-risk areas, ultimately obtaining initial adversarial scenario data that increases the risk indicators.

[0068] A perturbation neighborhood is constructed centered on the initial adversarial scenario data. Based on the scenario variable constraint data, perturbation ranges and distribution assumptions are set for each scenario variable. Latin hypercube sampling is performed on this neighborhood to generate a more comprehensive extended verification scenario set. Each extended verification scenario is then input into the simulation and risk assessment process to calculate the corresponding over-limit amplitude, over-limit duration, and protection violation severity, thereby obtaining the risk distribution of the extended verification scenario set. Based on this, the sample with the highest risk score or the worst-case scenario satisfying the lower confidence bound is selected as the "worst-case" scenario, and its risk value, corresponding scenario variable value, and offset from the initial scenario are output, forming the worst-case adversarial scenario data.

[0069] In this embodiment, adversarial seed parameters are obtained by inverting and locating risk indicators such as over-limit amplitude, over-limit duration, and severity of protection violations. This transforms the worst-case search from "blind enumeration" to a targeted search that "infers the direction of disturbance from the risk results," thereby significantly improving the efficiency of high-risk scenario discovery and reducing the computational load of invalid scenarios. Furthermore, adversarial search is performed based on adversarial seed parameters within the scenario variable constraint domain, which can quickly construct initial adversarial scenarios that meet the constraints and have significantly increased risks, enhancing the coverage of complex coupled uncertainties such as regeneration spikes, fault disturbances, and operational graph fluctuations. Finally, the initial adversarial scenarios are validated by neighborhood perturbation expansion through Latin hypercube sampling, which can achieve statistical confirmation of worst-case scenarios and robustness with limited samples, avoiding planning deviations caused by the randomness of single-point worst-case scenarios.

[0070] In an exemplary embodiment, adversarial search is performed on the power supply operation scenario data based on adversarial seed parameter data and scenario variable constraint data to obtain initial adversarial scenario data, including steps 402 to 408. Wherein:

[0071] Step 402: Project the boundary constraint data, coupling constraint data, and discrete value constraint data in the scenario variable constraint data to obtain feasible adversarial variable domain data.

[0072] Step 404: Based on the adversarial seed parameter data, combine the power supply variables of each scenario in the power supply operation scenario data to obtain the candidate adversarial scenario set data.

[0073] Step 406: Perform risk proxy screening on the candidate adversarial scenario set data to obtain a simplified adversarial scenario set data.

[0074] Step 408: Perform a local search based on the simplified adversarial scenario set data to obtain the initial adversarial scenario data.

[0075] Among them, boundary constraint data are upper and lower limits or interval rule data that limit the range of variables in each scenario (such as the range of regeneration peak rate, upper and lower limits of SOC, range of disturbance amplitude, etc.).

[0076] Among them, coupling constraint data is rule data that limits the consistency or linkage between multiple scenario variables (such as the matching relationship between fault location and switch status, feedback capability and SOC boundary, topology status and power supply arm configuration).

[0077] Among them, discrete value constraint data is constraint rule data that allows scenario variables to take values ​​only from a finite set or a set of integers (such as fault segment number, fault type enumeration, switch status enumeration, train formation type enumeration, etc.).

[0078] Among them, feasible adversarial variable domain data refers to the feasible search space or its parameterized description data of scene variables under the condition that boundary constraints, coupling constraints and discrete value constraints are satisfied simultaneously.

[0079] Among them, the scenario power supply variables are a set of disturbances and uncertainties used to characterize the power supply operation scenario, including at least timetable disturbances, regeneration peaks, fault location and type, parameter uncertainty, and SOC initial value / boundary variables.

[0080] Among them, the candidate adversarial scenario set data is a data set consisting of multiple candidate scenario records and their variable value vectors obtained by combining scenario power variables within the feasible adversarial variable domain.

[0081] Among them, risk proxy screening is a screening process that uses risk proxy models or fast approximation scorers to assess the risk level of candidate scenarios at low cost and then sorts and truncates them accordingly.

[0082] Among them, the streamlined adversarial scenario set data is a subset of high-risk and representative candidate scenario data that has been filtered by a risk proxy and is used for subsequent more refined local searches or high-fidelity verification.

[0083] Local search is an optimization search process that uses simplified candidate scenarios as initial values ​​and performs neighborhood perturbation and iterative improvement on scenario variables within the feasible adversarial variable domain to further increase the risk indicators.

[0084] Specifically, the boundary constraints in the scenario variable constraint data are represented as upper and lower limit intervals or allowable sets for each scenario variable; the coupling constraints in the scenario variable constraint data are represented as feasible relationships between variables (e.g., consistency constraints of "fault location - switch status - power supply arm topology", coupling constraints of "regenerative spike - SOC boundary - feedback capability"); and the discrete value constraints in the scenario variable constraint data are represented as enumeration sets or integer fields. Based on the above data, a feasibility projection is performed on the candidate scenario variable vectors. Continuous variables exceeding the boundaries are truncated back to intervals, combinations of variables that do not satisfy the coupling relationship are remapped to combinations that satisfy the relationship according to preset repair rules, and discrete variables are mapped to allowable value sets according to nearest neighbors or feasible enumerations. This yields feasible adversarial variable domain data describing the feasible search space (e.g., output in the form of constraint sets, feasible domain descriptors, or projectible operators).

[0085] Using the subset of highly sensitive variables, their value directions, and priorities provided by the adversarial seed parameter data as guidance for combination generation, the system prioritizes the discrete-continuous hybrid combination of power supply variables for each scenario within the feasible adversarial variable domain. These variables include timetable disturbance variables (such as train interval compression / expansion, departure phase shift), regeneration peak variables (such as peak multiplier, duration, number of trains regenerating simultaneously), fault variables (such as fault section, fault type, and occurrence time), parameter uncertainty variables (such as line resistance, catenary voltage drop, converter current limiting deviation), and initial / boundary SOC values. During combination, constrained enumeration or hierarchical sampling is used for discrete variables, while segmented value taking or low-difference sampling is used for continuous variables along the seed direction. Each variable combination is encapsulated as a candidate adversarial scenario record, ultimately forming a candidate adversarial scenario set data containing multiple candidate scenarios and their variable vectors, constraint labels, and initial risk priors.

[0086] For each candidate adversarial scenario, a proxy input feature vector is constructed (e.g., the scenario variables themselves and their consistency with the adversarial seed direction, the combined features of key coupling variables, etc.), and their risk level is quickly analyzed using a risk proxy model. The risk proxy model can employ a regression / classification model trained on historical simulation samples or a physics-inspired fast approximation scorer to output proxy estimates or a comprehensive proxy risk score for the out-of-limit amplitude, out-of-limit duration, and severity of protection violations. Subsequently, the proxy risk scores are sorted from high to low, and Top-K truncation or threshold filtering is performed in conjunction with coverage constraints (e.g., maintaining diversity across different fault locations / different regeneration peak intensities) to retain a high-risk and representative subset of candidates, thus obtaining a streamlined set of adversarial scenario data.

[0087] Each scenario variable combination in the simplified adversarial scenario dataset is used as the initial value for local search. Within the feasible adversarial variable domain, coordinate search, gradient ascent, or trust region perturbation are performed on continuous variables, while neighborhood enumeration or swapping operations are performed on discrete variables. At each candidate update, a fast risk assessment (which can be a low-cost simulation or a more accurate surrogate model) is invoked to calculate risk indicators. Iterative updates are performed according to the criterion of "accept if risk increases, otherwise back off / reduce step size." When the iteration count, improvement threshold, or risk convergence condition is met, the scenario variable combination with the highest risk and its corresponding risk assessment result are output as the initial adversarial scenario data.

[0088] In this embodiment, by first projecting the boundary, coupling, and discrete value constraints into a feasible adversarial variable domain, infeasible combinations can be eliminated at the initial stage of the search, reducing subsequent repair costs and thus improving the effectiveness and computational efficiency of adversarial scenario generation. Then, by using adversarial seed parameters to guide the combination generation of scenario power supply variables, search resources can be focused on sensitive variables that are more likely to trigger limit violations and protection breaches, significantly improving the hit rate of high-risk candidate scenarios. Subsequently, risk proxy is used to quickly screen candidate scenarios, enabling risk stratification and representative retention of a large-scale candidate set under low-cost evaluation, reducing the burden of full high-fidelity simulation. Finally, local search is performed on a simplified high-risk subset, which can further approximate the initial adversarial scenario with higher risk indicators while satisfying constraints, enhancing the ability to discover the most unfavorable working conditions under complex coupling perturbations.

[0089] In an exemplary embodiment, risk proxy screening is performed on the candidate adversarial scenario set data to obtain a simplified adversarial scenario set data, including steps 502 to 508. Wherein:

[0090] Step 502: Perform sensitivity injection feature processing on the candidate adversarial scenario set data to obtain risk agent feature data;

[0091] Step 504: Based on the violation constraint type data, violation boundary data, or time-series violation data in the constraint enhancement information, perform verification feedback calibration on the risk agency feature data to obtain calibrated risk agency feature data;

[0092] Step 506: Perform uncertainty definition analysis on the calibration risk agency characteristic data to obtain the lower confidence boundary data of agency risk;

[0093] Step 508: Sort and truncate the confidence boundary data under agent risk to obtain a simplified set of adversarial scenario data.

[0094] Among them, sensitivity injection feature processing is a process of fusing and mapping the scene variable features of candidate adversarial scenarios with sensitivity information obtained from differentiable simulation in order to construct proxy input features that can reflect the "scene trigger risk capability".

[0095] Among them, risk agency feature data is feature vector or feature set data formed after sensitivity injection featureization.

[0096] Among them, the violation constraint type data is the identification data of the category to which the violation identified by the formal safety verification belongs, including at least types such as voltage over-limit, temperature rise over-limit, protection maloperation / failure to operate, or protection selective mismatch.

[0097] Among them, violation boundary data are parametric representation data of the safety boundary or violation boundary output by formal safety verification, which at least include information such as boundary location, boundary margin or boundary normal direction used to characterize "approaching violation".

[0098] Among them, timing violation data is violation description data when the protection action or reconstruction process fails to meet the preset timing specifications, including at least the violation event chain, the trigger time, and the identifier of the link where the timing constraint is violated.

[0099] Among them, the verification feedback calibration is a calibration process that uses violation constraint type data, violation boundary data or time-series violation data to correct the risk agency characteristics or agency score, so that the agency assessment is closer to the real safety boundary and violation mechanism.

[0100] Among them, the calibrated risk proxy feature data is the risk proxy feature data after verification and feedback calibration, which makes it more able to distinguish high-risk candidate scenarios near the security boundary at the statistical and mechanistic levels.

[0101] Uncertainty definition analysis is an analytical process that quantifies the uncertainty of agent risk prediction and constructs confidence boundaries or confidence intervals accordingly to assess the credibility of the agent's outcome.

[0102] Among them, the confidence boundary data under proxy risk is a conservative estimate of the risk level of candidate scenarios given a confidence level. It is usually obtained by combining proxy risk point estimation and uncertainty for robust screening.

[0103] Among them, sorting truncation is a screening process that sorts candidate scenarios according to the confidence boundary under agency risk and retains subsets according to Top-K, threshold or quantile rules to form a simplified set.

[0104] Specifically, for each candidate adversarial scenario in the candidate adversarial scenario set data, its scenario variable vector (such as timetable perturbations, regeneration spikes, fault location and type, parameter uncertainty, SOC initial value / boundary, etc.) is constructed into basic features. Key sensitivity components related to the scenario (such as the gradient of risk indicators on planning variables like equipment capacity, energy storage power, switch topology, and protection settings, or the sensitive direction of voltage / thermal margin to scenario variable perturbations) are extracted from the sensitivity information output by differentiable coupled simulations. These sensitivity components are then injected into the basic features using methods such as weighted concatenation, cross-term construction, or directional consistency (the angle / inner product between the scenario perturbation direction and the gradient direction) to form risk proxy feature data that can simultaneously characterize both the "scenario itself" and the "triggering capability for system weaknesses."

[0105] The violation constraint type data extracted from the constraint enhancement information (such as overvoltage, undervoltage, temperature rise exceeding limits, protection malfunction / failure to operate, and non-compliance with action timing) are mapped to feature calibration labels, and the violation boundary data or barrier certificate boundary is characterized as the safety boundary normal direction, boundary distance, or boundary margin. Subsequently, the risk proxy feature data is subjected to verification feedback calibration. For example, the feature weights that are consistent with the boundary normal are increased for candidate scenarios near the violation boundary, the penalty or enhancement coefficients are applied to the corresponding feature dimensions for scenarios that trigger specific violation types, and the timing consistency defect features are introduced for scenarios that violate timing logic, so that the proxy features are more statistically consistent with the "real safety boundary" and "violation mechanism", thereby obtaining calibrated risk proxy feature data.

[0106] For each calibration risk proxy feature data, an uncertainty estimation mechanism is used to quantify the uncertainty of the proxy risk prediction. For example, the "prediction credibility" can be characterized by the ensemble model divergence (multi-model output variance), the prediction variance approximated by Monte Carlo dropout, or the out-of-distribution detection score based on feature space distance. Based on this, a confidence bound for the proxy risk is constructed (preferably a lower confidence bound to represent a conservative worst-case risk estimate, or an upper confidence bound constructed under a specific definition). For example, the proxy risk point estimate and uncertainty are combined according to a preset confidence level to form a confidence bound value, so that the corresponding confidence bound tends to be more conservative when the uncertainty is large, thereby obtaining the lower confidence bound data for the proxy risk corresponding to each candidate adversarial scenario.

[0107] The candidate adversarial scenario set is sorted in descending order using the lower confidence bound of the proxy risk corresponding to each candidate adversarial scenario as the sorting key. Then, truncation retention is performed according to the preset retention scale or threshold rules, such as Top-K retention, quantile threshold retention, or fixed risk threshold retention. At the same time, coverage constraints can be optionally introduced to avoid retaining only scenarios with the same type of fault location or the same type of regeneration peak intensity. This ensures that scenario diversity is taken into account while prioritizing high risk. The final output is a simplified set of adversarial scenario data.

[0108] In this embodiment, by performing sensitivity injection featureization on the candidate adversarial scenario set, the proxy evaluation not only reflects the scenario variables themselves, but also explicitly characterizes the scenario's ability to trigger system weaknesses and risk indicators, thereby improving the resolution of high-risk scenario identification. Furthermore, by using constraint enhancement information to verify and calibrate the proxy features, the violation mechanism and security boundary direction obtained from formal security verification can be embedded into the screening criteria, reducing misjudgments and biases caused by relying solely on statistical proxy models. Then, by defining the uncertainty, a confidence bound under the proxy risk is obtained, which can automatically adopt a more conservative risk estimate when the proxy model prediction is uncertain, improving the robustness of the screening results to out-of-distribution disturbances and parameter uncertainties. Finally, sorting and truncation based on the confidence bound can quickly retain a high-risk and credible representative subset from a large-scale candidate set at low cost, thereby significantly reducing the computational burden of subsequent high-fidelity simulation and local search and improving the efficiency of worst-case detection.

[0109] In an exemplary embodiment, based on sensitivity information, constraint enhancement information, and worst-case adversarial scenario data, a two-level planning iterative solution is performed on the flexible DC traction power supply to obtain power supply safety output planning data, including steps 602 to 606. Wherein:

[0110] Step 602: Perform feasible region shrinkage processing on the constraint enhancement information to obtain safe feasible region data;

[0111] Step 604: Based on the sensitivity information, pre-update the planning variables within the safe and feasible domain data to obtain pre-updated planning variable data;

[0112] Step 606: Using the pre-updated planning variable data as initial values, and based on the worst-case adversarial scenario data, perform a two-level planning iteration solution for the flexible DC traction power supply to obtain the power supply safety output planning data.

[0113] Among them, the feasible region shrinkage process is to transform the constraint enhancement information into new safety constraints and incorporate them into the original planning constraint system, so as to take the intersection or projection correction of the original planning feasible region and thus eliminate the unsafe operation area.

[0114] Among them, the safe and feasible domain data is the set of feasible planning variables that still satisfy the original engineering constraints and the new safety constraints after the feasible domain is shrunk, and its parameterized representation data.

[0115] Among them, the planning variables are the set of decision variables in the bi-level programming model that can be optimized and used to determine the system configuration and setting scheme, including at least the power supply device and energy storage capacity / control parameters, line configuration parameters, switch topology parameters and protection setting parameters.

[0116] Among them, the pre-updated planning variable data is the planning variable value data obtained as the initial value for the two-level iterative solution after adjusting the current planning variables according to the sensitivity information within the safe and feasible domain.

[0117] Specifically, the constraint enhancement information is structured into a new set of constraints that can be directly incorporated into the upper-level planning model. For example, cutting plane constraints are represented as linear inequalities, barrier certificate / invariant set boundary constraints are represented as differentiable or piecewise differentiable inequalities, and temporal logic constraints are represented as satisfiability constraints of action sequence and time window. These new constraints are then merged with existing constraints such as equipment capacity, site selection, topology, and tuning range to form an updated constraint set. Then, intersection operations or projection corrections are performed on the original planning feasible region to obtain safe feasible region data that satisfies the new safety constraints (which can be represented by constraint lists, feasible region descriptors, or projective operators). This allows for the early exclusion of scheme areas with verified violation mechanisms at the planning level and ensures that subsequent searches are conducted within the safe boundaries.

[0118] Gradient directions related to the planning objectives and safety margins are selected from sensitivity information (e.g., gradients of risk indicators with respect to equipment capacity, energy storage power, switch topology parameters, and protection setting parameters). Directional updates are then performed on the current planning variables under the constraints of the safe feasible region. Specifically, projected gradient steps, constrained trust region steps, or component updates sorted by sensitivity can be used. The update step size is adaptively controlled by the safety margin or the magnitude of risk reduction. After each update, the results are feasiblely projected to ensure they still fall within the safe feasible region, obtaining pre-updated planning variable data. This ensures that the optimization process proceeds from the outset along the effective direction of "risk reduction / margin improvement."

[0119] Worst-case scenario data is solidified into a set of input scenarios or equivalent worst-case constraints (including combinations of faults, regeneration spikes, timetable disturbances, and parameter uncertainties) for two-layer operational simulation. Pre-updated planning variable data serves as the initial values ​​for the two-layer solution, with alternating iterations performed between updating the upper-layer planning variables and solving the lower-layer operational simulation / control response. The upper layer updates decision variables such as equipment configuration, energy storage configuration, topology, and protection settings under the constraints of the safe feasible region. The lower layer solves the operational response under the worst-case scenario with fixed upper-layer variables and returns operational costs and risk indicators. During the iteration process, constraint enhancement information is continuously added and risk assessments are updated until convergence criteria are met and formal safety checks no longer produce violations. The final output includes power supply safety output planning data containing power supply device and energy storage configurations, line and switch topology schemes, protection setting parameters, and corresponding safety margins / proof evidence.

[0120] In this embodiment, by directly using the constraint enhancement information output by the formal security check to shrink the feasible region, the solution areas with exposed violation mechanisms can be eliminated in advance during the planning stage, forming a safe feasible region, thereby reducing invalid searches and rework. Furthermore, by using the sensitivity information obtained from differentiable coupled simulation to pre-update the planning variables within the safe feasible region, the iteration can proceed from the beginning along the effective direction of risk reduction and safety margin improvement, significantly improving convergence efficiency and reducing overly conservative redundant configurations. Finally, by using the pre-updated results as initial values ​​and performing bi-level planning iteration under worst-case adversarial scenario constraints, the output solution can be ensured to still meet safety requirements under extreme disturbances and uncertainties, thereby improving the safety reliability and resilience of the power supply system, and achieving better life-cycle cost and resource allocation efficiency while meeting safety requirements.

[0121] In an exemplary embodiment, using pre-updated planning variable data as initial values, and based on worst-case scenario data, the flexible DC traction power supply undergoes a two-level planning iteration solution to obtain power supply safety output planning data, including steps 702 to 710. Wherein:

[0122] Step 702: Solidify the fault data, regeneration peak data, timetable disturbance data and parameter uncertainty data in the worst-case scenario data to obtain the worst-case equivalent constraint data.

[0123] Step 704: The voltage margin data, thermal margin data, and protection action timing margin data corresponding to the worst-case equivalent constraint data are reconstructed in a target-oriented manner to obtain the safety margin objective function data.

[0124] Step 706: Combine the pre-updated planning variable data and the safety margin objective function data to perform differentiable bi-level equivalence, and obtain bi-level equivalent optimization problem data.

[0125] Step 708: Based on the violation boundary data or barrier certificate boundary data in the constraint enhancement information, perform evidence-driven backtracking update on the bi-layer equivalent optimization problem data to obtain the corrected initial value data.

[0126] Step 710: Using the corrected initial value data as the initial value, sequential quadratic programming is used to iteratively solve the two-level equivalent optimization problem data to obtain the power supply safety output planning data.

[0127] Among them, fault data is parameterized data describing the fault conditions of the traction power supply system, including at least the fault location, fault type, time of occurrence, and isolation / recovery related action constraints.

[0128] Among them, regenerative peak data is parameterized data that describes the characteristics of the sudden increase in feedback power caused by train regenerative braking, including at least the peak amplitude, duration, rise slope, and number of trains regenerating simultaneously.

[0129] Among them, timetable disturbance data are disturbance parameter data that describe the deviation of the train timetable from the planned operation state, including at least departure phase offset, train interval compression / expansion, and section running time deviation.

[0130] Among them, parameter uncertainty data are range or distribution description data that reflect the deviation or fluctuation of line and equipment parameters in actual operation, including at least line resistance / voltage drop coefficient, converter current limiting deviation, load model error, etc.

[0131] Scenario solidification is the process of converting the disturbances and uncertainties in the worst-case scenario into definite parameters, interval parameters, or constraints that can be directly incorporated into the optimization model.

[0132] Among them, the worst-case equivalent constraint data is a set of constraints or its parameterized representation data obtained by fixing the scenario and used to force the worst-case working condition to be reflected in the planning solution.

[0133] Among them, voltage margin data is the quantified data of the minimum safety margin of the voltage of key nodes in the operating trajectory relative to the upper and lower limits of the allowable range.

[0134] Among them, thermal margin data is the quantitative data of the minimum safety margin of temperature rise / temperature of conductor or device relative to thermal limit or current-carrying thermal stability limit.

[0135] Among them, the protection action timing margin data is the quantitative data of the minimum time margin of the actual triggering / completion time of the protection action (tripping, blocking, reconfiguration, etc.) relative to the specified time window and action sequence specification.

[0136] Among them, objective-oriented reconstruction is a process of converting multiple types of safety margin indices into objective function forms that can be used for optimization by means of maximization, minimization, penalty terms or smooth approximation.

[0137] Among them, the safety margin objective function data is the objective function expression and its parameter data obtained by objective-oriented reconstruction, which is used to drive the planning scheme to improve the worst-case safety margin.

[0138] Differentiable bilayer equivalence is a process that uses automatic differentiation or adjoint sensitivity to map the lower-level simulation and control response into differentiable equivalent constraints / equivalent objectives, thereby transforming the bilayer problem into an equivalent problem that can be numerically optimized.

[0139] Among them, the bi-level equivalent optimization problem data is the problem description data that can be directly input into the optimization solver, consisting of the equivalent objective function, equivalent constraint set, variable domain and initial value obtained through differentiable bi-level equivalence.

[0140] Among them, the violation boundary data is the parameterized representation data of the violation occurrence boundary output by the formal security verification, which includes at least the boundary position, boundary distance or boundary normal direction.

[0141] Among them, the barrier certificate boundary data is the parameterized representation data of the security set boundary characterized by the barrier certificate or the control barrier function, which includes at least the boundary conditions with a barrier function value of zero and their gradient information.

[0142] Evidence-driven backtracking update is an update process that backtracks, projects, or directionally corrects the initial value of the solution based on violation boundary data or barrier certificate boundary data, so as to make the iteration starting point far away from the violation region and more conducive to convergence.

[0143] Among them, the corrected initial value data is the initial value data of the planning variables obtained after evidence-driven backtracking update, which satisfies the safe and feasible region and is used for subsequent iterative solutions.

[0144] Sequential quadratic programming is a numerical optimization method that approximates the optimal solution of the original problem by performing quadratic / linear approximations on nonlinear objectives and constraints and iteratively solving a series of quadratic programming subproblems.

[0145] Specifically, the variables describing disturbances and uncertainties in the worst-case scenario data are explicitly transformed into parameters or constraints that can be incorporated into the optimization model. Fault data is solidified into constraint parameters such as fault location and type, fault occurrence time, and isolation / recovery action window. Regeneration peak data is solidified into the boundary or injection envelope of regeneration power peak, duration, and number of trains regenerating simultaneously. Timetable disturbance data is solidified into the limiting intervals of departure phase offset, train interval compression / expansion, and section running time deviation. Parameter uncertainty data is solidified into the interval set or scenario values ​​of uncertain parameters such as line resistance / inductance, converter current limiting deviation, and contact network voltage drop coefficient. By uniformly expressing the above solidified results as a set of inequalities or scenario parameter sets that "must be satisfied in the worst case", the equivalent constraint data for the worst case is formed.

[0146] Under worst-case equivalent constraints, key safety margin indicators are extracted from operational simulations or proxy assessments. These include the minimum margin of pantograph / bus voltage relative to upper and lower limits, the minimum margin of conductor and device temperature rise relative to thermal limits, and the minimum sequential margin of protection actions (tripping, blocking, reconfiguration) relative to specified time windows and sequence specifications. These margins, with their multiple indicators and different dimensions, are then reconstructed into an optimizable objective function. For example, a max-min objective of "maximizing minimum margin" can be used, a differentiable objective can be constructed using a smooth approximation of the minimum margin (such as log-sum-exp or p-norm), or each margin can be converted into a risk penalty term and weighted summed to obtain the safety margin objective function data.

[0147] Using pre-updated planning variable data as initial values ​​for upper-level decisions, the lower-level operational response (train power injection—network voltage and current—converter control—energy storage SOC / temperature dynamics) is considered a differentiable mapping of upper-level variables and scenario parameters. Through automatic differentiation or adjoint sensitivity, the lower-level optimality conditions / simulation implicit equations are transformed into equivalent gradients, equivalent constraints, or equivalent objective terms that can be directly used by the upper level. Based on this, the safety margin objective function, worst-case equivalent constraints, and the original upper-level engineering constraints are uniformly aggregated to form bi-level equivalent optimization problem data that can be processed by a numerical optimizer (e.g., output in the form of a single-level nonlinear programming, an optimization form with implicit constraints, or a differentiable constraint system).

[0148] When the formal safety verification outputs violation boundaries or barrier certificate boundaries, the boundary normal, boundary distance, or barrier function gradient information representing the direction of the safety boundary are extracted from the safety margin objective function data, and these are used as the initial value correction process of the equivalent optimization problem as a "backtracking update" rule. Specifically, a backtracking or projection step is performed near the pre-updated planning variables in a direction away from the violation boundary, while priority correction weights are applied to the key variable components that trigger violations (such as insufficient capacity, excessively tight current limiting, unreasonable SOC boundaries, and excessive protection settings). By projecting the corrected variables back onto the safe feasible region and satisfying the worst-case equivalent constraints, corrected initial value data that is closer to the inside of the safety boundary and conducive to convergence is obtained.

[0149] The initial value data input sequence quadratic programming (SQP) solution framework is modified. In each iteration, the two-level equivalent optimization problem is approximated by a quadratic approximation and a linearized constraint approximation to obtain the search direction and step size and update the planning variables. At the same time, the satisfaction of the worst-case equivalent constraints and safety margin objectives is continuously checked during the iteration process, and the monotonic improvement of the objectives and the maintenance of feasibility are ensured by combining line search or trust region strategies. When the convergence criteria (such as KKT residuals, objective improvement thresholds and constraint violation thresholds) are met and formal safety check violations are no longer triggered, the output includes power supply safety output planning data containing power supply device and energy storage configuration, line and switch topology schemes, protection setting parameters, and corresponding safety margins / proof evidence.

[0150] In this embodiment, by solidifying worst-case adversarial scenarios such as faults, regeneration spikes, timetable disturbances, and parameter uncertainties into equivalent constraints, extreme operating conditions can be directly embedded into the planning and solving process, avoiding safety blind spots caused by optimizing only under typical scenarios from the source. Furthermore, voltage margin, thermal margin, and protection action timing margin are reconstructed into a safety margin objective function, elevating the optimization objective from simple cost or loss optimization to improving safety margins for worst-case conditions, thereby reducing overly conservative redundant configurations and improving the efficiency of safety boundary utilization. Finally, through differentiable two-layer equivalence, the operation simulation and control response are integrated. The problem should be mapped to an optimizable equivalent problem, allowing the impact of planning variables on the safety margin to be utilized in a gradient manner, significantly accelerating convergence and improving solution stability. Simultaneously, evidence-driven backtracking updates using violation boundaries or barrier certificate boundaries can proactively avoid violation regions and provide better initial values ​​in the early stages of iteration, reducing the iteration cost of repeatedly triggering violations. Finally, sequential quadratic programming is used to iteratively solve the equivalent optimization problem, obtaining power supply safety output planning data that satisfies worst-case constraints and safety margin objectives within a mature numerical optimization framework. This enhances the provable safety and system resilience of the planning results under extreme disturbances and uncertainties.

[0151] In an exemplary embodiment, based on the bi-level programming variable data and power supply operation scenario data, a differentiable coupled simulation of the flexible DC traction power supply is performed to obtain operation trajectory data and sensitivity information, including steps 802 to 808. Wherein:

[0152] Step 802: Perform continuous relaxation or smoothing approximation on the switch state switching data, fault isolation data, and recovery action data in the power supply operation scenario data to obtain differentiable topology event data.

[0153] Step 804: Based on the bi-level programming variable data, perform differentiable uniformity processing on the traction load power injection data corresponding to the flexible DC traction power supply to obtain differentiable power injection data.

[0154] Step 806: Based on the differentiable topology event data and differentiable power injection data, perform multi-physics coupling simulation to solve the power supply status of the flexible DC traction power supply and obtain the running trajectory data.

[0155] Step 808: Perform sensitivity calculation on the trajectory data to obtain sensitivity information.

[0156] Among them, the switch status switching data is the data on the time, duration window and corresponding topology change information of the opening and closing switching of switchgear such as sectionalizing switches and tie switches in the operation scenario.

[0157] Among them, fault isolation data includes data such as the location and type of fault, the sequence of isolation actions, and the timing of actions involved in isolating the faulty section after a fault occurs.

[0158] Among them, the recovery action data is the action sequence, action time and recovery topology status of the power supply reconstruction or switch closing and opening recovery operations performed after fault isolation to restore power supply.

[0159] Among them, the continuous relaxation or smooth approximation process is to replace the discrete switch / action state with continuous variables and smooth transition functions so that the topology switching is continuously differentiable in numerical calculation.

[0160] Among them, differentiable topology event data are topology event parameter data that can be used for differentiable simulation after being formed by continuous relaxation or smoothing approximation, characterizing the continuous evolution of the effect of switching and fault isolation / recovery on the network equivalent parameters.

[0161] Among them, the traction load power injection data is the time-varying power absorption and regeneration feedback power sequence data generated at each injection point of the traction power supply network during the train traction and braking process.

[0162] Among them, differentiable uniformity processing is a process of smoothing out non-smooth links such as limiting, saturation and feedback switching in power injection under the action of equipment control capability and boundary constraints, so as to make them consistent with the planning variable constraints and differentiable.

[0163] Among them, the differentiable power injection data is a continuous differentiable power injection sequence data obtained after differentiable uniformization processing, which is used to simultaneously satisfy the control boundary and support sensitivity calculation in the simulation.

[0164] Among them, the power supply status is a collection of electrical and energy states of the traction power supply system during operation, including at least node voltage, branch current, power flow, control status of conversion devices, and state of energy storage (SOC).

[0165] Among them, multi-physics coupled simulation solution is a solution process that couples the electric network equations, converter control, energy storage dynamics and optional thermal models in a unified framework and obtains the system state evolution through numerical integration or implicit solution.

[0166] Sensitivity calculation is the process of calculating the derivative, gradient, or equivalent accompanying sensitivity of the operating index or safety margin relative to the planning variables or key parameters, in order to quantify the impact of variable changes on operating risks and margins.

[0167] Specifically, the switch state switching data, fault isolation data, and recovery action data, which were originally represented as discrete states in the power supply operation scenario data, are rewritten into event descriptions that can be used for continuous calculation. For example, the switch state is replaced by a discrete variable of {0,1} with a continuous relaxed variable with values ​​of [0,1], and the switching process is characterized by a smooth step function or homotopy transition function, so that the branch admittance, tie relationship, or equivalent impedance changes continuously with the event variable. At the same time, the triggering time and duration window of the fault isolation and recovery actions are encoded as differentiable gating functions, so that the influence of the event on the network topology and parameters is continuously differentiable within the simulation step, thereby obtaining differentiable topology event data containing event variables, transition parameters, gating functions, and evolution rules of topology equivalent parameters.

[0168] Based on the train traction / braking power curves and operation diagrams corresponding to flexible DC traction power supply, an original power injection sequence is constructed. Constraints such as the voltage and current limiting capabilities of the power supply device, energy storage power and SOC boundaries, feedback absorption capabilities, and voltage control targets from the two-layer planning variable data are introduced into the power injection model. Non-smooth elements in power injection, such as limiting, saturation, feedback switching, and energy abandonment (e.g., braking resistor activation), are standardized using smooth approximations (e.g., Softplus, smoothed saturation functions, or piecewise differentiable substitutions), ensuring that traction absorption power and regenerative feedback power form a continuously differentiable injection function under constraints. This yields differentiable power injection data that simultaneously satisfies the control capabilities and boundary constraints defined by the planning variables.

[0169] A unified multi-physics state vector and state equations are constructed for flexible DC traction power supply. This couples the node voltage-branch current equations of the traction power supply network, the current / voltage control and current / voltage limiting logic of the flexible converter, the charging and discharging power constraints and SOC dynamics of energy storage (including temperature / internal resistance variations with SOC when necessary), and the thermal-electric coupling relationships between conductors and key components (current-induced temperature rise, temperature-affected resistance and current-carrying capacity) within the same simulation framework. At each time step, the network equivalent parameters are updated based on differentiable topology event data. Train power injection and feedback are applied based on differentiable power injection data. An event-consistent stepping strategy and numerical integration are used to solve the state equations, outputting the operational trajectory data of node voltage, branch current, power flow, SOC, and temperature evolving over time, forming a complete operational state sequence.

[0170] Based on the differentiable computational graph or implicit equation system corresponding to the trajectory generation process, voltage margin, thermal margin, loss, over-limit duration, or comprehensive risk score are selected as output indicators. Automatic differentiation or adjoint sensitivity methods are used to calculate the gradient or equivalent sensitivity of these output indicators relative to the bi-level programming variable data (such as converter capacity and current limiting parameters, energy storage power and energy parameters, switch topology parameters, line parameters, and protection setting parameters). For power flow / dynamic equations with implicit solutions, stable sensitivity is obtained using implicit function theorems or the solver's built-in adjoint module. When necessary, smooth substitution or subgradient approximation is applied to discontinuities to obtain sensitivity information.

[0171] In this embodiment, the beneficial effects of adopting the above-mentioned differentiable coupled simulation process are as follows: By continuously relaxing or smoothing discrete topological events such as switch switching, fault isolation and recovery actions, the topological changes are made continuous and differentiable in the simulation calculation, thereby avoiding the sensitivity distortion caused by the non-differentiability at the switching point in traditional event-driven simulation; further, by differentiating and unifying the traction load power injection under the constraints of bi-level planning variables, non-smooth links such as current limiting and voltage limiting, feedback switching and saturation can be transformed into differentiable expressions, so that the power injection is consistent with the equipment control capability and can be used for gradient calculation; on this basis, multi-physics coupled simulation is carried out based on differentiable topological events and differentiable power injection, which can simultaneously obtain the full operating trajectory of electrical state, control state and energy storage dynamics, and improve the accuracy of characterizing complex coupled conditions; finally, by solving the sensitivity, the gradient or equivalent sensitivity of the operating index to the planning variables can be output, which can provide a direct and usable quantitative guide for subsequent constraint enhancement, adversarial scenario construction and bi-level planning iteration, thereby significantly improving the convergence efficiency of planning optimization and the pertinence of safety margin improvement.

[0172] In an exemplary embodiment, based on differentiable topological event data and differentiable power injection data, a multi-physics coupling simulation is performed to solve the power supply state of the flexible DC traction power supply, obtaining the running trajectory data, including steps 902 to 910. Wherein:

[0173] Step 902: Perform joint state expansion on the differentiable topology event data and the differentiable power injection data to obtain expanded state initialization data.

[0174] Step 904: Perform differentiable substitution processing on the control saturation data in the extended state initialization data to obtain differentiable control constraint data.

[0175] Step 906: Based on the differentiable controllable constraint data, perform homomorphic coupling modeling on the thermal-electric coupling relationship of flexible DC traction power supply to obtain multi-physics coupled state equation data.

[0176] Step 908: Based on the differentiable topological event data, perform homotopy gating on the simulation step of the multi-physics coupled simulation to obtain event-consistent step strategy data.

[0177] Step 910: Perform implicit trapezoidal integration on the multi-physics coupled state equation data and the event-consistent stepping strategy data to obtain the received trajectory data.

[0178] Among them, joint state extension is a process of concatenating topological event variables, power injection variables, electrical states, control states, energy storage states, and thermal states into a unified state vector and establishing a consistent state space representation.

[0179] Among them, the extended state initialization data is a set of data formed by the initial values, boundaries and consistency conditions given to each state component at the start of the simulation after the joint state extension.

[0180] Among them, the control saturation data are descriptive data such as saturation threshold, trigger state and saturation amount generated by mechanisms such as current limiting, voltage limiting, power limiting and anti-integral saturation in the control of flexible conversion devices and energy storage.

[0181] Among them, differentiable substitution is a process of replacing the original non-smooth saturation / switching logic with a smooth saturation function, a soft threshold function, or complementary relaxation, etc., to maintain differentiability.

[0182] Among them, the differentiable control constraint data is the parameterized constraint data of the differentiable control boundary and control logic obtained after differentiable substitution processing, which is used to impose control constraints in simulation and optimization.

[0183] Among them, the thermo-electric coupling relationship is the data on the interaction between current and the temperature rise caused by losses, and the temperature in turn affects electrical characteristics such as resistance, current carrying capacity or device losses.

[0184] Homomorphic coupling modeling is a modeling process that expresses thermal subsystems and electronic systems as a unified dynamic model using consistent state variables and coupling terms within the same state space and time-progressing framework.

[0185] Among them, the multi-physics coupled state equation data is a unified set of state equations and its parameter data formed by homomorphically coupling the electric network equations, control equations, energy storage dynamic equations and thermal equations.

[0186] Among them, homotopy gating consistency processing is a process of introducing homotopy transition parameters within the event window of topology switching and fault isolation / recovery, and combining gating rules to perform consistent control of step size and parameter updates.

[0187] Among them, the event-consistent stepping strategy data is a set of strategy data consisting of step size adaptive rules, event triggering conditions, homotopy path parameters and gating update rules output by homotopy gating consistency processing.

[0188] Implicit trapezoidal integral is a method for solving the stable numerical integral of a rigidly coupled system by implicitly discretizing the state equation using the trapezoidal formula and iteratively solving for the state at the next time step.

[0189] Specifically, at the start of the simulation, the node voltage and branch current states of the traction power supply network, the internal control states of the flexible converter (such as the current loop / voltage loop integral states, current-limiting and voltage-limiting state variables), the SOC and power states of the energy storage device, and the event state variables (such as switching relaxation variables, homotopy transition parameters) used to characterize the continuous evolution of topology events, as well as the injection states related to power injection (such as regenerative gating variables, smoothing and limiting variables) are uniformly concatenated into an extended state vector. Based on the differentiable topology event data and differentiable power injection data, the initial values, boundaries, and consistency conditions of each state component (such as initial topology equivalent parameters, initial injected power, and initial SOC values) are given, thereby generating extended state initialization data.

[0190] Identify non-smooth elements related to the control of the flexible converter in the extended state initialization data (such as current / voltage limiting triggering, power limiting, controller anti-integral saturation, energy storage charging and discharging power limiting, etc.), and replace the original control saturation data, i.e. hard saturation, segmented switching, or complementary constraints, with smooth saturation functions, soft threshold functions, or complementary relaxations (such as introducing smooth penalty functions or continuous relaxation variables), so that the control constraints are numerically continuous and differentiable; at the same time, write the replaced constraints into the control model in the form of inequality constraints, equivalent dynamic equation terms, or gating coefficients, thereby obtaining differentiable control constraint data.

[0191] Based on the node voltage-branch current equations, converter control equations, and energy storage SOC dynamic equations of the electrical subsystem, differentiable controllable constraint data is used as constraints. The thermal subsystem state (such as conductor temperature, device junction temperature, or equivalent temperature rise) is introduced, and a current-induced heating and heat dissipation balance equation is established. Simultaneously, the influence of temperature on resistance, current-carrying capacity, current-limiting threshold, or device loss model is fed back into the electrical and control equations. "Homomorphic coupling" is manifested in expressing the thermal and electrical models in the same state space and time scale using consistent state variables and coupling terms. This allows the thermal-electric coupling to be embedded into a unified dynamic system in the form of differentiable state equations, thus forming multi-physics coupled state equation data.

[0192] Information such as the event timing, event window, and event intensity of switchover, fault isolation, and recovery actions is read from differentiable topology event data. An event gating function is constructed to adjust the step size and update the topology equivalent parameters within the event window. A homotopy transition parameter is introduced to transform topology switching from an "instantaneous jump" to a homotopy path that "evolves continuously within a finite window." Based on this, step consistency rules are formulated. For example, smaller or adaptive step sizes are used within the event window to control local truncation errors, while larger step sizes are restored outside the window to improve efficiency. Homotopy parameters and gating coefficients are synchronously advanced in each update step to ensure consistent updates of topology equivalent parameters, control saturation substitution terms, and power injection gating. This results in event-consistent stepping strategy data that includes event triggering conditions, adaptive step size rules, homotopy path parameters, and gating update rules.

[0193] Using the extended state initialization data as initial values, the current step size and gating / homotopy parameter updates are determined according to an event-consistent stepping strategy at each simulation step. Then, the multi-physics coupled state equations are discretized using implicit trapezoidal integrals. The next-time state is used as an unknown to establish a system of nonlinear equations, which are then solved using Newton iteration or an equivalent nonlinear solver under differentiable control constraints and topological homotopy transition constraints to obtain the next-time state. The above integration and solution process is repeated until the entire simulation time domain is covered, outputting the state sequence of node voltage, branch current, power flow, control state, SOC, and temperature as they evolve over time, thus forming the running trajectory data.

[0194] In this embodiment, by jointly extending the state of differentiable topological events and differentiable power injection to form a unified extended state initialization, the network, power electronic control, energy storage dynamics, and event evolution can be consistently characterized within the same state space, reducing coupling errors caused by split modeling. Furthermore, differentiable substitution is applied to the control saturation stage, enabling non-smooth control logic such as current limiting, voltage limiting, power limiting, and anti-saturation to possess continuous differentiability while maintaining engineering meaning, thereby improving the stability of numerical solutions and laying the foundation for subsequent sensitivity calculations. Finally, homomorphic modeling of thermo-electric coupling is used to correlate current-induced temperature rise with temperature reversal. Incorporating effects such as feed resistance / current carrying capacity into the unified state equation can more realistically reflect the impact of thermal constraints on power supply capacity and risks, enhancing the physical reliability of the operating trajectory. At the same time, adopting a homotopy-gated consistent stepping strategy to achieve continuous transition and step size adaptation within event windows such as topology switching and fault isolation / recovery can significantly reduce numerical oscillations and error accumulation at event points. Finally, combining implicit trapezoidal integrals to perform robust integral solutions to the coupled state equations can obtain stable and accurate operating trajectory data in rigid systems and strongly coupled scenarios, thereby improving the reliability of safety verification, risk assessment, and planning optimization.

[0195] Based on the same inventive concept, this application also provides a flexible DC traction power supply dual-layer planning and optimization device that integrates safety verification and operation simulation for implementing the aforementioned flexible DC traction power supply dual-layer planning and optimization method. For example... Figure 3 As shown, a flexible DC traction power supply dual-layer planning optimization device integrating safety verification and operation simulation is provided, including: a data adjustment module, a power supply simulation module, a data verification module, an adversarial analysis module, and a planning solution module. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the flexible DC traction power supply dual-layer planning optimization device integrating safety verification and operation simulation provided below can be found in the limitations of the flexible DC traction power supply dual-layer planning optimization method integrating safety verification and operation simulation above, and will not be repeated here.

[0196] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. This computer device includes a processor, memory, input / output interfaces (I / O), and communication interfaces.

[0197] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0198] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0199] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.

[0200] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0201] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0202] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A two-layer planning and optimization method for flexible DC traction power supply that integrates safety verification and operation simulation, characterized in that, The method includes: Spatiotemporal unified adjustment of multi-source planning data and basic operational data of flexible DC traction power supply is performed to obtain dual-layer planning variable data and power supply operation scenario data; Based on the bi-level planning variable data and the power supply operation scenario data, a differentiable coupling simulation is performed on the flexible DC traction power supply to obtain the operation trajectory data and sensitivity information. The operational trajectory data is subjected to formal security verification to obtain constraint enhancement information and risk indicator data; Based on the risk indicator data, the flexible DC traction power supply is subjected to a counter-scenario analysis to obtain the worst-case counter-scenario data; The constraint enhancement information is subjected to feasible region shrinkage processing to obtain safe feasible region data; Based on the sensitivity information, the planning variables within the safe and feasible domain data are pre-updated to obtain pre-updated planning variable data; The fault data, regeneration spike data, timetable disturbance data and parameter uncertainty data in the worst-case scenario data are solidified to obtain the worst-case equivalent constraint data. The voltage margin data, thermal margin data, and protection action timing margin data corresponding to the worst-case equivalent constraint data are reconstructed in a target-oriented manner to obtain the safety margin objective function data. By combining the pre-updated planning variable data and the safety margin objective function data, a differentiable bi-level equivalent transformation is performed to obtain bi-level equivalent optimization problem data; Based on the violation boundary data or barrier certificate boundary data in the constraint enhancement information, the two-layer equivalent optimization problem data is updated using evidence-driven backtracking to obtain corrected initial value data. Using the corrected initial value data as the initial value, sequential quadratic programming is used to iteratively solve the two-level equivalent optimization problem data to obtain the power supply safety output planning data.

2. The method according to claim 1, characterized in that, The step involves analyzing the worst-case scenario of the flexible DC traction power supply based on the risk indicator data, resulting in worst-case scenario data, including: The risk indicator data, including the out-of-limit amplitude data, out-of-limit duration data, and protection violation severity data, are inverted and located to obtain the adversarial seed parameter data. Based on the adversarial seed parameter data and scenario variable constraint data, an adversarial search is performed on the power supply operation scenario data to obtain the initial adversarial scenario data; The worst-case adversarial scenario data is obtained by perturbing the initial adversarial scenario data through Latin hypercube sampling.

3. The method according to claim 2, characterized in that, The step of performing adversarial search on the power supply operation scenario data based on the adversarial seed parameter data and scenario variable constraint data to obtain initial adversarial scenario data includes: Projecting the boundary constraint data, coupling constraint data, and discrete value constraint data in the scenario variable constraint data yields feasible adversarial variable domain data. Based on the adversarial seed parameter data, the power supply variables of each scenario in the power supply operation scenario data are combined to obtain the candidate adversarial scenario set data; Risk proxy screening is performed on the candidate adversarial scenario set data to obtain a simplified adversarial scenario set data; The initial adversarial scenario data is obtained by performing a local search based on the simplified set of adversarial scenarios.

4. The method according to claim 3, characterized in that, The step of performing risk proxy screening on the candidate adversarial scenario set data to obtain a simplified adversarial scenario set data includes: Sensitivity injection feature processing is performed on the candidate adversarial scenario set data to obtain risk agent feature data; Based on the violation constraint type data, violation boundary data, or time-series violation data in the constraint enhancement information, the risk agency feature data is verified and calibrated to obtain calibrated risk agency feature data. Uncertainty definition analysis is performed on the calibration risk proxy characteristic data to obtain the lower confidence boundary data of proxy risk; The confidence boundary data under the aforementioned proxy risk is sorted and truncated to obtain a simplified set of adversarial scenario data.

5. The method according to claim 1, characterized in that, The process involves performing a differentiable coupled simulation of the flexible DC traction power supply based on the bi-level programming variable data and the power supply operation scenario data to obtain operation trajectory data and sensitivity information, including: The switch state switching data, fault isolation data, and recovery action data in the power supply operation scenario data are continuously relaxed or smoothed to obtain differentiable topology event data. Based on the bi-level planning variable data, the traction load power injection data corresponding to the flexible DC traction power supply is processed to be differentiable and consistent, so as to obtain differentiable power injection data. Based on the differentiable topology event data and the differentiable power injection data, the power supply status of the flexible DC traction power supply is solved by multi-physics coupling simulation to obtain the running trajectory data; The sensitivity information is obtained by performing sensitivity calculation on the trajectory data.

6. The method according to claim 5, characterized in that, The process involves performing multi-physics coupling simulation to determine the power supply status of the flexible DC traction power supply based on the differentiable topology event data and the differentiable power injection data, thereby obtaining the operating trajectory data, including: Joint state expansion is performed on the differentiable topology event data and the differentiable power injection data to obtain expanded state initialization data; Differentiable substitution processing is performed on the control saturation data in the extended state initialization data to obtain differentiable control constraint data; Based on the differentiable controllable constraint data, the thermal-electric coupling relationship of the flexible DC traction power supply is modeled in a homomorphic coupling manner to obtain multi-physics coupling state equation data. Based on the differentiable topological event data, the simulation step of the multi-physics coupled simulation is homotopically gated to achieve consistency, thereby obtaining event-consistent step strategy data. The running trajectory data is obtained by performing implicit trapezoidal integration on the multi-physics coupled state equation data and the event-consistent step strategy data.

7. A flexible DC traction power supply dual-layer planning and optimization device integrating safety verification and operation simulation, characterized in that, The device includes: The data adjustment module is used to perform spatiotemporal unified adjustment of the multi-source planning data and operational basic data of flexible DC traction power supply to obtain dual-level planning variable data and power supply operation scenario data. The power supply simulation module is used to perform differentiable coupling simulation of the flexible DC traction power supply based on the dual-level planning variable data and the power supply operation scenario data, so as to obtain the operation trajectory data and sensitivity information. The data verification module is used to perform formal security verification on the running trajectory data to obtain constraint enhancement information and risk indicator data; The adversarial analysis module is used to perform adversarial scenario analysis on the flexible DC traction power supply based on the risk index data, and obtain the worst-case adversarial scenario data. The planning and solving module is used to perform feasible region shrinkage processing on the constraint enhancement information to obtain safe feasible region data; Based on the sensitivity information, the planning variables within the safe and feasible domain data are pre-updated to obtain pre-updated planning variable data; The fault data, regeneration spike data, timetable disturbance data and parameter uncertainty data in the worst-case scenario data are solidified to obtain the worst-case equivalent constraint data. The voltage margin data, thermal margin data, and protection action timing margin data corresponding to the worst-case equivalent constraint data are reconstructed in a target-oriented manner to obtain the safety margin objective function data. By combining the pre-updated planning variable data and the safety margin objective function data, a differentiable bi-level equivalent transformation is performed to obtain bi-level equivalent optimization problem data; Based on the violation boundary data or barrier certificate boundary data in the constraint enhancement information, the two-layer equivalent optimization problem data is updated using evidence-driven backtracking to obtain corrected initial value data. Using the corrected initial value data as the initial value, sequential quadratic programming is used to iteratively solve the two-level equivalent optimization problem data to obtain the power supply safety output planning data.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

Images