A power distribution network planning method based on decision-oriented learning

Abstract

The present application relates to the technical field of power distribution network control, and specifically includes a power distribution network planning method based on decision-oriented learning, which comprises: obtaining a power distribution network planning dataset; training a scene generation model, the inner layer optimization connection layer performs feature analysis and constraint embedding on the power distribution network planning dataset, and the outer layer optimization connection layer takes the weak boundary region as the priority orientation domain; and outputting a power distribution network planning scheme. The technical problem of being unable to fully respond to the uncertainty of new energy output fluctuation and load space-time change and being difficult to adapt to the complex demand of new power systems is solved, the technical effect of deep coupling of data and electrical constraints through feature analysis and constraint embedding is achieved, the weak boundary region is taken as the priority orientation domain to strengthen the scene coverage of high-risk areas, the response capability to uncertainty such as new energy fluctuation and load change is improved, the planning strategy is dynamically optimized with the goal of minimizing the expected decision loss, and the reliability of the power distribution network planning scheme is improved.

Description

Technical Field

[0001] This invention relates to the field of distribution network control technology, specifically to a distribution network planning method based on decision-oriented learning. Background Technology

[0002] Distribution network planning must adapt to the temporal and spatial uncertainties of load growth and the volatility of renewable energy output, ensuring the safe operation of the power grid under extreme scenarios such as N-1 faults. At the same time, it must meet electrical constraints such as voltage quality and line current carrying capacity, avoiding resource waste or power supply gaps caused by planning deviations, and ensuring the long-term stable operation of the power grid.

[0003] Current distribution network planning methods are mostly based on static load forecast data or limited typical scenarios, without fully considering uncertainties such as fluctuations in renewable energy output and changes in the spatiotemporal distribution of load. In areas with weak grid boundaries, such as high load density areas and concentrated renewable energy access points, electrical constraints are easily breached, leading to power supply risks. In addition, the decision optimization and loss control mechanisms are imperfect. When forecast deviations trigger loss thresholds, planning strategies cannot be quickly adjusted to suppress losses. Distribution network planning schemes are difficult to adapt to the complex needs of new power systems, further restricting the reliability of distribution network planning.

[0004] In summary, existing technologies suffer from technical problems due to their reliance on static load forecasting data or limited typical scenarios. These technologies are unable to adequately address the uncertainties of new energy output fluctuations and load changes in time and space, and are therefore ill-suited to the complex needs of new power systems. Summary of the Invention

[0005] This application provides a distribution network planning method based on decision-oriented learning, aiming to solve the technical problem that existing technologies rely solely on static load forecast data or limited typical scenarios, which cannot fully cope with the uncertainties of new energy output fluctuations and load spatiotemporal changes, and are difficult to adapt to the complex needs of new power systems.

[0006] In view of the above problems, the technical solution to achieve the present application is as follows:

[0007] This application provides a distribution network planning method based on decision-oriented learning. The method includes: acquiring a distribution network planning dataset of a target distribution network; training a scenario generation model under a two-layer optimization architecture based on mapping relationships, with the goal of minimizing the expected decision loss of the decision loss function; wherein the two-layer optimization architecture includes an inner optimization connection layer and an outer optimization connection layer, the inner optimization connection layer performing feature parsing and constraint embedding on the distribution network planning dataset, and the outer optimization connection layer prioritizing weak boundary regions; simultaneously, after each iteration, the expected decision loss over the entire lifecycle is re-evaluated to determine whether convergence has occurred: if the decrease in expected decision loss during the iteration process is less than a set tolerance, training is stopped, and a distribution network planning scheme under decision-oriented learning is output.

[0008] Preferably, the outer optimization connection layer is based on the mapping relationship and decision loss chain, with the weak boundary region as the priority guiding domain; at the same time, the constraint contraction coefficient of the inner optimization connection layer is optimized to coordinate the adaptation of the feasible operating domain and the feasibility boundary, reduce the decision loss sensitivity corresponding to the weak boundary region, and iteratively execute the dynamic adjustment process based on the TVS diode protection mechanism.

[0009] Preferably, the distribution characteristics of the virtual prediction error and the scene generation weights are dynamically adjusted by differentiably optimizing the backpropagation decision loss gradient, thereby enhancing the scene coverage density of the weak boundary area; and the mapping relationship is determined according to the power grid topology model of the target distribution network.

[0010] Preferably, using the power grid topology model of the target distribution network, multiple error combinations are configured with virtual prediction errors under positive and negative deviations, and a distribution network planning strategy is executed for each of the multiple error combinations to obtain the mapping relationship from the input load prediction error to the planning increment.

[0011] Preferably, the planning increment is determined by tracing the decision loss chain throughout the entire life cycle; the planning increment is used to characterize the additional cumulative difference between the distribution network planning strategy and the globally optimal decision in the prediction deviation scenario.

[0012] Preferably, the feasible operating domain is defined by the first electrical constraint condition and the second electrical constraint condition; the feasible operating domain and the feasibility boundary are constrained and contracted using a coupling constraint matrix, and weak boundary regions are marked.

[0013] Preferably, the power distribution network planning dataset includes energy application data and technology development data, power supply configuration data and load space planning data, power grid basic data, and historical time series data; the first electrical constraint condition is set using the power grid basic data; and the second electrical constraint condition is set using the historical time series data.

[0014] Preferably, the feasibility boundary of the power distribution network planning strategy is set based on the energy application data and technology development data; and the coupling constraint matrix is ​​constructed based on the power supply configuration data and load space planning data.

[0015] Preferably, when the virtual prediction error triggers the loss threshold of the weak boundary region, the decision loss is suppressed to a safe range by step adjustment of the constraint contraction coefficient, and then the steady-state adjustment stage is entered. The virtual prediction error distribution and constraint parameters are finely adjusted in the gradient descent direction to gradually approach the optimal solution and determine the distribution network planning scheme.

[0016] Preferably, using the typical power distribution scenario set output by the scenario generation model as the initial search starting point, multiple state transition paths are constructed along the load growth trajectory, the renewable energy output fluctuation path, and the N-1 fault transfer section. Small disturbances are introduced into each state transition path to simulate the dynamic drift process at the edge of the feasible operating domain, monitoring the gradient of decision loss changes in the corresponding power distribution network planning scheme. In the first state transition path, if the decision loss growth rate of the current state node exceeds a preset loss sensitivity threshold, the current state node and M associated state nodes are combined to form a locally highly sensitive section. Clustering expansion is then performed based on the geometric continuity and gradient consistency corresponding to the decision loss gradient distribution of the first state transition path to determine the risk manifold. The first state transition path is any one of the multiple state transition paths. Based on the geometric features of the risk manifold, including curvature and expansion direction, the sampling density function of the scenario generation model is adjusted inversely.

[0017] In summary, one or more technical solutions provided in this application achieve deep coupling between data and electrical constraints through feature parsing and constraint embedding, strengthen the coverage of high-risk area scenarios by prioritizing weak boundary areas, enhance the ability to cope with uncertainties such as new energy fluctuations and load changes, minimize expected decision losses, combine full life cycle loss assessment and iterative convergence judgment, dynamically optimize planning strategies, avoid long-term resource waste and power supply risks, and improve the reliability of distribution network planning schemes. Attached Figure Description

[0018] Figure 1 This application provides a flowchart illustrating a distribution network planning method based on decision-oriented learning. Detailed Implementation

[0019] The embodiments are described in detail below with reference to the accompanying drawings, such as... Figure 1 As shown, this application provides a distribution network planning method based on decision-oriented learning, wherein the method includes:

[0020] S1: Obtain the distribution network planning dataset of the target distribution network; S2: Based on the mapping relationship, train the scene generation model under the two-layer optimization architecture with the goal of minimizing the expected decision loss of the decision loss function; S3: Wherein, the two-layer optimization architecture includes an inner optimization connection layer and an outer optimization connection layer. The inner optimization connection layer performs feature parsing and constraint embedding on the distribution network planning dataset, and the outer optimization connection layer takes the weak boundary region as the priority guiding domain.

[0021] Specifically, the distribution network planning dataset is a comprehensive collection of data containing information about the target distribution network, covering aspects such as energy application and technology development, power supply configuration and load space planning, basic power grid data, and historical time-series data. It forms the foundation for distribution network planning. By analyzing the distribution network planning dataset, we can understand the current status of the distribution network, load characteristics, power supply distribution, and other key information, providing data support for subsequent planning decisions. In distribution network planning, the mapping relationship refers to the correspondence between the input load forecasting error and the planning increment. Through the mapping relationship, uncertainties such as load forecasting errors are transformed into specific planning strategy adjustments, thereby achieving dynamic optimization of the distribution network planning scheme. The mapping relationship is obtained through model training and can reflect the optimal planning response under different input conditions.

[0022] The decision loss function is a quantitative indicator for evaluating the quality of distribution network planning schemes. It comprehensively considers various costs and risks throughout the entire life cycle, including operation and maintenance costs, power outage losses, and the cost of abandoning renewable energy. The goal of minimizing the expected decision loss of the decision loss function is to find the statistically optimal distribution network planning scheme. The two-layer optimization architecture is a hierarchical optimization framework, including an inner optimization connection layer and an outer optimization connection layer. The inner optimization connection layer is mainly responsible for feature parsing and constraint embedding of the distribution network planning dataset. By extracting key features from the data and combining them with electrical constraints, it provides data support and constraint conditions for the scenario generation model. The outer optimization connection layer prioritizes weak boundary areas and strengthens the scenario coverage of high-risk areas by optimizing the weights and distribution characteristics of the scenario generation model, thereby improving the distribution network planning scheme's ability to cope with uncertainties.

[0023] Execution Steps: It is necessary to collect and organize various data from the target distribution network, including energy application and technology development data, power supply configuration and load spatial planning data, grid basic data, and historical time-series data. The distribution network planning dataset for the target distribution network is obtained through various methods such as historical record queries and sensor data collection. Specifically, real-time data collected through smart meters and distributed energy management systems provides accurate information for load forecasting and power supply configuration. After obtaining the distribution network planning dataset, this data is used to train a scenario generation model. During training, the scenario generation model transforms input uncertainties such as load forecasting errors into specific planning increments based on mapping relationships. Simultaneously, the goal of training the scenario generation model is to minimize the expected value of the decision loss function, meaning that the scenario generation model needs to comprehensively consider various costs and risks throughout its entire lifecycle to find the optimal distribution network planning scheme. Specifically, in a distribution network that includes distributed photovoltaic power generation and electric vehicle charging loads, the scenario generation model needs to consider the intermittency of photovoltaic power generation and the spatiotemporal distribution changes of electric vehicle charging loads, minimizing long-term decision losses by optimizing the distribution network planning scheme.

[0024] Preferably, the dual-layer optimization architecture achieves deep coupling between data and electrical constraints and enhanced coverage of high-risk areas through the collaborative work of the inner and outer optimization connection layers. The inner optimization connection layer performs feature analysis and constraint embedding on the distribution network planning dataset to ensure that the actual operational constraints of the power grid, such as voltage quality and line current carrying capacity, are fully considered during model training. The outer optimization connection layer prioritizes weak boundary areas, improving the scenario coverage density of these high-risk areas by optimizing the weights and distribution characteristics of the scenario generation model. Furthermore, in weak boundary areas such as high load density areas or concentrated renewable energy access points, the scenario generation model generates more scenarios to simulate potential extreme situations, thereby improving the distribution network planning scheme's ability to cope with uncertainties. Through these steps, the uncertainties of renewable energy output fluctuations and load spatiotemporal variations can be effectively addressed, enhancing the reliability of the distribution network planning scheme.

[0025] S4: At the same time, after each iteration, the expected decision loss over the entire life cycle is re-evaluated to determine whether convergence has occurred. If the decrease in the expected decision loss during the iteration process is less than the set tolerance, then training is stopped and a decision-oriented distribution network planning scheme is output.

[0026] Specifically, the expected decision loss over the entire lifecycle refers to the expected value after comprehensively considering various costs and risks throughout the entire lifecycle of the distribution network planning scheme, including initial investment costs, operation and maintenance costs, power outage losses, and renewable energy abandonment costs. By evaluating the expected decision loss over the entire lifecycle, the long-term reliability of the distribution network planning scheme can be more comprehensively measured. In the iterative optimization process, convergence judgment refers to determining whether the model has approached the optimal solution by setting a tolerance value. If, after a certain iteration, the decrease in expected decision loss is less than this set tolerance value, the model can be considered to have converged, that is, the optimal distribution network planning scheme in the statistical process has been found. Convergence judgment is an important part of the optimization process, determining when to stop model training to avoid overcomputation and waste of resources.

[0027] Execution steps: During each iteration, the scenario generation model recalculates the expected decision loss over the entire lifecycle based on the current parameters and the generated scenario. As the model parameters are updated and the scenario is optimized, the expected decision loss gradually decreases. Specifically, if the initial expected decision loss is 100 units, after the first iteration, the expected decision loss decreases to 80 units, after the second iteration it decreases to 65 units, and so on. In this way, the model can continuously optimize and find a better distribution network planning scheme. After each iteration, the decrease in expected decision loss is calculated and compared with the set tolerance value. If the decrease is less than the tolerance value, for example, if the tolerance value is set to 1 unit and the decrease after a certain iteration is only 0.5 units, then the model can be considered to have converged. At this time, the model stops training and outputs the current decision-oriented distribution network planning scheme.

[0028] In the above steps, by dynamically evaluating the expected decision loss over the entire life cycle and stopping training based on convergence judgment, the optimal distribution network planning scheme in the statistical process is found, which improves the reliability of the distribution network planning scheme, avoids over-computation and resource waste, and preferably satisfies electrical constraints and minimizes long-term decision loss within limited computing resources in distribution networks containing multiple distributed power sources and complex loads. This enables the distribution network planning scheme to better adapt to the complex needs of new power systems and improve the long-term stable operation capability of the distribution network.

[0029] Furthermore, the method used in this application for training a scene generation model with a two-layer optimization architecture includes:

[0030] The outer optimization connection layer is based on the mapping relationship and decision loss chain, and takes the weak boundary region as the priority guidance domain. At the same time, it optimizes the constraint contraction coefficient of the inner optimization connection layer, performs collaborative adaptation between the feasible operating domain and the feasibility boundary, reduces the decision loss sensitivity corresponding to the weak boundary region, and iteratively executes the dynamic adjustment process based on the TVS diode protection mechanism.

[0031] Specifically, the decision loss chain refers to a series of causal relationships from input uncertainties to the final decision loss. Common uncertainties include load forecasting errors and fluctuations in new energy output. By tracing the decision loss chain, the contribution of each link to the final decision loss can be clarified, thus providing direction for optimization. The constraint contraction coefficient is a parameter used in the inner optimization connection layer to adjust the strictness of constraints. By optimizing the constraint contraction coefficient, the shape of the feasible operating domain and feasibility boundary can be dynamically adjusted to better fit the current optimization objective. A TVS (Transient Voltage Suppressor) diode is a component used for circuit protection. It can withstand high-energy pulse voltages for a short time, thereby protecting the circuit from damage caused by transient voltages. Specifically, the TVS diode protection mechanism adapts to the dynamic adjustment process. When the decision loss triggers the threshold, parameters such as the constraint contraction coefficient are quickly adjusted to suppress the decision loss to a safe range.

[0032] Execution steps: The outer optimization connection layer analyzes the mapping relationship and decision loss chain to identify weak boundary areas as priority optimization areas, because weak boundary areas are usually high-risk areas in the power grid, such as high load density areas or concentrated access points for new energy sources. By strengthening the scenario coverage of weak boundary areas, the ability of the distribution network planning scheme to cope with uncertainties is improved. Specifically, in a certain distribution network where the high penetration rate of distributed photovoltaic power generation makes it a weak boundary area, the outer optimization connection layer will increase the scenario generation weight of this area to ensure that the scenario generation model can fully consider the extreme situations that may occur in this area.

[0033] The inner optimization connection layer dynamically adjusts the shape of the feasible operating domain and the feasible boundary by adjusting the constraint contraction coefficient. This process is collaborative and adaptive, meaning that while adjusting the constraint contraction coefficient, changes in other relevant parameters are also considered to ensure the overall optimization effect. Furthermore, by adjusting the constraint contraction coefficient, the boundary of the feasible operating domain is contracted in a safer direction, thereby reducing the sensitivity of decision loss corresponding to weak boundary areas and improving the reliability of the distribution network planning scheme.

[0034] When the virtual prediction error triggers the loss threshold in the weak boundary region, parameters such as the constraint contraction coefficient are adjusted to suppress the decision loss to a safe range. Furthermore, TVS diodes in the distribution network are located. Through the role of TVS diodes in circuit protection, excessive transient currents are quickly diverted, and the voltage level in the weak boundary region is stabilized. This allows for a rapid response in a short time, preventing the decision loss from further expanding. After entering the steady-state adjustment stage, the distribution of the virtual prediction error and constraint parameters are finely adjusted in the gradient descent direction to gradually approach the optimal solution, and finally, the optimal distribution network planning scheme is determined.

[0035] In the above steps, the optimization of weak boundary areas is strengthened through the coordinated work of the outer optimization connection layer and the inner optimization connection layer, which improves the ability of the distribution network planning scheme to cope with uncertainties. At the same time, through the dynamic adjustment process based on the TVS diode protection mechanism, it can respond quickly when the decision loss trigger threshold is reached, prevent the loss from expanding further, and ensure the reliability and stability of the distribution network planning scheme.

[0036] Furthermore, the outer optimization connection layer is based on mapping relationships and decision loss chains, and the method of this application also includes:

[0037] By differentiably optimizing the backpropagation decision loss gradient, the distribution characteristics of the virtual prediction error and the scene generation weights are dynamically adjusted to enhance the scene coverage density of the weak boundary area; and the mapping relationship is determined according to the power grid topology model of the target distribution network.

[0038] Specifically, differentiable optimization allows model parameters to be adjusted by calculating the gradient of the loss function. In distribution network planning, differentiable optimization enables the model to dynamically adjust parameters through backpropagation to minimize the decision loss function. Backpropagation of the decision loss gradient is a step in differentiable optimization. By calculating the gradient of the decision loss function relative to the model parameters, it determines how to adjust the parameters to reduce losses. In distribution network planning, this means that the distribution characteristics of the virtual prediction error and the scenario generation weights can be dynamically adjusted according to the current decision loss situation.

[0039] In distribution network planning, virtual forecasting errors include load forecasting errors and renewable energy output forecasting errors. By adjusting the distribution characteristics of virtual forecasting errors, uncertainties in actual operation can be better simulated. Scenario generation weights refer to the weights assigned to each scenario when generating different operating scenarios. By adjusting these weights, the scenario generation model can control the importance of different scenarios in the optimization process, thereby strengthening scenario coverage of weak boundary areas. The power grid topology model is the physical structure model of the distribution network, describing the connection relationships between various nodes and lines in the distribution network. By analyzing the power grid topology model, mapping relationships are determined, that is, the correspondence between input uncertainties and planning increments.

[0040] Execution steps: In each iteration, the gradient of the decision loss function relative to the virtual prediction error and the scene generation weights is calculated using differentiable optimization techniques. Based on these gradients, the distribution characteristics of the virtual prediction error and the scene generation weights are dynamically adjusted. Furthermore, if the decision loss is high in a certain weak boundary area, the scene generation weights in that area are increased, and the distribution of the virtual prediction error is adjusted to be closer to the actual possible error distribution. In this way, weak boundary areas are covered more effectively, and the ability of the distribution network planning scheme to cope with uncertainties is improved.

[0041] When determining the mapping relationship, the power grid topology model is used to analyze the physical structure of the distribution network. By analyzing the power grid topology model, the connection relationship between different nodes and lines is determined, thereby establishing a mapping relationship from input uncertainties to planned increments. Uncertainties include load forecasting errors and new energy output forecasting errors. Furthermore, in distribution networks containing multiple distributed power sources and complex loads, by analyzing the power grid topology model, it is determined how load forecasting errors at different nodes and lines affect the operating status of the entire distribution network, thereby establishing the corresponding mapping relationship.

[0042] In the above steps, by dynamically adjusting the distribution characteristics of virtual prediction errors and the scene generation weights, the scene coverage of weak boundary areas is strengthened, and the ability of the distribution network planning scheme to cope with uncertainties is improved. At the same time, by using the power grid topology model to determine the mapping relationship, the model can more accurately reflect the actual operating characteristics of the distribution network, thereby improving the reliability and adaptability of the distribution network planning scheme.

[0043] Furthermore, based on the power grid topology model of the target distribution network, the mapping relationship is determined. The method of this application includes:

[0044] Using the power grid topology model of the target distribution network, multiple error combinations are configured with virtual prediction errors under positive and negative deviations, and a distribution network planning strategy is executed for each of the multiple error combinations to obtain the mapping relationship from the input load prediction error to the planning increment.

[0045] Specifically, the power grid topology model is the physical structure model of the distribution network, describing the connection relationships between various nodes and lines in the distribution network. By analyzing the power grid topology model, we can understand the physical structure and operating characteristics of the distribution network. In the process of prediction error, positive deviation refers to the situation where the actual value is higher than the predicted value, and negative deviation refers to the situation where the actual value is lower than the predicted value. Positive and negative deviations reflect the uncertainty of prediction. For distribution network planning, considering positive and negative deviations can help better cope with the uncertainty in actual operation. Error combination refers to combining different positive and negative deviations to form a variety of possible prediction error scenarios. By analyzing these error combinations, we can gain a more comprehensive understanding of the operating status of the distribution network under different prediction error conditions. Planning increment refers to the additional adjustments or changes made in distribution network planning to cope with prediction errors. Specifically, this includes increasing line capacity and adjusting power supply configuration to ensure the reliable operation of the power grid under different prediction error conditions.

[0046] Execution steps: In distribution network planning, to comprehensively evaluate the operating status of the power grid under different prediction error conditions, it is necessary to consider both positive and negative deviations. Multiple error combinations are configured in the power grid topology model. The power grid topology model can generate operating status under different scenarios. Specifically, if there are two cases of load prediction with positive deviation of 10% and negative deviation of 10%, two error combinations will be generated: one is the case where the load prediction value increases by 10%, and the other is the case where the load prediction value decreases by 10%. In this way, a wider range of operating scenarios can be covered, improving the adaptability and reliability of the distribution network planning scheme.

[0047] After generating multiple error combinations, a distribution network planning strategy is executed for each combination. This includes adjusting power supply configuration, increasing line capacity, and optimizing switch configuration to ensure reliable operation of the power grid under different prediction error conditions. The execution of the distribution network planning strategy yields a mapping relationship from the input load prediction error to the planning increment, providing a foundation for subsequent optimization. In the above steps, by considering multiple error combinations with both positive and negative deviations, the operating status of the distribution network under different prediction error conditions is comprehensively evaluated. The execution of the distribution network planning strategy obtains a mapping relationship from the input load prediction error to the planning increment, which reflects the actual operating characteristics of the distribution network and provides data support for dynamically adjusting the distribution network planning scheme.

[0048] Furthermore, to obtain the mapping relationship from the input load forecasting error to the planning increment, the method of this application includes:

[0049] The planning increment is determined by tracing the decision loss chain throughout the entire life cycle; the planning increment is used to characterize the additional cumulative difference between the distribution network planning strategy and the globally optimal decision in the prediction deviation scenario.

[0050] Specifically, the decision loss chain in the whole life cycle refers to the causal chain of accumulated decision losses at all decision points from the initial stage of distribution network planning to the end of its entire operation cycle. This includes decision losses during the initial planning and additional decision losses caused by various other factors during operation, such as load forecasting errors and fluctuations in renewable energy output. The planning increment refers to the additional adjustments or changes required to make the distribution network planning strategy achieve an effect similar to the global optimal decision under specific forecasting deviation scenarios. It reflects the difference between the planning strategy and the global optimal decision under non-ideal forecasting conditions.

[0051] Execution steps: In distribution network planning, to evaluate the performance of a planning strategy throughout its entire lifecycle, it is necessary to trace its decision loss chain, involving all decision points from the initial planning stage to the end of operation. This involves analyzing the losses at each decision point and determining the cumulative process of those losses. Furthermore, in the initial planning stage, errors in load growth forecasts lead to deviations in initial investment decisions; during operation, fluctuations in renewable energy output result in additional operating costs and power outage losses. By tracing these decision losses, the additional adjustments required to achieve the globally optimal decision-making effect of the current planning strategy, i.e., the planning increment, are determined.

[0052] Planning increment is an indicator that measures the gap between the current planning strategy and the globally optimal decision. It reflects the additional adjustments needed to achieve the effect of the globally optimal decision under specific prediction deviation scenarios. Specifically, in prediction deviation scenarios, if the current planning strategy leads to additional operating costs and power outage losses, planning increment includes measures such as increasing line capacity and optimizing power supply configuration to reduce these additional losses. By calculating planning increment, the performance of the planning strategy can be evaluated more accurately, and direction can be provided for subsequent optimization. In the above steps, the planning increment is determined by tracing the decision loss chain throughout the entire life cycle. The planning increment reflects the gap between the current planning strategy and the globally optimal decision, providing direction for dynamically adjusting the planning strategy. Specifically, in distribution networks containing multiple distributed power sources and complex loads, by tracing the decision loss chain, the planning increment required under different prediction deviation scenarios is determined, thereby optimizing the planning strategy and reducing long-term decision losses.

[0053] Furthermore, using the aforementioned weak boundary region as the priority guiding domain, the method of this application also includes:

[0054] The feasible operating domain is defined by the first electrical constraint and the second electrical constraint; the feasible operating domain and the feasibility boundary are constrained and contracted using the coupling constraint matrix, and weak boundary regions are marked.

[0055] Specifically, the first electrical constraint is a constraint set based on the basic data of the power grid, which usually includes restrictions on basic electrical parameters such as voltage level, line capacity, and short-circuit current. The first electrical constraint is used to ensure that the target distribution network meets basic electrical safety and performance requirements under normal operating conditions. The second electrical constraint is a constraint set based on historical time-series data, which usually involves dynamic factors such as load changes and fluctuations in renewable energy output. The second electrical constraint considers the electrical characteristics of the power grid at different times and operating states, ensuring that the target distribution network can operate safely and stably under various possible operating scenarios. The feasible operating domain refers to the set of all possible states in which the power grid can operate safely and stably under given electrical constraints, defining the operating boundary of the target distribution network under normal and abnormal conditions.

[0056] The coupling constraint matrix is ​​used to represent and handle the interrelationships between multiple constraints. In distribution network planning, the coupling constraint matrix can be used to comprehensively consider the first and second electrical constraints to form a unified constraint system. Constraint contraction refers to adjusting the strictness of constraints to make the boundary of the feasible operating domain more compact, thereby improving the reliability and economy of the distribution network planning scheme. Constraint contraction can dynamically adjust constraints to better adapt to the current optimization objectives. Weak boundary areas are high-risk areas in the power grid, which are prone to operational problems due to high load density, concentrated access of new energy sources, or stringent electrical constraints. Marking weak boundary areas helps to concentrate optimization resources and improve the pertinence and effectiveness of distribution network planning schemes.

[0057] Execution steps: First, based on the power grid's basic data, set the first electrical constraint, stipulating that the voltage of all nodes must be within the allowable range. Commonly, the first electrical constraint can be 95% to 105% of the rated voltage, and the current of all lines must not exceed their rated capacity. The first electrical constraint provides a basic electrical safety boundary for the normal operation of the power grid. Second, based on historical time-series data, set the second electrical constraint. Further, considering load fluctuations and uncertainties in renewable energy output over different time periods, it stipulates that at any given time, the power balance of the power grid must meet the requirements, and voltage and current must not exceed safe limits. Through the first and second electrical constraints, the feasible operating domain of the target distribution network is defined, limiting the range within which the power grid can operate safely and stably under various operating conditions.

[0058] After defining the feasible operating domain, a coupling constraint matrix is ​​used to comprehensively process the first and second electrical constraints. The coupling constraint matrix can dynamically adjust the strictness of the constraints. Through constraint contraction processing, the boundary of the feasible operating domain is made more compact. Furthermore, by increasing the constraint contraction coefficient, the boundary of the feasible operating domain is contracted in a safer direction, thereby improving the reliability of the distribution network planning scheme. During the constraint contraction process, weak boundary areas in the power grid are identified. Weak boundary areas are usually high load density areas, concentrated access points of new energy sources, or places with more stringent electrical constraints. Marking weak boundary areas helps to concentrate resources for optimization and improve the pertinence and effectiveness of the distribution network planning scheme.

[0059] Preferably, by defining feasible operating domains and performing constraint contraction processing, the feasibility and reliability of the distribution network planning scheme under electrical constraints are ensured; by marking weak boundary areas, resources can be concentrated and optimized, and scenario coverage and optimization processing of these high-risk areas can be strengthened, thereby improving the adaptability and flexibility of the distribution network planning scheme and ensuring the stability and security of the power grid under various operating conditions.

[0060] Furthermore, by defining the feasible operating domain through the first electrical constraint and the second electrical constraint, the method of this application also includes:

[0061] The power distribution network planning dataset includes energy application data and technology development data, power supply configuration data and load space planning data, power grid basic data, and historical time series data; the first electrical constraint condition is set using the power grid basic data; and the second electrical constraint condition is set using the historical time series data.

[0062] Specifically, energy application data and technology development data reflect the application of energy and the development trends of related technologies. Specifically, this includes the installed capacity of renewable energy, the application of energy storage technology, and the development of smart grid technology, which helps to predict the impact of future energy use patterns and technological innovations on the power grid. Power supply configuration data and load spatial planning data include the type, location, and capacity of power sources, as well as the geographical distribution of loads. Specifically, this includes the configuration of solar power plants and wind farms connected to the target distribution network, as well as the load distribution in residential, commercial, and industrial areas, which is crucial for the rational planning of the power grid layout.

[0063] Basic power grid data comprises the fundamental technical parameters and configuration information of the power grid, including line length, capacity, resistance, transformer rated power, and voltage level. It forms the basis for setting the first electrical constraints, ensuring that the power grid meets basic electrical safety and performance requirements during normal operation. The first electrical constraints are electrical constraints set based on the basic power grid data. These are typically static constraints based on design standards. They are calculated using AC power flow equations to determine electrical parameters such as node voltage, line current, and short-circuit capacity, and then embedded with the first electrical constraints, including voltage, current, and short-circuit capacity. This ensures that the power grid meets basic safety and performance requirements under normal operating conditions, including ensuring that line current does not exceed its rated value and node voltage remains within permissible ranges.

[0064] Historical time-series data comprises operational data of the power grid over different time periods, including measurements of load curves and power generation. It reflects the actual operating state of the power grid at different points in time and serves as the basis for setting the second electrical constraint. The second electrical constraint, based on historical time-series data, considers the dynamic characteristics of the power grid under different operating conditions. It models the time-series fluctuation characteristics of load and distributed generation to support multi-period planning simulations. The second electrical constraint ensures the safe and stable operation of the power grid under various possible operating scenarios, guaranteeing grid stability during peak load periods or fluctuations in renewable energy output.

[0065] Execution steps: The distribution network planning dataset is the foundation for comprehensive planning and optimization of the distribution network. It covers all aspects of the distribution network, from energy application and technological development trends to power source configuration and load geographic distribution, as well as basic technical parameters and historical operating data of the power grid. By integrating these data, we can gain a comprehensive understanding of the current status and future development trends of the distribution network, providing data support for formulating scientific and reasonable distribution network planning schemes.

[0066] The power grid basic data provides the basic physical and electrical characteristics of the power grid. By setting the first electrical constraints based on this data, the basic electrical safety and performance requirements of the power grid are ensured to be met during normal operation. Specifically, based on the rated current and resistance of the lines, the line current is set to not exceed the rated value to prevent line overload and overheating. Furthermore, based on the rated power and voltage level of the transformers, the node voltage is set to be maintained within the allowable range, typically 95% to 105% of the rated voltage. Voltage constraints can be calculated using AC power flow equations to ensure that the voltage of each node is within the allowable range. Current constraints ensure that the current of all lines does not exceed their rated values ​​to prevent line overload and overheating. Current constraints can be calculated using AC power flow equations to ensure that the current of each line does not exceed its rated value. Short-circuit capacity constraints ensure that in the event of a short-circuit fault, the short-circuit current does not exceed the short-circuit capacity of the equipment. Short-circuit capacity constraints can be determined by short-circuit calculations to ensure that the short-circuit current does not exceed the short-circuit capacity of the equipment. By setting the first electrical constraints based on the power grid basic data, the basic electrical safety and performance requirements of the power grid are ensured to be met during normal operation.

[0067] Historical time-series data reflects the actual operating status of the power grid in different time periods. By setting the second electrical constraint condition using this historical time-series data, considering the dynamic characteristics of the power grid under different operating conditions, and based on historical load curves and power generation data, it is set that the voltage and current of the power grid remain within a safe range during load peaks or fluctuations in renewable energy output. Specifically, by analyzing load changes in different time periods, it is set that the voltage and current of the power grid must remain within a safe range during peak load periods. This can be achieved by setting dynamic voltage and current constraints to ensure stable operation of the power grid during peak load periods. Based on the power generation in the historical time-series data, the time-series fluctuations of distributed power sources, including solar and wind power, are modeled. Furthermore, by analyzing the fluctuations in renewable energy output, it is determined that the power balance and voltage stability of the power grid must meet requirements when renewable energy output fluctuates significantly. This can be achieved by setting dynamic power balance and voltage stability constraints, ensuring stable grid operation during renewable energy output fluctuations. By modeling the time-series fluctuation characteristics of load and distributed power sources, multi-period planning simulations are supported. Furthermore, a natural day can be divided into multiple time periods, each with different electrical constraints set based on historical data. During peak load periods, line capacity is increased and power supply configuration is optimized; during periods of renewable energy output fluctuation, the charging and discharging strategies of the energy storage system are adjusted to ensure safe and stable grid operation under different operating scenarios. By setting a second electrical constraint based on historical time-series data, the safe and stable operation of the power grid under various possible operating scenarios is ensured.

[0068] Preferably, by integrating the distribution network planning dataset and setting the first and second electrical constraints, scientific data support and a strict constraint framework are provided for distribution network planning. Furthermore, the first electrical constraint ensures the basic safety and performance of the power grid during normal operation, while the second electrical constraint considers the dynamic characteristics of the power grid under various possible operating scenarios. Specifically, in distribution networks containing multiple power sources and complex load distributions, by reasonably setting electrical constraints, risks and losses in long-term operation can be effectively reduced, ensuring the safe and reliable operation of the power grid.

[0069] Furthermore, the method of this application further includes constraining and shrinking the feasible operating domain and feasibility boundary using a coupling constraint matrix:

[0070] Based on the energy application data and technology development data, the feasibility boundary of the power distribution network planning strategy is set; based on the power supply configuration data and load space planning data, the coupling constraint matrix is ​​constructed.

[0071] Specifically, energy application data and technology development data cover energy usage patterns, the current application status of various energy technologies, and future development trends. Specifically, this includes the installed capacity of renewable energy, the configuration of energy storage systems, and the application of smart grid technologies. This helps predict the impact of future energy usage patterns and technological innovations on the power grid. The feasibility boundary of distribution network planning strategies refers to the optimal performance range achievable by a distribution network planning strategy under given energy application and technological development conditions. It defines the maximum benefit and minimum cost that a distribution network planning scheme can achieve under current technological and resource constraints. Furthermore, considering the capacity and charge / discharge rate of current energy storage technologies, it determines the power supply capacity and reliability of the distribution network under different load demands. The coupling constraint matrix is ​​used to represent and handle the interrelationships between multiple constraints. In distribution network planning, the coupling constraint matrix can comprehensively consider multiple factors such as power supply configuration, load distribution, and electrical constraints to form a unified constraint system. This helps optimize the operation and planning of the distribution network, ensuring that electrical safety and performance requirements are met under various conditions.

[0072] Execution steps: Analyze the energy application data and technology development data to understand the current and future energy usage patterns and the impact of technological innovations on the power grid. Specifically, if the data shows that the installed capacity of renewable energy will increase significantly in the next few years, then the planning strategy needs to consider how to better integrate these energy sources, determine the feasibility boundary of the distribution network planning strategy, that is, the maximum benefit that the distribution network planning scheme can achieve under the constraints of existing technology and resources, and determine the power supply capacity and reliability under different load demands, taking into account the current capacity and charge / discharge rate of energy storage technology. By setting the feasibility boundary, ensure that the distribution network planning scheme is both forward-looking and in line with the actual technical and resource conditions.

[0073] By utilizing power supply configuration data and load space planning data, a coupling constraint matrix is ​​constructed. The power supply configuration data includes information such as the type, location, and capacity of the power supply, while the load space planning data describes the geographical distribution of the load. Through these two sets of data, a constraint matrix that comprehensively considers the distribution of power supply and load is built. Furthermore, the coupling constraint matrix ensures that at any given time, the output of the power supply can meet the load demand while also satisfying electrical safety and performance requirements. By constructing this coupling constraint matrix, the operation and planning of the distribution network can be optimized, ensuring safe and stable operation under various conditions.

[0074] Preferably, by setting feasibility boundaries for distribution network planning strategies and constructing coupling constraint matrices, the feasibility and optimization of distribution network planning schemes under technical and resource constraints are ensured. The feasibility boundaries provide clear performance targets for the distribution network planning schemes, while the coupling constraint matrix ensures that electrical safety and performance requirements are met under various operating conditions. By rationally setting feasibility boundaries and constructing coupling constraint matrices, risks and losses in long-term operation are effectively reduced, ensuring the safe, reliable, and economical operation of the power grid.

[0075] Furthermore, the method of this application also includes iteratively executing a dynamic adjustment process based on the TVS diode protection mechanism, further comprising:

[0076] When the virtual prediction error triggers the loss threshold of the weak boundary region, the decision loss is suppressed to a safe range by step adjustment of the constraint contraction coefficient. Then, the steady-state adjustment stage is entered, and the distribution of the virtual prediction error and the constraint parameters are finely adjusted in the gradient descent direction to gradually approach the optimal solution and determine the distribution network planning scheme.

[0077] Specifically, in distribution network planning, virtual forecasting error includes load forecasting error and renewable energy output forecasting error, reflecting the deviation between the forecast value and the actual value; the loss threshold of weak boundary area refers to the decision loss reaching a critical value in the weak boundary area. Exceeding the critical value may lead to a significant increase in the risk of grid operation. Weak boundary areas include high load density areas, centralized renewable energy access points, etc. The loss threshold is a preset safety boundary used to trigger the protection mechanism.

[0078] A step adjustment of the constraint contraction coefficient refers to suddenly increasing the value of the constraint contraction coefficient to make the constraints more stringent, thereby quickly suppressing the decision loss to within a safe range. Step adjustment is a non-gradual adjustment method that can significantly change the constraints in a short period of time. After the decision loss is suppressed to a safe range, the system enters the steady-state adjustment phase. In the steady-state adjustment phase, more refined adjustments are made to gradually approach the optimal solution. Adjustments in the steady-state adjustment phase are usually gradual, optimizing the decision loss by fine-tuning the parameters. Fine-tuning along the gradient descent direction refers to calculating the gradient of the decision loss function and adjusting the parameters along the gradient descent direction to gradually reduce the decision loss. In the steady-state adjustment phase, fine-tuning along the gradient descent direction can help the model find a better parameter configuration.

[0079] Execution steps: In distribution network planning, when virtual prediction errors cause decision losses in weak boundary areas to exceed the preset loss threshold, it is necessary to take rapid measures to prevent further expansion of losses. Specifically, by step-wise adjustment of the constraint contraction coefficient, the constraint conditions can be tightened quickly. Furthermore, the TVS diodes quickly conduct under overvoltage conditions to clamp the voltage within a safe range. The rapid response mechanism can quickly reduce decision losses in weak boundary areas and prevent grid operation risks caused by transient errors. In particular, the step-wise adjustment of the constraint contraction coefficient can quickly suppress decision losses to a safe range and prevent grid operation risks caused by transient errors.

[0080] After the decision loss is suppressed to a safe range, the system enters the steady-state adjustment phase. In this phase, the decision loss is gradually optimized by fine-tuning the virtual prediction error distribution and constraint parameters. Specifically, the TVS diode maintains low leakage current under normal operating voltage to ensure the system operates in a stable state, allowing for fine-tuning in steady state to gradually optimize the decision loss. Specifically, the gradient of the decision loss function is calculated, and parameters are adjusted along the gradient descent direction. Further, the gradient distribution of the virtual prediction error is fine-tuned, and parameters are adjusted along the gradient descent direction to gradually optimize the decision loss, making it closer to the actual possible error distribution. Simultaneously, constraint parameters are fine-tuned to better suit the current optimization objective. Through this gradual adjustment, the model can gradually approach the optimal solution, determining the distribution network planning scheme. The distribution network planning scheme includes line expansion timing, power source access ratio, energy storage configuration strategy, and dynamic adjustment rules to address prediction deviations.

[0081] TVS diode's fast response mechanism: In power distribution networks, TVS diodes, as a protective device, can quickly conduct under overvoltage conditions, clamping the voltage within a safe range. This fast response mechanism can rapidly reduce decision-making losses in weak boundary areas and prevent power grid operation risks caused by transient errors. For example, when the voltage of a node rises sharply due to transient errors, the TVS diode can quickly conduct, clamping the voltage within a safe range, thereby protecting the equipment in that node and its surrounding area from high voltage damage.

[0082] The stability of TVS diodes: Under normal operating voltage, TVS diodes maintain low leakage current, ensuring that the system operates in a stable state. This stability allows the system to make fine adjustments in steady state and gradually optimize decision loss. For example, in the steady-state adjustment stage, by fine-tuning the virtual prediction error distribution and constraint parameters, the decision loss can be gradually optimized to make it closer to the actual possible error distribution. At the same time, the constraint parameters can be fine-tuned to make them more suitable for the current optimization objective.

[0083] Preferably, through rapid response and fine-tuning, the stability and reliability of the distribution network planning scheme are ensured when facing high-risk situations. Specifically, the step adjustment can take measures quickly when the decision loss exceeds the threshold to prevent the risk of grid operation from increasing further; the fine-tuning in the steady-state adjustment stage can gradually optimize the decision loss and find the optimal distribution network planning scheme. Through this dynamic adjustment mechanism, the uncertainty brought about by the fluctuation of new energy output and load changes can be effectively dealt with.

[0084] Furthermore, the method of this application also includes:

[0085] Using the typical power distribution scenario set output by the scenario generation model as the initial search starting point, multiple state transition paths are constructed along the load growth trajectory, the new energy output fluctuation path, and the N-1 fault transfer section. Small disturbances are introduced into each state transition path to simulate the dynamic drift process at the edge of the feasible operating domain, monitoring the gradient of decision loss changes in the corresponding power distribution network planning scheme. In the first state transition path, if the decision loss growth rate of the current state node exceeds a preset loss sensitivity threshold, the current state node and M associated state nodes are combined to form a locally highly sensitive section. Clustering expansion is then performed based on the geometric continuity and gradient consistency corresponding to the decision loss gradient distribution of the first state transition path to determine the risk manifold. The first state transition path is any one of the multiple state transition paths. Based on the geometric features of the risk manifold, including curvature and expansion direction, the sampling density function of the scenario generation model is adjusted inversely.

[0086] Specifically, a typical distribution scenario set refers to a set of representative distribution network operation scenarios output by a scenario generation model, reflecting the operating status of the distribution network under different loads, renewable energy output, and fault conditions; a state transition path refers to the transition path from one operating state to another. In distribution network planning, state transition paths can be constructed along load growth trajectories, renewable energy output fluctuation paths, and N-1 fault transfer sections; the decision loss change gradient refers to the rate of change of decision loss along the state transition path. By monitoring this gradient, we can understand how decision loss changes with state changes.

[0087] Locally sensitive regions refer to areas along the state transition path where the growth rate of decision loss exceeds a preset loss sensitivity threshold. These regions are particularly sensitive to changes in decision loss. Risk manifolds describe the distribution characteristics of decision loss in the state space. They include geometric features such as curvature and expansion direction, which can help identify and quantify the changing trends of decision loss. The sampling density function is a function in the scene generation model used to determine the density of scene generation in different state space regions. By adjusting the sampling density function, the scene generation density in specific regions can be increased or decreased.

[0088] Execution steps: The typical distribution scenario set output by the scenario generation model is used as the initial search starting point. This set contains representative states of the distribution network under different operating conditions. Multiple state transition paths are constructed along the load growth trajectory, the renewable energy output fluctuation path, and the N-1 fault transfer section, reflecting the possible transitions of the distribution network from one operating state to another. On each state transition path, small perturbations are introduced to simulate the dynamic drift process of the distribution network at the edge of the feasible operating domain. In this way, the stability and reliability of the distribution network can be tested when approaching its operating limits. Simultaneously, the gradient of decision loss changes is monitored to understand how the decision loss changes with state changes. This helps identify sensitive areas of decision loss.

[0089] In the state transition path, if the growth rate of the decision loss of a certain state node exceeds a preset loss sensitivity threshold, then this node and its M associated state nodes will be combined into a locally high-sensitivity segment. Based on the geometric continuity and gradient consistency of the decision loss gradient distribution, these high-sensitivity segments are clustered and expanded to determine the risk manifold. The geometric characteristics of the risk manifold, such as curvature and expansion direction, are used to identify and quantify the changing trend of decision loss. According to the geometric characteristics of the risk manifold, such as curvature and expansion direction, the sampling density function of the scene generation model is adjusted in reverse. Specifically, if the curvature of the risk manifold is large or the expansion direction points to a high-risk area, then the sampling density of these areas can be increased to study the operating state and decision loss of these areas in more detail.

[0090] Preferably, by constructing state transition paths, introducing small disturbances, monitoring the gradient of decision loss changes, identifying locally highly sensitive sections, determining risk manifolds, and adjusting the sampling density function, the stability and reliability of the distribution network planning scheme under different operating conditions are comprehensively evaluated. In this way, high-risk areas in the distribution network can be identified and addressed more accurately, the distribution network planning scheme can be optimized, and the long-term stable operation capability of the distribution network can be improved.

[0091] In summary, the beneficial effects of the embodiments of this application are:

[0092] This application provides a distribution network planning method based on decision-oriented learning. It achieves deep coupling between data and electrical constraints through feature analysis and constraint embedding, strengthens high-risk area scenario coverage by prioritizing weak boundary areas, enhances the ability to cope with uncertainties such as renewable energy fluctuations and load changes, and dynamically optimizes planning strategies by minimizing the expected decision loss function, aiming to minimize the expected decision loss. This avoids long-term resource waste and power supply risks, and improves the reliability of distribution network planning schemes by combining full-lifecycle loss assessment and iterative convergence judgment. The method aims to obtain a distribution network planning dataset of the target distribution network; based on mapping relationships, it trains a scenario generation model under a two-layer optimization architecture, including an inner optimization connection layer and an outer optimization connection layer. The inner optimization connection layer performs feature parsing and constraint embedding on the distribution network planning dataset, while the outer optimization connection layer prioritizes weak boundary areas as the guiding domain. Simultaneously, after each iteration, the expected decision loss over the entire lifecycle is re-evaluated to determine convergence. If the expected decision loss decreases less than a set tolerance during iteration, training stops, and a decision-oriented distribution network planning scheme is output.

[0093] In summary, any step can be stored as a computer instruction or program in an unrestricted computer memory and can be called and recognized by an unrestricted computer processor; no further restrictions are imposed here.

[0094] Furthermore, the above technical solutions only embody the preferred technical solutions of the embodiments of this application. Any changes that those skilled in the art may make to certain parts of these solutions embody the novel principles of the embodiments of this application. Obviously, those skilled in the art can make various modifications and variations to this application without departing from the scope of this application.

Claims

1. A decision-oriented learning based power distribution network planning method, characterized in that, The method comprises: obtaining power distribution network planning data set of target power distribution network; training scenario generation model under double-layer optimization architecture based on mapping relationship, taking minimization of expected decision loss of decision loss function as target; wherein, the double-layer optimization architecture comprises inner-layer optimization connection layer and outer-layer optimization connection layer, the inner-layer optimization connection layer performs feature analysis and constraint embedding on the power distribution network planning data set, and the outer-layer optimization connection layer takes weak boundary region as priority orientation domain; At the same time, re-evaluate expected decision loss under whole life cycle after each iteration, and judge whether to converge: if the expected decision loss in the iteration process decreases by less than a set tolerance, stop training, and output power distribution network planning scheme under decision orientation; wherein, training scenario generation model under double-layer optimization architecture comprises: the outer-layer optimization connection layer takes the weak boundary region as the priority orientation domain based on the mapping relationship and the decision loss chain; At the same time, optimize the constraint shrinkage coefficient of the inner-layer optimization connection layer, and cooperatively adapt the feasible operating domain and the feasibility boundary, reduce the sensitivity of the decision loss corresponding to the weak boundary region, and iteratively execute the dynamic adjustment process based on the TVS diode protection mechanism; wherein, the outer-layer optimization connection layer based on the mapping relationship and the decision loss chain further comprises: By differentiable optimization back propagation decision loss gradient, dynamically adjust the distribution characteristics of virtual prediction error and scenario generation weight, and strengthen the scenario coverage density of the weak boundary region; determine the mapping relationship according to the power grid topology model of the target power distribution network; wherein, determining the mapping relationship according to the power grid topology model of the target power distribution network comprises: Through the power grid topology model of the target power distribution network, configure multiple error combinations with virtual prediction errors under positive and negative deviations, and each error combination in the multiple error combinations executes power distribution network planning strategy to obtain the mapping relationship from input load prediction error to planning increment; wherein, taking the weak boundary region as the priority orientation domain further comprises: define the feasible operating domain through the first electrical constraint condition and the second electrical constraint condition; use the coupling constraint matrix to perform constraint shrinkage processing on the feasible operating domain and the feasibility boundary, and mark the weak boundary region; wherein, iteratively executing the dynamic adjustment process based on the TVS diode protection mechanism further comprises: When the virtual prediction error triggers the loss threshold of the weak boundary region, after suppressing the decision loss to a safe interval with step adjustment of the constraint shrinkage coefficient, enter the steady state adjustment stage, fine-tune the virtual prediction error distribution and the constraint parameter in the gradient descent direction, gradually approach the optimal solution, and determine the power distribution network planning scheme.

2. The decision-directed learning based power distribution network planning method of claim 1, wherein, Obtaining the mapping relationship from the input load prediction error to the planning increment, the method comprises: trace back the decision loss chain under the whole life cycle to determine the planning increment; The planning increment is used to represent the additional cumulative difference of the power distribution network planning strategy corresponding to the prediction deviation scene compared with the global optimal decision.

3. The decision-directed learning based power distribution network planning method of claim 1, wherein, Defining the feasible operating domain through the first electrical constraint condition and the second electrical constraint condition, the method further comprises: The power distribution network planning dataset includes energy application data and technical development data, power supply configuration data and load space planning data, power grid foundation data, and historical time series data; The power grid foundation data is used to set the first electrical constraint condition; The historical time series data is used to set the second electrical constraint condition.

4. The decision-directed learning based power distribution network planning method of claim 3, wherein, The method further includes: Based on the energy application data and technical development data, the feasibility boundary of the power distribution network planning strategy is set; Based on the power supply configuration data and load space planning data, the coupling constraint matrix is constructed.

5. The decision-directed learning based power distribution network planning method of claim 4, wherein, The method further includes: A typical power distribution scenario set output by the scenario generation model is used as an initial search starting point, and a plurality of state transition paths are constructed along a load growth trajectory, a new energy output fluctuation path, and an N-1 fault transfer section; A small perturbation is introduced on each state transition path to simulate a dynamic drift process at the edge of the feasible operation domain, and a decision loss change gradient of the corresponding power distribution network planning scheme is monitored; In the first state transition path, if the decision loss growth rate of the current state node exceeds a preset loss sensitivity threshold, the current state node and M associated state nodes are combined to form a local high-sensitivity section, and based on the geometric continuity and gradient consistency corresponding to the decision loss gradient distribution of the first state transition path, a risk manifold is determined, and the first state transition path is any one of the plurality of state transition paths; Based on the geometric features of the risk manifold including curvature and expansion direction, the sampling density function of the scenario generation model is adjusted in reverse.

Images