Method for improving distribution network photovoltaic carrying capacity considering multi-agent security constraints

Abstract

This invention relates to a method for enhancing the carrying capacity of distribution networks' photovoltaic (PV) power generation considering multi-entity security constraints, belonging to the field of power communication technology. First, this invention constructs a cloud-edge collaborative source-load-storage scheduling and multi-level security constraint framework to achieve security-aware scheduling. Second, it constructs a collaborative optimization problem for distribution network PV carrying capacity based on multi-objective trade-offs, maximizing distributed PV carrying capacity while minimizing costs. Finally, it proposes a distributed PV carrying capacity enhancement method based on cloud-edge collaborative security federation deep reinforcement learning, achieving cloud-edge collaborative iterative optimization. This invention can significantly improve the distribution network's capacity to accommodate distributed PV while reducing operating costs and grid losses, under the premise of strictly ensuring multi-level security constraints within and between distribution areas and protecting user data privacy. It is suitable for distribution network security and economical scheduling scenarios involving multiple stakeholders under high-penetration PV access.

Description

Technical Field

[0001] This invention relates to a method for improving the photovoltaic carrying capacity of distribution networks that takes into account the safety constraints of multiple entities, and belongs to the field of power communication technology. Background Technology

[0002] Currently, with the deepening of the global energy transition and the continuous reduction of dependence on fossil fuels, the penetration rate of renewable energy, represented by distributed photovoltaic (PV), in distribution networks is rapidly increasing. However, due to the randomness, intermittency, and source-load correlation of PV output, distribution networks are facing safety risks such as node voltage exceeding limits and line reverse power flow overload, becoming a key bottleneck restricting the capacity for distributed PV adoption. Fully exploring the regulation potential of flexible source-load-storage resources, while ensuring grid security constraints and operational economics, is the core requirement for improving the PV adoption capacity of distribution networks. Related optimized scheduling places stringent demands on multi-regional collaboration, privacy protection, and globally optimal decision-making.

[0003] Federated Deep Reinforcement Learning (FDRL), with its advantages in privacy protection and collaborative optimization, has been widely used in improving the distributed photovoltaic (PV) integration capacity of multi-regional distribution networks, becoming a mainstream technology for data-driven intelligent scheduling. However, existing FDRL research has significant shortcomings: it fails to fully integrate multi-level security constraints within / between distribution areas; agents are prone to sacrificing security boundaries for short-term rewards; unreasonable federated aggregation mechanisms lead to global model distortion; and it is difficult to adapt to the safe operation requirements of distribution networks with high PV penetration.

[0004] Addressing the dual demands of multi-level security constraints and collaborative optimization, the integration of cloud-edge collaborative architecture and secure reinforcement learning offers a feasible solution. Cloud-edge collaboration is naturally suited to distributed scheduling scenarios, with the edge handling local fine-grained control and the cloud enabling global collaboration and model aggregation, balancing data privacy with cross-regional complementary learning. Security awareness mechanisms can quantify and embed the risk of exceeding limits into the learning process, guiding strategies to proactively avoid security violations. However, existing methods do not deeply integrate intra- and inter-station security limit violations into policy learning, and federated aggregation does not weight edge models from a security perspective, easily leading to sudden global performance changes and deterioration of security indicators. Therefore, there is an urgent need to construct a cloud-edge collaborative secure federated deep reinforcement learning method. Through cloud-edge hierarchical constraint mechanisms and security adaptive model aggregation, this method can achieve collaborative optimization that maximizes the photovoltaic acceptance capacity of the distribution network, minimizes operating costs and network losses, and effectively suppresses security limit violations.

[0005] In view of the above-mentioned shortcomings, the present invention aims to create a method for improving the carrying capacity of photovoltaic distribution networks that takes into account the safety constraints of multiple entities, so as to make it more valuable for industrial application. Summary of the Invention

[0006] To address the aforementioned technical problems, the purpose of this invention is to provide a method for enhancing the photovoltaic carrying capacity of distribution networks that takes into account the safety constraints of multiple stakeholders.

[0007] The present invention provides a method for enhancing the photovoltaic carrying capacity of a distribution network considering multiple security constraints, comprising the following steps:

[0008] S1: Construct a cloud-edge collaborative source-load-storage scheduling and multi-level security constraint framework, complete the modeling of photovoltaic, load, and energy storage characteristics and the full constraint quantification of power flow, voltage, branch, and common connection point, ensure local hard constraints on the edge side, and coordinate global coupled security in the cloud to achieve security-aware scheduling; S2: Construct a collaborative optimization problem of distribution network photovoltaic carrying capacity based on multi-objective trade-offs. By jointly optimizing photovoltaic access capacity, photovoltaic output, reduceable load power, and energy storage charging and discharging power, the distributed photovoltaic carrying capacity is maximized while minimizing source-load-storage operation costs, grid loss costs, and safety default penalty costs. S3: Proposes a distributed photovoltaic carrying capacity improvement method based on cloud-edge collaborative secure federation deep reinforcement learning, including Markov decision process modeling considering security limit violation penalties, secure federation deep actor-commentator algorithm based on cloud-edge hierarchical constraint guarantee, and cloud-edge collaborative optimization method based on secure adaptive aggregation, to achieve cloud-edge collaborative iteration and effectively improve the photovoltaic carrying capacity of distribution network.

[0009] Furthermore, step S1 includes: S1.1: Source-load-storage characteristic modeling, including distributed photovoltaic carrying capacity modeling, load reduction characteristic modeling and energy storage characteristic modeling; S1.2: Source-load-storage scheduling constraint modeling, including distributed photovoltaic power output constraints, load reduction coefficient constraints, energy storage power output constraints, power flow constraints, low-voltage distribution area common connection point power constraints and power balance constraints, node voltage deviation constraints and branch current carrying capacity constraints.

[0010] Furthermore, in step S2, the optimization problem is constructed as follows:

[0011] The constraints include: source-load-storage coordinated dispatch constraints, power flow constraints within the distribution area, power constraints and power balance constraints at the point of common coupling, node voltage deviation constraints, and branch current carrying capacity constraints. In the formula, 1) Indicates the area i The set of all photovoltaic access capacity; 2) Indicates the area i The sum of all photovoltaic power outputs; 3) Indicates the area i The collection of all load-cutting power; 4) Indicates the area iThe sum of all battery energy storage charging power; 5) Indicates the area i The sum of all battery energy storage discharge power. , , , These are the weighting coefficients for distributed photovoltaic carrying capacity, equipment operating costs, voltage over-limit penalty costs, and grid loss costs, respectively.

[0012] Furthermore, step S3 includes: S3.1: Model a Markov decision process that considers the penalty for exceeding safety limits, embedding the violation of safety constraints into the state space so that the agent can accurately perceive the violation situation during the policy optimization process; S3.2: A secure federated deep actor critic algorithm based on cloud-edge hierarchical constraint guarantee, which transforms edge-side physical constraints into quantifiable penalty terms and embeds a reward function; S3.3: A cloud-edge collaborative optimization method based on security adaptive aggregation quantifies the edge model weights by fusing security violation penalty items within and between transformer substations, and adopts a soft update strategy to fuse historical parameters.

[0013] Furthermore, step S3.3 includes: Step 1: Initialize the agent network parameters on each source-load-storage edge control device; Step 2: Each source-load-storage edge control device downloads the global model from the cloud-based capacity enhancement control device; Step 3: Execute actions based on the current strategy and calculate rewards, then transition to the next time period; Step 4: Each agent in each station area calculates the temporal difference error based on the empirical samples and updates the parameters of the actor network and the critic network; Step 5: After the edge model training is completed, upload the trained edge model parameters to the cloud; Step 6: The cloud performs federated aggregation based on the received parameters of each edge model; Step 7: Repeat steps 2 to 6 until the preset time period is reached, and output the optimal distribution network photovoltaic carrying capacity improvement decision.

[0014] A distribution network photovoltaic carrying capacity enhancement system considering multi-entity security constraints, and an operation method comprising a cloud server layer, an edge server layer, and a device layer: The cloud server layer includes a cloud-based capacity enhancement and control device, which is responsible for model aggregation and security coordination across distribution substations. It receives model parameters uploaded by the source-load-storage edge control devices of each substation, performs model fusion based on a security adaptive aggregation mechanism, and distributes the optimized global model to the source-load-storage edge control devices of each substation. The edge server layer includes the source-load-storage edge control device for distribution areas, which is deployed in each distribution area and undertakes the responsibility of local decision-making and hard constraint guarantee. It performs local fine scheduling of source-load-storage based on real-time data from the equipment layer and forcibly guarantees that hard constraints such as photovoltaic output, energy storage charging and discharging, and load reduction within the distribution area do not exceed the limits through the boundary pruning function. The equipment layer includes distributed photovoltaics, load shedding systems, energy storage systems, power distribution lines, transformers, and other equipment. It uploads real-time collected operating status data such as power, voltage, current, and energy storage charge status to the source-load-storage edge control device in the distribution area and strictly executes the dispatch instructions issued by the source-load-storage edge control device.

[0015] Furthermore, the cloud-based capacity enhancement and control device is based on a safety adaptive aggregation mechanism, weighted and fused according to the safety penalty weight of the transformer area, and adopts a soft update strategy to fuse historical parameters. It only updates the global model when the safety performance is improved, and comprehensively handles the voltage, current and power coupling constraints between transformer areas. The source-load-storage edge control device of the transformer area is equipped with an intelligent agent. This intelligent agent is based on an actor-critic architecture, which ensures hard constraints through a boundary pruning function, transforms safety limit violations into penalty terms embedded in the learning process, and only uploads model parameters rather than raw data.

[0016] A source-load-storage edge control device for operating a distribution system includes: The power module provides a stable power supply to the various modules in the edge control device; The communication module enables secure data interaction between the cloud-based capacity enhancement and control device and the source-load-storage edge control device, as well as between the source-load-storage edge control device and terminal equipment. The source-load-storage characteristic modeling module establishes accurate mathematical characteristic models for distributed photovoltaic, load-shaving, and energy storage systems within the transformer area. The source-load-storage scheduling constraint calculation module generates equipment-level hard constraint boundaries such as photovoltaic output, load reduction, energy storage charging and discharging, and state of charge. The power flow calculation module performs accurate power flow calculations for low-voltage distribution areas based on the power flow model, and calculates key electrical quantities such as power at each node, branch loss, voltage and current. The load-bearing capacity cost weighted difference calculation module calculates the load-bearing capacity gain and the weighted difference of comprehensive cost, transforming multi-objective optimization into a single-value reward item; A Markov decision modeling module considering security defaults embeds security default information within and between transformer areas into the state space and reward function. The boundary function-based clipping module performs double-layer boundary clipping on the original scheduling actions output by the actor network. The source load storage scheduling action generation module obtains the original scheduling action by performing Gaussian sampling based on the actor network output, and generates compliant instructions after boundary function pruning. The safety violation penalty calculation module within the transformer area calculates and sums penalty values ​​for three types of violations: power exceeding the limit at the point of common coupling, voltage deviation exceeding the limit at the node, and branch current overload. The time series error calculation module calculates the time series difference error based on the current reward, discount factor, and the value of the next state. The actor network update module uses temporal difference error as the optimization basis and adopts the gradient ascent algorithm to iteratively update the actor network parameters. The commentator network update module uses the gradient descent algorithm to minimize the squared temporal difference error and iteratively optimizes the commentator network parameters.

[0017] Furthermore, the source-load-storage characteristic modeling module is connected to the source-load-storage scheduling constraint calculation module; the source-load-storage scheduling constraint calculation module is connected to the power flow calculation module and the boundary function-based pruning module; the power flow calculation module is connected to the source-load-storage scheduling constraint calculation module and the bearing capacity cost weighted difference calculation module. The boundary function-based pruning module performs two-layer pruning on scheduling actions. The first layer satisfies the hard constraints of the equipment, and the second layer adapts to the characteristics of energy storage and load, mapping infeasible actions to safe and feasible ranges. The module for calculating penalties for breach of contract within the station area is connected to the module for calculating the weighted difference of bearing capacity cost and the module for Markov decision modeling that considers breach of contract; the module for calculating time-series errors is connected to the module for updating the actor network, the module for updating the critic network, and the module for Markov decision modeling that considers breach of contract.

[0018] A cloud-based capacity enhancement and control device for operating a system, comprising: The power module provides stable power to all modules in the cloud-based capacity enhancement and control device. The communication module enables secure data exchange between the cloud-based capacity enhancement and control device and the edge devices of each distribution area; The overload penalty calculation module for reverse load rate of the transformer area takes the common connection point of the transformer area as the object, calculates the power interaction between the transformer area and the upstream power grid, and calculates the forward and reverse load rates in combination with the rated capacity of the transformer. The inter-station node voltage over-limit penalty calculation module calculates the voltage deviation between ordinary nodes and aggregated interconnection nodes within the station area in a hierarchical manner, distinguishing between voltage over-limit scenarios within the station area and those coupled between station areas. The overload penalty calculation module for branch current carrying capacity between distribution substations calculates the current amplitude of branches within the distribution substation and interconnecting branches across distribution substations based on the distribution network topology and thermal stability limits. The penalty normalization module performs exponential normalization on the safety default penalty values ​​within and between transformer areas uploaded by each source-load-storage edge control device. The cloud-based federated aggregation module receives network parameters of actors and critics uploaded by edge control devices of various source load storage, and performs weighted fusion of edge models of multiple stations according to security weights; The soft update module based on security penalty weights uses the security penalty weights as coefficients to integrate the parameters of the historical global model and the current aggregated model, and decides whether to update the global model after comparing the security performance.

[0019] Furthermore, the penalty term normalization module converts the penalty value into a weight in the 0-1 range using the Softmax function, providing a safe weighting basis for federated aggregation; The soft update module based on security penalty weights adopts a soft update strategy, which merges the new aggregated parameters with historical parameters and updates the global model only when the current model has better security performance, thus avoiding large fluctuations in security performance. The communication module is connected to the edge communication module, the penalty item normalization module, the cloud federated aggregation module, and the soft update module based on security penalty weights; the transformer area reverse load rate overload penalty calculation module, the transformer area node voltage over-limit penalty calculation module, and the transformer area branch current carrying capacity overload penalty calculation module are all connected to the penalty item normalization module, the cloud federated aggregation module, and the soft update module based on security penalty weights.

[0020] By means of the above-described solution, the present invention has at least the following advantages: (1) This invention proposes a method for improving the carrying capacity of photovoltaic distribution networks that considers the security constraints of multiple stakeholders. It achieves global optimization based on cloud-edge collaborative architecture and security awareness reinforcement learning, specifically including: 1) Cloud-Edge Collaborative Source-Load-Storage Scheduling Architecture: This layered architecture aims to enhance the acceptance capacity of distributed photovoltaic (PV) systems while minimizing operating costs, eliminating network losses and voltage over-limit issues, and ensuring the economical and safe operation of the system. DAC (Digital Access Control) agents are deployed on the source-load-storage edge control devices in the distribution area to achieve inter-distribution scheduling of PV, adjustable loads, and energy storage. The cloud-based capacity-enhancing control device manages inter-distribution coupling constraints and optimizes the edge scheduling model through federated aggregation technology.

[0021] 2) Security FDAC Based on Security Violation Penalties and Cloud-Edge Layered Constraint Guarantee: Security constraint violations within and between power distribution units are embedded into the MDP state space to achieve security awareness. These violations are further transformed into quantified penalties and embedded in a reward function to guide the DAC agent to proactively avoid inter-power distribution violation risks during policy learning. Constraint guarantees within the power distribution unit are achieved through hard constraint pruning technology on the source-load-storage edge control device.

[0022] 3) Security Adaptive Global Model Aggregation Mechanism: The proposed mechanism quantifies edge model weights by fusing security violation penalties within and between substations. This mechanism integrates current aggregation parameters (hard update) with historical model parameters to maintain consistency, and updates are only performed when the global model's security performance is better. This soft update approach effectively preserves reliable historical security characteristics while incorporating high-quality local experience, thereby mitigating the problem of sudden security performance changes caused by poor edge model quality.

[0023] (2) The present invention designs a distribution network photovoltaic carrying capacity enhancement system that considers the safety constraints of multiple entities. It consists of a distribution area source-load-storage edge regulation device and a cloud-based carrying capacity enhancement regulation device. The functions and connections of each module in the device are as follows: 1) Source-load-storage edge control device in the distribution area The device is based on MDP modeling with safety limit violation penalties, constructs a safe FDAC algorithm with cloud-edge hierarchical constraint guarantee to carry out local agent training, uses an Actor-Critic network to complete policy iteration and value evaluation, updates network parameters through temporal differential error, ensures that scheduled actions meet hard constraints through double-layer pruning of boundary functions, and achieves local optimization under privacy protection by uploading only model parameters.

[0024] 2) Cloud-based capacity enhancement and control device Based on a secure adaptive federated aggregation mechanism, the device performs security penalty weighted fusion on the edge models of each distribution area, and completes global model iterative optimization by relying on a secure soft update strategy. The global model is updated only when the security performance is better, so as to achieve cross-distribution area collaborative scheduling and global security optimization under the protection of the privacy of multiple subjects, and support the stable improvement of the photovoltaic carrying capacity of the distribution network.

[0025] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show a certain embodiment of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a diagram of the distribution network photovoltaic carrying capacity enhancement system considering multiple subject safety constraints, as described in this invention. Figure 2 This is a schematic diagram of the source-load-storage edge regulation device of the present invention; Figure 3This is a schematic diagram of the cloud-based capacity enhancement and control device of the present invention; Figure 4 This is a flowchart of the method for improving the load-bearing capacity of photovoltaic distribution networks that considers the safety constraints of multiple entities, as proposed in this invention. Detailed Implementation

[0028] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0029] This invention is a method for improving the carrying capacity of photovoltaic distribution networks considering multiple security constraints, comprising three parts: system, device, and method, as described in detail below.

[0030] This invention proposes a distribution network photovoltaic carrying capacity enhancement system that considers multiple subject safety constraints, such as... Figure 1 As shown, it includes the cloud server layer, edge server layer, and device layer. Each layer is described below: Cloud Server Layer: This layer is the core of global collaborative control. Its core device is the cloud-based capacity enhancement and regulation device, responsible for model aggregation and security coordination across distribution substations. The cloud-based capacity enhancement and regulation device receives model parameters uploaded by the source-load-storage edge regulation devices of each substation. Based on a security-adaptive aggregation mechanism, it weights and fuses models according to substation security penalty weights and employs a soft update strategy to fuse historical parameters. The global model is updated only when security performance is improved, comprehensively handling voltage, current, and power coupling constraints between substations to avoid global security mutations caused by inferior local models. Finally, the optimized global model is distributed to the source-load-storage edge regulation devices of each substation, achieving multi-substation collaborative optimization.

[0031] Edge Server Layer: This layer includes the source-load-storage edge control devices deployed in each distribution substation, responsible for local decision-making and ensuring compliance with hard constraints. These devices are equipped with intelligent agents that perform fine-grained local scheduling of source, load, and storage based on real-time data from the device layer. They enforce strict limits on photovoltaic output, energy storage charging and discharging, and load reduction within the substation through boundary pruning functions, and convert safety limit violations into penalties embedded in the learning process. The edge control devices only upload model parameters, not raw data, completing local training while protecting privacy and providing reliable local model parameter data for cloud aggregation.

[0032] Equipment Layer: This layer is the physical execution layer, containing distributed photovoltaic (PV) systems, load shedding systems, energy storage systems, power distribution lines, transformers, and other equipment. This layer uploads real-time operational status data, such as power, voltage, current, energy storage, and state of charge, to the source-load-storage edge control device in the distribution area. It strictly executes the dispatching instructions issued by the source-load-storage edge control device, completing actions such as PV output adjustment, load shedding, and energy storage charging and discharging. Simultaneously, it adheres to the equipment's own operational boundary constraints, providing real operational feedback for upper-level decision-making.

[0033] This invention proposes a source-load-storage edge control device for a transformer substation, such as... Figure 2 As shown. The modules in the device are described below: Power supply module: Provides stable power supply to each module in the source-load-storage edge control device of the distribution area, ensuring the continuous operation of the device.

[0034] Communication module: Enables secure data interaction between the cloud-based capacity enhancement and control device and the source-load-storage edge control device, and between the source-load-storage edge control device and terminal equipment. It completes global model download, local model upload, transmission of scheduling instructions, and collection of source-load-storage operation data.

[0035] Source-load-storage characteristic modeling module: used to establish accurate mathematical characteristic models of distributed photovoltaic, load shedding and energy storage systems in the transformer area, calculate photovoltaic access capacity, describe load shedding laws, update energy storage state of charge, and convert physical characteristics into algorithm-recognizable expressions.

[0036] The source-load-storage scheduling constraint calculation module generates equipment-level hard constraint boundaries such as photovoltaic output, load reduction, energy storage charging and discharging, and state of charge. It converts the safe operating range into numerical thresholds, provides constraint intervals for action trimming, and prevents scheduling from exceeding limits and damaging equipment.

[0037] Power flow calculation module: Based on the power flow model, it completes accurate power flow calculation for low-voltage distribution areas, solves the power, branch loss, voltage and current of each node, satisfies power balance and topology constraints, and provides data for constraint calculation and safety penalty calculation.

[0038] The load-bearing capacity cost weighted difference calculation module aims to maximize photovoltaic load-bearing capacity, minimize operating costs, grid loss and safety default penalties, calculate the load-bearing capacity gain and the comprehensive cost weighted difference, and transform multi-objective optimization into a single-value reward item.

[0039] The Markov decision modeling module considering security defaults embeds security default information within and between transformer substations into the state space and reward function, constructs a security-aware Markov decision process, integrates operational data to form a state vector, defines scheduling actions and reward functions, and supports security decision-making.

[0040] The boundary function-based pruning module performs two-layer boundary pruning on the original scheduling actions output by the Actor network. The first layer satisfies the hard constraints of the equipment, and the second layer adapts to the characteristics of energy storage and load, mapping the inoperable actions to the safe and feasible range.

[0041] The source-load-storage scheduling action generation module obtains the original scheduling action by performing Gaussian sampling based on the Actor network output. After boundary function pruning, it generates compliant instructions and outputs complete scheduling instructions for photovoltaics, loads, and energy storage to drive terminal equipment to execute scheduling.

[0042] The safety violation penalty calculation module within the transformer area calculates and sums penalty values ​​for three types of violations: power exceeding the limit at the point of common coupling, voltage deviation exceeding the limit at the node, and branch current overload. It transforms safety constraints into quantitative penalties and embeds a reward function to form negative feedback.

[0043] The temporal error calculation module calculates the temporal difference error based on the current reward, discount factor, and next state value, quantifies the deviation between actor strategy and critic value assessment, and provides a basis for updating the parameters of the Actor and Critic networks.

[0044] The Actor network update module uses temporal differential error as the optimization basis and adopts the gradient ascent algorithm to iteratively update the actor network parameters, continuously optimizing the source-load-storage scheduling strategy to make the actions conform to the goals of high carrying capacity, low default rate, and low cost.

[0045] The Critic network update module uses the gradient descent algorithm to minimize the squared temporal difference error, iteratively optimizes the Critic network parameters, improves the accuracy of long-term action value assessment, provides accurate feedback to the Actor network, and accelerates model convergence.

[0046] This invention proposes a cloud-based capacity enhancement and control device. The modules within the device are described below: Power module: Provides stable power to all modules in the cloud-based capacity enhancement and control device, ensuring the continuous operation of the device.

[0047] Communication module: Used to realize secure data interaction between the cloud-based capacity enhancement and control device and the edge devices of each distribution area, undertake the function of cloud-edge collaboration hub, uniformly distribute the global optimization model downwards, receive local model parameters uploaded by each edge device upwards, and complete cloud-edge model synchronization and data transmission.

[0048] Transformer Area Reverse Load Rate Overload Penalty Calculation Module: Taking the transformer area's common connection point as the object, it calculates the power interaction between the transformer area and the upstream power grid, calculates the forward / reverse load rate in combination with the transformer's rated capacity, identifies the risk of reverse power overload and transformer over-limit operation, and outputs safety judgment indicators.

[0049] Inter-station node voltage over-limit penalty calculation module: calculates the voltage deviation between ordinary nodes and aggregated interconnected nodes within the transformer area in layers, distinguishes between voltage over-limit scenarios coupled between transformer areas, determines the over-limit status and outputs the over-limit magnitude, and provides a basis for safety penalties.

[0050] The overload penalty calculation module for branch current carrying capacity between distribution substations calculates the current amplitude of branches within the distribution substation and interconnecting branches across distribution substations based on the distribution network topology and thermal stability limits, verifies the safe operating range of current, identifies potential overload hazards, and generates safety violation marks.

[0051] Penalty Normalization Module: Performs exponential normalization on the safety default penalty values ​​within and between distribution zones uploaded by each source-load-storage edge control device. The penalty values ​​are converted into 0-1 interval weights using the Softmax function, providing a safety weighting basis for federated aggregation.

[0052] Cloud-based federated aggregation module: Receives Actor and Critic network parameters uploaded by edge control devices of various sources, loads and storage, performs weighted fusion of edge models of multiple distribution areas according to security weights, generates a preliminary global model, and realizes global collaborative training.

[0053] The soft update module based on security penalty weights: using security penalty weights as coefficients, it integrates the parameters of the historical global model and the current aggregated model, compares the security performance, and decides whether to update the global model, thus achieving smooth iteration of the global model.

[0054] This invention proposes a method for enhancing the carrying capacity of photovoltaic distribution networks considering multiple security constraints. The specific process is as follows: Figure 4 As shown.

[0055] S1: Construct a cloud-edge collaborative source-load-storage scheduling and multi-level security constraint framework, complete the modeling of photovoltaic, load, and energy storage characteristics and the full constraint quantification of power flow, voltage, branches, and common connection points, etc. The edge side ensures local hard constraints, and the cloud coordinates global coupled security to achieve security-aware scheduling.

[0056] S2: Construct a collaborative optimization problem of distribution network photovoltaic carrying capacity based on multi-objective trade-offs. By jointly optimizing photovoltaic access capacity, photovoltaic output, load that can be reduced, and energy storage charging and discharging power, the distributed photovoltaic carrying capacity is maximized while minimizing source-load-storage operation costs, grid loss costs, and safety default penalty costs.

[0057] S3: Proposes a distributed photovoltaic carrying capacity improvement method based on cloud-edge collaborative secure federated deep reinforcement learning, including a Markov decision model considering security limit violation penalties, a secure FDAC (Federated Deep Actor-Critic, FDAC) algorithm based on cloud-edge hierarchical constraint guarantee, and a cloud-edge collaborative optimization method based on secure adaptive aggregation, to achieve cloud-edge collaborative iteration and effectively improve the photovoltaic carrying capacity of the distribution network.

[0058] In a further embodiment, step S1 includes: S1.1 Source-Load-Storage Characteristic Modeling Distribution network by I The distribution radio station area is composed of a set, denoted as a set. . No. i Individual distribution radio area includes J i Taiwanese photovoltaic units, K i A load that can be reduced andL i A set of energy storage systems, denoted as a collection. , and Optimization time includes T Each time period is denoted as a set. , The duration of a single time period.

[0059] This section models the operating characteristics of distributed photovoltaic systems, load shedding, and energy storage batteries, as detailed below.

[0060] S1.1.1 Distributed Photovoltaic Carrying Capacity Modeling To meet constraints such as branch current carrying capacity and node voltage deviation, and considering the temporal randomness of user loads, distributed photovoltaic power output has different allowable capacities at different nodes during different operating periods.

[0061] definition for t Time zone i No. j The permitted capacity of a photovoltaic power station, the area i No. t Distributed photovoltaic carrying capacity in a given time period Represented as (1) S1.1.2 Modeling of Reduced Load Characteristics To mitigate node voltage exceedance issues caused by heavy load on the load side, demand-side response technology is used to mobilize reduceable loads within the distribution area for coordinated dispatch. By introducing a load reduction factor, the operating power of reduceable loads is dynamically adjusted. Definition Taiwan District i The Middle The power reduction of a load is expressed as (2) In the formula, The load power before reduction, This is the load reduction factor.

[0062] S1.1.3 Energy Storage Characteristic Modeling Battery energy storage controllers enhance the distributed photovoltaic carrying capacity of the power grid by adjusting charging and discharging behavior. The dynamic update of the energy storage system's State of Charge (SoC) is represented as follows: (3) In the formula, Taiwan District i No. Energy storage period tBattery state of charge; and For the charging and discharging efficiency of the energy storage system; and These are the charging power; This refers to the rated capacity of the energy storage system.

[0063] S1.2 Source-Load-Storage Scheduling Constraints Photovoltaic output, load shedding, and energy storage charging and discharging are characterized by randomness and volatility. Without constraints, this can easily lead to risks such as voltage exceeding limits and line overload. Constraints can regulate equipment operating boundaries, preventing equipment damage. Simultaneously, they can balance power flow, reduce grid losses and operating costs, and promote photovoltaic power consumption. Details are as follows.

[0064] S1.2.1 Distributed Photovoltaic Output Constraints The output of photovoltaic inverters must not exceed the upper limit of the permitted access capacity approved by the grid dispatch. The output constraint for distributed photovoltaic systems is expressed as follows: (4) S1.2.2 Load Reduction Factor Constraint The load reduction factor constraint is expressed as (5) In the formula, This represents the upper limit of the load reduction factor.

[0065] S1.2.3 Energy Storage Output Constraints To protect energy storage devices and prevent overcharging and discharging, the charging and discharging of energy storage systems must meet capacity constraints. Energy storage output constraints consist of charging power constraints, discharging power constraints, and energy storage capacity constraints, expressed as follows: (6) (7) (8) In the formula, and These represent the maximum allowable charging and discharging power of the battery. and These are the upper and lower limits of the battery's state of charge, respectively. Equation (8) ensures that the energy storage battery cannot be charged and discharged simultaneously at the same time.

[0066] S1.2.4 Power Flow Constraints Power flow constraints within a transformer substation consist of two parts: power flow constraints within the substation area and power flow constraints between substations. (Based on the substation area...) i For example, based on the Distribution Flow (Distflow) model, flow constraints are constructed and represented as follows: (9) In the formula, and They are respectively the Taiwan area i A set of nodes and lines. and They are nodes Flow to Node The active and reactive power. and They are respectively from nodes The outflow of active and reactive power. and The lines are respectively The resistance and reactance. For nodes The root mean square of the voltage. For the line The root mean square of the current. Represents a node All downstream nodes A set of. Represents a node All upstream nodes A set of.

[0067] S1.2.5 Power Constraints and Power Balance Constraints at the Point of Common Coupling in Low-Voltage Distribution Areas Low-voltage distribution substations are connected to the medium-voltage distribution network via distribution transformers. To prevent forward or reverse overload of the transformers, the power exchange between the low-voltage substations and the medium-voltage distribution network needs to be constrained. The power at the point of common coupling (PCC) of the low-voltage substations... The constraint is expressed as (10) In the formula, and Taiwan District i The maximum forward and reverse power that the transformer is allowed to transmit.

[0068] The power balance constraint of the low-voltage distribution area is expressed as: (11) In the formula, Taiwan District i exist t Network loss power during a given time period.

[0069] S1.2.6 Node voltage deviation constraint High-proportion distributed photovoltaic (PV) grid integration can easily cause node voltage fluctuations, therefore, strict limits must be placed on the voltage deviation of each node. Node voltage deviation constraints consist of two parts: voltage deviation constraints for each node within a distribution area, and aggregation node voltage deviation constraints for low-voltage distribution areas connected to the medium-voltage distribution network, expressed as follows: (12) In the formula, Taiwan District i Middle node o Voltage deviation; The voltage deviation of the aggregation node in the transformer area; and Taiwan District i Middle node o The maximum negative voltage deviation and the maximum positive voltage deviation; and They are respectively the Taiwan area i Middle node o The equivalent resistance and equivalent reactance; and For injection into the substation area i Middle node o The active and reactive power limits; Taiwan District i Middle node o The rated voltage. The constraints for low-voltage distribution areas connecting to medium-voltage distribution network nodes are similar and will not be repeated.

[0070] S1.2.7 Branch current carrying capacity constraint To prevent line overload, the current amplitude flowing through each branch must not exceed the thermal stability limit of the conductor. Branch current-carrying constraints include inter-substation current-carrying constraints and intra-substation current-carrying constraints, expressed as: (13) In the formula, and In order to cooperate with Taiwan i node o The current in all connected branches; and This represents the current-carrying limit of the corresponding branch. Taiwan District i Taiwan The current in the interconnecting branches, This represents the current carrying capacity limit of the interconnected branch.

[0071] S2 optimization problem modeling Enhancing the carrying capacity of distributed photovoltaic (PV) power is crucial for achieving high penetration rates of renewable energy while reducing carbon emissions. The randomness and intermittency of distributed PV easily lead to node voltage exceedances and line overloads, which are core bottlenecks restricting its absorption. Reducing the operating costs of source-load-storage (SHS) equipment can optimize economic feasibility and promote the large-scale deployment of flexible control resources; minimizing network losses can improve energy utilization efficiency and reduce energy waste in the transmission process; suppressing voltage exceedances can eliminate safety risks and operational constraints that force PV curtailment. The synergistic effect of these measures can expand the safe operation boundaries of the distribution network and fully tap the potential of flexible control of source-load-storage resources.

[0072] Therefore, the optimization objective of this invention is to maximize the distributed photovoltaic carrying capacity while reducing operating costs, network losses, and voltage over-limit penalty costs through source-load-storage coordinated scheduling. The optimization variables are defined as follows: 1) Indicates the area i The set of all photovoltaic access capacity; 2) Indicates the area i The sum of all photovoltaic power outputs; 3) Indicates the area i The collection of all load-cutting power; 4) Indicates the area i The sum of all battery energy storage charging power; 5) Indicates the area i The set of all battery energy storage discharge powers. The optimization problem is formulated as follows: (14) In the formula, , , , These are the weighting coefficients for distributed photovoltaic carrying capacity, equipment operating costs, voltage over-limit penalty costs, and grid loss costs, respectively. Constraints for coordinated scheduling of energy sources, loads, and storage. For the tidal current constraint within the area. For PCC power constraints and power balance constraints. For node voltage deviation constraints. For branch line current carrying capacity constraints.

[0073] The equipment operating cost is expressed as (15) In the formula, Unit operating cost of distributed photovoltaic power; The cost of supplying electricity per unit load; The unit operating cost of charging and discharging the energy storage system; The configuration depreciation cost per unit of photovoltaic capacity; This reduces the unit penalty cost for load reduction.

[0074] Taiwan District i The voltage over-limit penalty cost, which consists of both the inter-station and intra-station voltage over-limit penalty items, is expressed as follows: (16) In the formula, and These represent the unit penalty costs for voltage exceeding limits between and within transformer substations, respectively.

[0075] Network loss cost for a distribution transformer area reflects the energy utilization efficiency of the area's operation. Distributed photovoltaic (PV) grid connection may alter power flow distribution and even cause power flow backflow, thus affecting line losses. Network loss cost is expressed as... (17) In the formula, Cost per unit of network loss.

[0076] This invention proposes a method for enhancing the carrying capacity of distribution network photovoltaics (PV) systems while considering multi-entity safety constraints. First, the method constructs a cloud-edge collaborative source-load-storage scheduling architecture to dynamically integrate topology and operational status information from multiple distribution areas. It then combines source-load-storage characteristics to model and collaboratively optimize PV access capacity, load reduction, and energy storage charging and discharging behavior. A multi-level safety constraint mechanism is introduced, quantifying constraints such as node voltage deviation, branch current carrying capacity, and transformer reverse load rate as boundary conditions and embedding them into the optimization model, enabling the system to possess safety operation boundary awareness and compliant control capabilities. Second, a multi-objective optimization objective construction mechanism is designed. By comprehensively balancing PV carrying capacity, operating costs, voltage exceedance penalties, and network loss costs, it effectively balances capacity and operational safety while reducing control complexity. The proposed method, through a dynamic balance mechanism between safety and efficiency, achieves collaborative optimization and efficient utilization of distribution network source-load-storage resources, overcoming the key challenge of balancing safety constraints and optimization objectives in traditional methods. The specific implementation steps are detailed in S3, including three parts: "Markov Decision Process (MDP) Modeling Considering Security Limit Exceedance Penalties", "FDAC Algorithm Based on Cloud-Edge Constraint Mapping", and "Cloud-Edge Collaborative Model Optimization Based on Security Adaptive Aggregation". The specific implementation steps are as follows.

[0077] S3.1 MDP Modeling Considering Safety Limit Penalties Based on a safety penalty mechanism, this invention improves the traditional MDP model by embedding safety constraint violation into the state space. This enables the agent to accurately perceive violation situations during policy optimization, thereby allowing the generated scheduling strategy to proactively avoid safety violation risks while improving the grid-connected capacity of photovoltaic systems.

[0078] S3.1.1 State Space State space definition i Individual distribution radio station area t The operational status of a time period includes the following key vectors: t Access capacity of distributed photovoltaic power during the -1 time period Photovoltaic power output It can reduce load power. Battery energy storage charging / discharging power / , represented as (18) In the formula, Indicates the first i Energy storage system in individual distribution substation area t The charged state vector for a given time period; Indicates the first i Individual distribution radio station area t A vector formed by the PCC voltage deviation over a time period of 1; Indicates the first i Individual distribution radio station area t Total PCC voltage deviation during period 1; , , They are respectively t- 1st time zone i Operating costs, voltage deviation penalty costs, and network loss costs.

[0079] S3.1.2 Motion Space Define the area i The action space is It contains a set of decision actions for source-load-storage coordinated scheduling, and its expression is: (19) S3.1.3 Reward Function definition For the first i The intelligent agent in each distribution radio zone t The reward obtained during a time period, the reward function provides positive and negative feedback to the agent's actions, guiding the learning process towards the optimal policy, is represented as... (20) In the formula, and These are the safety performance weighting coefficients within and between transformer substations, used to quantify the degree of importance attached to different levels of safety limits. The larger the weight value, the higher the safety requirements for that level. and These are the sum of safety over-limit penalties within and between transformer substations, respectively. The specific calculation method will be explained in detail in S3.2.

[0080] S3.2 Secure FDAC Algorithm Based on Cloud-Edge Hierarchical Constraints For the MDP model constructed above, this invention proposes a secure FDAC algorithm based on cloud-edge hierarchical constraint assurance. This algorithm aggregates the parameters of all distribution area optimization models to solve for the optimal source-load-storage collaborative scheduling strategy; it transforms edge-side physical constraints into quantifiable penalty terms, enabling edge agents to satisfy local hard constraints through boundary functions; and it further embeds these penalty terms into the reward function, guiding the distributed Actor-Critic (DAC) agent to actively avoid inter-station boundary exceedance risks during policy learning, thereby achieving global collaboration between cloud and inter-station coupling.

[0081] S3.2.1 Actor-Critic Architecture The proposed algorithm employs the Actor-Critic framework, where the Actor network is in its current state. As input, output the current policy. , that is, the mean and variance of each action in the action space. Wherein, for t Time zone i The parameters of the Actor network model. Taking the charging and discharging of an energy storage system as an example, the sets of its mean and variance are respectively expressed as... and Similarly, the distribution areas can be obtained separately. i The set of mean and variance of the access capacity for distributed photovoltaic power. and The set of mean and variance of photovoltaic power output and This can reduce the set of mean and variance of load power. and The Critic network is used to evaluate the value of the current action. ,in for t Time zone i The parameters of the Critic network model are determined. This architecture enables the co-optimization of policy generation and value assessment to maximize the optimization objective.

[0082] S3.2.2 Action Generation Mechanism Taking an energy storage system as an example, its action generation is represented as follows: (twenty one) In the formula, Taiwan District i The charging and discharging power vector of the energy storage system. This represents a normal distribution. To ensure the actions satisfy... This section uses the charging / discharging power constraints of an energy storage system as an example. Boundary functions are used to limit the battery's charging / discharging power within a range that satisfies the constraints, resulting in the final charging / discharging power, expressed as: (twenty two) (twenty three) In the formula, It is a boundary function. and express The results of the first and last cuts correspond to equations (6) and (3). The first cut ensures... The second cut emphasizes meeting energy storage output constraints. Consistency with energy storage characteristics. When hour, , indicating the first l The energy storage system is in a charging state when it is charging, and in a discharging state when it is discharging. Formulas (22) and (23) respectively make the energy storage system in a charging state when it is discharging. and It meets the charging and discharging power constraints and energy storage capacity constraints of the energy storage system. , , , , They are respectively the Taiwan area i A collection of energy storage rated capacity, battery capacity lower limit, battery capacity upper limit, discharge power, and charging power.

[0083] Similarly, the photovoltaic access capacity action can be obtained. ,satisfy China's photovoltaic power output constraints ,satisfy The load power constraint can be reduced. Source-load-storage scheduling strategy It can be represented as (twenty four) S3.2.3 Calculation of Safety Limit Exceeding Penalties To guide the algorithm to satisfy constraints such as power flow, node voltage deviation, branch current carrying capacity, and transformer reverse load rate as much as possible, a sum of safety over-limit penalties is set for both within and between transformer substations. and , represented as (25) (26) In the formula, , , The first i Penalty coefficients corresponding to PCC power constraints, node voltage deviation constraints, and branch current carrying capacity constraints in each distribution substation area. , These are the penalty coefficients corresponding to the inter-station node voltage deviation constraint and the inter-station branch current carrying capacity constraint, respectively. , , The first i In each distribution substation area, the PCC power constraint, the node voltage deviation constraint, and the branch current carrying capacity constraint penalty indicator variables are: , These are the penalty indicator variables for inter-station node voltage deviation constraints and branch current carrying capacity constraints, respectively. The penalty indicator variable is 0 when the corresponding constraint is satisfied, and 1 otherwise.

[0084] S3.3 Cloud-Edge Collaborative Optimization Method Based on Security Adaptive Aggregation After training the source-load-storage scheduling model on the source-load-storage edge control device, the intelligent agents in each distribution substation upload the trained model parameters to the cloud-based capacity enhancement control device for federated aggregation. A safety adaptive mechanism is incorporated into the aggregation process. The core idea is to transform physical safety constraints into quantifiable safety indicators by utilizing safety over-limit penalties within and between substations. By comparing the new aggregated model with the global model from the previous time period, the global model is updated only if the current model has better safety performance (i.e., a lower penalty value). Subsequently, the global model is synchronously distributed to the source-load-storage edge control devices in all distribution substations to update the edge-side scheduling model. The updated model can accurately adapt to the random fluctuation characteristics of source-load and the coupling characteristics of multiple substations while strictly meeting multi-level physical constraints. This safety adaptive aggregation mechanism effectively avoids sudden changes in safety performance, thereby significantly improving the photovoltaic carrying capacity of the distribution network. Specific details are as follows: Step 1: Initialization. Initialize the agent network parameters on each source-load-storage edge control device, and initialize the reward function to 0.

[0085] Step 2: Global Model Download. At the start of optimization, each source-load-storage edge control device downloads the global model from the cloud-based capacity enhancement control device and sets local model parameters based on the global model, represented as follows: and .

[0086] Step 3: Action Execution. Obtain the action based on formulas (21)-(23), execute the action, and calculate the reward based on formulas (20), (25), and (26). The state is transformed based on equation (18) to .

[0087] Step 4: Edge model training and update. t During the time period, each district's intelligent agent calculates TDerror based on empirical samples, which is represented as... (27) In the formula, Discount factor and . These are the parameters for the Critic network model. Let be the value function of the Critic network. The larger the absolute value of the TD error, the greater the estimation bias during model training. Then, the parameters of the Actor network and Critic network models are updated. The Actor network is updated using gradient ascent, and the Critic network is updated using gradient descent, as shown below. (28) (29) In the formula, and The learning rates are updated for the Actor network and the Critic network, respectively.

[0088] Step 5: Edge Model Upload. After the edge model training is completed, the source-load-storage edge control device uploads the trained edge model parameters to the cloud-based capacity enhancement and control device for global model updates.

[0089] Step 6: Cloud-based Federated Aggregation. The cloud-based capacity enhancement and control device performs federated aggregation based on the received parameters from each edge scheduling model. The edge penalty term is calculated and normalized, and expressed as... (30) In the formula, and These are the normalized weights for safety overrun penalties within and between transformer substations, respectively.

[0090] Then, based on the normalized weights mentioned above, the model parameters are federated and aggregated. The aggregated global model Actor network parameters are then expressed as follows: (31) Directly adopting the hard update strategy in equation (31) will lead to a sharp drop in security performance. This is because edge model parameters with poor security performance are included in the aggregation without screening, which will cause disturbance to the global model. To address this, this invention proposes a soft update method based on security penalty weights, which integrates the new aggregation parameters with historical parameters. This method absorbs high-quality local scheduling experience while retaining reliable historical security characteristics, thereby avoiding large fluctuations in security performance and ensuring the stable and safe operation of the distribution network. The soft update expression for the global model parameters is as follows: (32) In the formula, These are parameters for the cloud-side aggregation model. If the transformer area during this time period... i The optimization performance of the agent model is worse than in the previous period, that is... If not, then no global model parameter update will be performed during this period. Critic network model parameters The update method is the same.

[0091] Step 7: Iteration and Termination. Repeat steps 2-6 until... At this point, the optimal decision for improving the photovoltaic carrying capacity of the distribution network is output.

[0092] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for improving the carrying capacity of a distribution network photovoltaic system considering multi-agent security constraints, characterized in that, Includes the following steps: S1: Construct a cloud-edge collaborative source-load-storage scheduling and multi-level security constraint framework, complete the modeling of photovoltaic, load, and energy storage characteristics and the full constraint quantification of power flow, voltage, branch, and common connection point, ensure local hard constraints on the edge side, and coordinate global coupled security in the cloud to achieve security-aware scheduling; S2: Construct a collaborative optimization problem of distribution network photovoltaic carrying capacity based on multi-objective trade-offs. By jointly optimizing photovoltaic access capacity, photovoltaic output, reduceable load power, and energy storage charging and discharging power, the distributed photovoltaic carrying capacity is maximized while minimizing source-load-storage operation costs, grid loss costs, and safety default penalty costs. S3: Proposes a distributed photovoltaic carrying capacity improvement method based on cloud-edge collaborative secure federation deep reinforcement learning, including Markov decision process modeling considering security limit violation penalties, secure federation deep actor-commentator algorithm based on cloud-edge hierarchical constraint guarantee, and cloud-edge collaborative optimization method based on secure adaptive aggregation, to achieve cloud-edge collaborative iteration and effectively improve the photovoltaic carrying capacity of distribution network.

2. The method of claim 1, wherein the method further comprises: determining a power flow of the power grid; and determining a power flow of the power grid after the power grid is connected to the power source. Step S1 includes: S1.1: Source-load-storage characteristic modeling, including distributed photovoltaic carrying capacity modeling, load reduction characteristic modeling and energy storage characteristic modeling; S1.2: Source-load-storage scheduling constraint modeling, including distributed photovoltaic power output constraints, load reduction coefficient constraints, energy storage power output constraints, power flow constraints, low-voltage distribution area common connection point power constraints and power balance constraints, node voltage deviation constraints and branch current carrying capacity constraints. 3.The method of claim 1, wherein the method further comprises: determining a power flow of the power grid; and determining a power flow of the power grid after the power grid is connected to the PV power station. In step S2, the optimization problem is constructed as follows: The constraints include: source-load-storage coordinated dispatch constraints, power flow constraints within the distribution area, power constraints and power balance constraints at the point of common coupling, node voltage deviation constraints, and branch current carrying capacity constraints. In the formula, 1) representing the substation i All photovoltaic access capacity; 2) representing the substation i All photovoltaic output; 3) representing the substation i All load power reduction; 4) representing the substation i All battery energy storage charging power; 5) representing the substation i All battery energy storage discharge power, , , , The weight coefficients of distributed photovoltaic carrying capacity, equipment operation cost, voltage out-of-limit penalty cost and grid loss cost, respectively.

4. The method for improving the photovoltaic carrying capacity of a distribution network considering multiple subject safety constraints as described in claim 1, characterized in that: Step S3 includes: S3.1: Model a Markov decision process that considers the penalty for exceeding safety limits, embedding the violation of safety constraints into the state space so that the agent can accurately perceive the violation situation during the policy optimization process; S3.2: A secure federated deep actor critic algorithm based on cloud-edge hierarchical constraint guarantee, which transforms edge-side physical constraints into quantifiable penalty terms and embeds a reward function; S3.3: A cloud-edge collaborative optimization method based on security adaptive aggregation quantifies the edge model weights by fusing security violation penalty items within and between transformer substations, and adopts a soft update strategy to fuse historical parameters.

5. The method for improving the photovoltaic carrying capacity of a distribution network considering multiple subject safety constraints as described in claim 1, characterized in that: Step S3.3 includes: Step 1: Initialize the agent network parameters on each source-load-storage edge control device; Step 2: Each source-load-storage edge control device downloads the global model from the cloud-based capacity enhancement control device; Step 3: Execute actions based on the current strategy and calculate rewards, then transition to the next time period; Step 4: Each agent in each station area calculates the temporal difference error based on the empirical samples and updates the parameters of the actor network and the critic network; Step 5: After the edge model training is completed, upload the trained edge model parameters to the cloud; Step 6: The cloud performs federated aggregation based on the received parameters of each edge model; Step 7: Repeat steps 2 to 6 until the preset time period is reached, and output the optimal distribution network photovoltaic carrying capacity improvement decision.

6. A distribution network photovoltaic carrying capacity enhancement system considering multi-entity safety constraints, used to operate the method of any one of claims 1-5, characterized in that: It includes the cloud server layer, edge server layer, and device layer: The cloud server layer includes a cloud-based capacity enhancement and control device, which is responsible for model aggregation and security coordination across distribution substations. It receives model parameters uploaded by the source-load-storage edge control devices of each substation, performs model fusion based on a security adaptive aggregation mechanism, and distributes the optimized global model to the source-load-storage edge control devices of each substation. The edge server layer includes the source-load-storage edge control device for distribution areas, which is deployed in each distribution area and undertakes the responsibility of local decision-making and hard constraint guarantee. It performs local fine scheduling of source-load-storage based on real-time data from the equipment layer and forcibly guarantees that hard constraints such as photovoltaic output, energy storage charging and discharging, and load reduction within the distribution area do not exceed the limits through the boundary pruning function. The equipment layer includes distributed photovoltaics, load shedding systems, energy storage systems, power distribution lines, transformers, and other equipment. It uploads real-time collected operating status data such as power, voltage, current, and energy storage charge status to the source-load-storage edge control device in the distribution area and strictly executes the dispatch instructions issued by the source-load-storage edge control device.

7. The distribution network photovoltaic carrying capacity enhancement system considering multi-entity safety constraints according to claim 6, characterized in that: The cloud-based capacity enhancement and control device is based on a safety adaptive aggregation mechanism. It weights and fuses models according to the safety penalty weight of the transformer area, and uses a soft update strategy to fuse historical parameters. It only updates the global model when the safety performance is improved, and coordinates the voltage, current and power coupling constraints between transformer areas. The source-load-storage edge control device of the transformer area is equipped with an intelligent agent. This intelligent agent is based on an actor-critic architecture, which ensures hard constraints through a boundary pruning function, transforms safety limit violations into penalty terms embedded in the learning process, and only uploads model parameters rather than raw data.

8. A source-load-storage edge control device for operating the system described in claim 7, characterized in that, include: The power module provides a stable power supply to the various modules in the edge control device; The communication module enables secure data interaction between the cloud-based capacity enhancement and control device and the source-load-storage edge control device, as well as between the source-load-storage edge control device and terminal equipment. The source-load-storage characteristic modeling module establishes accurate mathematical characteristic models for distributed photovoltaic, load-shaving, and energy storage systems within the transformer area. The source-load-storage scheduling constraint calculation module generates equipment-level hard constraint boundaries such as photovoltaic output, load reduction, energy storage charging and discharging, and state of charge. The power flow calculation module performs accurate power flow calculations for low-voltage distribution areas based on the power flow model, and calculates key electrical quantities such as power at each node, branch loss, voltage and current. The load-bearing capacity cost weighted difference calculation module calculates the load-bearing capacity gain and the weighted difference of comprehensive cost, transforming multi-objective optimization into a single-value reward item; A Markov decision modeling module considering security defaults embeds security default information within and between transformer areas into the state space and reward function. The boundary function-based clipping module performs double-layer boundary clipping on the original scheduling actions output by the actor network. The source load storage scheduling action generation module obtains the original scheduling action by performing Gaussian sampling based on the actor network output, and generates compliant instructions after boundary function pruning. The safety violation penalty calculation module within the transformer area calculates and sums penalty values ​​for three types of violations: power exceeding the limit at the point of common coupling, voltage deviation exceeding the limit at the node, and branch current overload. The time series error calculation module calculates the time series difference error based on the current reward, discount factor, and the value of the next state. The actor network update module uses temporal difference error as the optimization basis and adopts the gradient ascent algorithm to iteratively update the actor network parameters. The commentator network update module uses the gradient descent algorithm to minimize the squared temporal difference error and iteratively optimizes the commentator network parameters.

9. The source-load-storage edge control device for a transformer substation according to claim 8, characterized in that: The source-load-storage characteristic modeling module is connected to the source-load-storage scheduling constraint calculation module; the source-load-storage scheduling constraint calculation module is connected to the power flow calculation module and the boundary function-based pruning module; the power flow calculation module is connected to the source-load-storage scheduling constraint calculation module and the bearing capacity cost weighted difference calculation module. The boundary function-based pruning module performs two-layer pruning on scheduling actions. The first layer satisfies the hard constraints of the equipment, and the second layer adapts to the characteristics of energy storage and load, mapping infeasible actions to safe and feasible ranges. The module for calculating penalties for breach of contract within the station area is connected to the module for calculating the weighted difference of bearing capacity cost and the module for Markov decision modeling that considers breach of contract; the module for calculating time-series errors is connected to the module for updating the actor network, the module for updating the critic network, and the module for Markov decision modeling that considers breach of contract.

10. A cloud-based capacity enhancement and control device for operating the system of claim 7, characterized in that, include: The power module provides stable power to all modules in the cloud-based capacity enhancement and control device. The communication module enables secure data exchange between the cloud-based capacity enhancement and control device and the edge devices of each distribution area; The overload penalty calculation module for reverse load rate of the transformer area takes the common connection point of the transformer area as the object, calculates the power interaction between the transformer area and the upstream power grid, and calculates the forward and reverse load rates in combination with the rated capacity of the transformer. The inter-station node voltage over-limit penalty calculation module calculates the voltage deviation between ordinary nodes and aggregated interconnection nodes within the station area in a hierarchical manner, distinguishing between voltage over-limit scenarios within the station area and those coupled between station areas. The overload penalty calculation module for branch current carrying capacity between distribution substations calculates the current amplitude of branches within the distribution substation and interconnecting branches across distribution substations based on the distribution network topology and thermal stability limits. The penalty normalization module performs exponential normalization on the safety default penalty values ​​within and between transformer areas uploaded by each source-load-storage edge control device. The cloud-based federated aggregation module receives network parameters of actors and critics uploaded by edge control devices of various source load storage, and performs weighted fusion of edge models of multiple stations according to security weights; The soft update module based on security penalty weights uses the security penalty weights as coefficients to integrate the parameters of the historical global model and the current aggregated model, and decides whether to update the global model after comparing the security performance.

11. The cloud-based capacity enhancement and control device according to claim 10, characterized in that: The penalty term normalization module converts the penalty value into a weight in the 0-1 range using the Softmax function, providing a safe weighting basis for federated aggregation; The soft update module based on security penalty weights adopts a soft update strategy, which merges the new aggregated parameters with historical parameters and updates the global model only when the current model has better security performance, thus avoiding large fluctuations in security performance. The communication module is connected to the edge communication module, the penalty item normalization module, the cloud federated aggregation module, and the soft update module based on security penalty weights; the transformer area reverse load rate overload penalty calculation module, the transformer area node voltage over-limit penalty calculation module, and the transformer area branch current carrying capacity overload penalty calculation module are all connected to the penalty item normalization module, the cloud federated aggregation module, and the soft update module based on security penalty weights.

Images