Power distribution network distributed energy collaborative management system based on deep learning

Abstract

This invention relates to the field of collaborative management technology, and in particular to a deep learning-based distributed energy collaborative management system for power distribution networks. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it performs the following steps: It obtains the energy imbalance driving force by subtracting the total absorption capacity boundary of a node from the actual photovoltaic injected power; this driving force is used to generate gating weights; these weights are multiplied by the voltage amplitude deviation to construct voltage response coupling characteristics; an activation mechanism is triggered when the energy imbalance driving force is greater than 0; the voltage response coupling characteristic differences between adjacent nodes are weighted and summed using the branch impedance modulus to generate spatial propagation imbalance characteristics; finally, these characteristics are input into a deep neural network model to identify overvoltage risk levels; and power allocation commands are issued based on the adequacy of dispatchable capacity. This accurately reflects the physical causal driving source and transmission path of overvoltage, achieving precise closed-loop control of the power distribution network.

Description

Technical Field

[0001] This invention relates to the field of collaborative management technology, and in particular to a distributed energy collaborative management system for power distribution networks based on deep learning. Background Technology

[0002] Distribution network status analysis and overvoltage risk assessment are crucial for ensuring the stable operation of the power grid. In scenarios where a large number of distributed photovoltaic power sources are connected to the distribution network, the grid connection of power sources changes the traditional operating status of the power grid.

[0003] Currently, conventional methods for analyzing the status of distribution networks mainly involve directly collecting basic operational data such as real-time voltage amplitude, photovoltaic output, load power, and energy storage charge status at each node. Then, a simple numerical difference between photovoltaic and load data is calculated to reflect the degree of power surplus. Alternatively, the voltage time-series data during this period can be directly fed into a data-driven model for statistical feature extraction to assess the operational risks of the power grid.

[0004] This conventional method has a simple structure and low implementation cost. In scenarios where the number of distributed power sources connected is small and the system power flow is characterized by simple load consumption, it can better reflect the changing patterns of electrical state.

[0005] However, with the increase in photovoltaic (PV) grid integration, especially in scenarios where PV power generation is high at midday but user electricity load is low, the power flow in the grid has shifted from unidirectional load consumption to a bidirectional flow dominated by PV injection. In such complex operating scenarios, existing analysis methods rely solely on direct surface observations of voltage and power, failing to consider the conduction principles of physical circuits to trace the true causes behind data changes. This results in characteristic parameters failing to accurately reflect the source of overvoltage. Specifically, when a voltage increase or excessive power is detected at a node, existing methods cannot distinguish whether this voltage increase is directly caused by excessive PV power generation at the local node or by the passive influence of voltage anomalies from distant nodes transmitted along transmission branches. Simultaneously, they cannot accurately determine whether the current power value has exceeded the maximum absorption limit that the node itself can physically withstand. This distortion in representation prevents the grid from accurately identifying the true driving source and conduction path of overvoltage risk in scenarios with high PV penetration. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a distributed energy collaborative management system for power distribution networks based on deep learning, in order to solve the problem that existing power distribution network state analysis methods cannot accurately identify the real driving source and transmission path of overvoltage risk, resulting in the inability to achieve precise closed-loop control in high photovoltaic penetration scenarios.

[0007] This invention provides a deep learning-based distributed energy collaborative management system for power distribution networks, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it performs the following steps: The busbar and grid connection point of the distribution network that are connected to photovoltaic equipment and at least one of the following: energy storage equipment, electric vehicle charging pile equipment, basic electrical equipment and flexible electrical equipment are used as nodes. The rigid load of the linearly superimposed node, the maximum charging power of the energy storage, the dispatchable charging load, and the adjustable margin of the flexible load are used to construct the total absorption capacity boundary describing the maximum absorption capacity of the node at the current moment. The energy imbalance driving quantity reflecting the degree of power overflow is calculated by subtracting the actual photovoltaic injected power from the total absorption capacity boundary. The voltage amplitude deviation of the node relative to the rated voltage is calculated, and an energy imbalance gating weight is generated based on the energy imbalance driving quantity. The energy imbalance gating weight is multiplied by the voltage amplitude deviation to construct a voltage response coupling feature that indicates whether the voltage anomaly is directly caused by the local power imbalance. When the energy imbalance driving force of a node is greater than 0, the activation mechanism is triggered. Based on the branch impedance magnitude between the node and its neighboring nodes, the difference in voltage response coupling characteristics between the node and its neighboring nodes is weighted and summed to generate spatial propagation imbalance characteristics. The spatial propagation imbalance characteristics of all nodes are input into a preset deep neural network model to identify and output the current overvoltage risk level of each node. For nodes with high risk levels, a collaborative peak-shaving power allocation instruction is generated based on the adequacy of local schedulable capacity and sent to the hardware execution terminal to perform closed-loop control.

[0008] Preferably, the generation energy imbalance gating weights include: The rated capacity parameter of the distribution transformer connected to the node is multiplied by a fixed proportional coefficient to determine the capacity conversion benchmark value of the node; the energy imbalance driving quantity is divided by the capacity conversion benchmark value of the node, and the energy imbalance gating weight is calculated by using the hyperbolic tangent function.

[0009] Preferably, the formula for calculating the voltage response coupling characteristic is: ; In the formula, The voltage response coupling characteristics of the node; This refers to the voltage amplitude deviation at the node; The rated voltage of the node; It is the hyperbolic tangent function; The driving force for energy imbalance at the node; The baseline value for calculating the capacity of a node; The energy imbalance gating weight for the node.

[0010] Preferably, the activation mechanism is triggered when the energy imbalance driving force of the determination node is greater than 0, including: If the value of the energy imbalance driving force of a node is greater than 0, the value of the source node hard threshold activation function of the node is set to 1. If the value of the energy imbalance driving force of a node is less than or equal to 0, the value of the source node hard threshold activation function of the node is set to 0.

[0011] Preferably, the formula for calculating the spatial propagation imbalance characteristic is: ; In the formula, for Spatial propagation unevenness characteristics; For the purpose of and An operator that sums the values ​​of all directly connected adjacent nodes in the set of all adjacent nodes; The sequence number of the adjacent nodes directly connected to the node; To and The set of all directly connected adjacent nodes; for With the The branch impedance magnitude between adjacent nodes; for With the The electrical coupling strength weight between adjacent nodes; for Voltage response coupling characteristics; For the first Voltage response coupling characteristics of adjacent nodes; for The voltage response coupling characteristics and the first The difference in voltage response coupling characteristics of adjacent nodes; for The source node hard threshold activation function.

[0012] Preferably, the data collected includes the node's basic rigid load, maximum energy storage charging power, dispatchable charging load and flexible load adjustment margin, actual photovoltaic injection power, and voltage amplitude, including: Extract the amplitude component of the AC voltage signal monitored by the synchronous phasor measuring device in the polar coordinate system as the voltage amplitude of the node; collect the actual active power injected into the distribution network by the photovoltaic equipment at the current moment as the actual photovoltaic injected power of the node; Extract the power consumed by the basic electrical equipment currently in operation and in an uninterruptible state within the node as the node's basic rigid load; read in real time the maximum active power limit that the energy storage device can absorb at the current moment as the node's maximum energy storage charging power; retrieve the maximum power absorbed by various electric vehicle charging pile devices connected to the node at the current moment as the node's dispatchable charging load; extract the upper limit of active power consumption that flexible electrical equipment such as temperature-controlled loads and industrial transferable processes can instantaneously increase at the current operating point as the node's flexible load adjustment margin.

[0013] Preferably, the step of generating collaborative peak-shaving power allocation instructions based on the sufficiency of locally schedulable capacity includes: The maximum energy storage charging power, dispatchable charging load, and flexible load adjustment margin of high-risk nodes and their adjacent nodes are summed to obtain the dispatchable capacity adequacy of each node, and a resource dispatchable list is generated in descending order of dispatchable capacity adequacy. The larger of the energy imbalance driving force of high-risk level nodes and 0 is taken as the power to be allocated. According to the resource schedulable list, the smaller of the current remaining power to be allocated and the current node's schedulable capacity sufficiency is taken as the power to be allocated to the current node. The power to be allocated to the current node is deducted from the current remaining power to be allocated until the remaining power to be allocated is 0 or the resource schedulable list is traversed to the end. Generate a collaborative peak shaving power allocation instruction that includes at least an instruction identifier, generation time, target node address, target device address, resource type, allocated power, start execution time, effective duration, and checksum, and send the collaborative peak shaving power allocation instruction to the energy storage controller, charging pile group control terminal, or flexible load controller of the corresponding target node.

[0014] Preferably, after the step of sending the data to the hardware execution terminal to perform closed-loop control, the method further includes: In the next data sampling cycle after the hardware execution terminal responds to the collaborative peak-shaving power allocation instruction, the multi-source data acquisition is re-triggered and the spatial propagation imbalance characteristics of the node are calculated again in real time. The trained deep neural network model is used to perform online verification of the current overvoltage risk level of the updated nodes; If the current overvoltage risk level of a node fails to fall back to the level of concern or normal, a risk warning signal containing a prompt to expand the scope of collaborative management will be issued and uploaded to the dispatch center.

[0015] Preferably, the preset deep neural network model adopts a multilayer perceptron architecture, including having The system consists of an input layer with 10 neurons, three hidden layers with 256, 128, and 64 neurons respectively, and a layer with 3... The output layer has 10 neurons, and the three hidden layers all use linear rectified activation functions. The output layer is divided according to the node order in the topology routing matrix. There are three output groups, each containing three neurons. A normalized exponent operation is performed independently on each output group to obtain the probability that the corresponding node belongs to the normal level, the attention level, or the high-risk level. This represents the number of valid nodes in the topology routing matrix.

[0016] The beneficial effects of the embodiments of the present invention compared with the prior art are as follows: In this invention, a clear physical absorption limit is constructed by superimposing the adjustment limits of basic power consumption, energy storage, and flexible regulation resources of nodes. The difference between this limit and the actual photovoltaic injection power can accurately calculate whether the power exceeds the node's tolerance limit, avoiding misjudgments caused by conventional surface numerical comparisons. By combining the power overflow situation with the degree of voltage deviation to generate gating weights, it is possible to distinguish whether the voltage rise is directly caused by local photovoltaic power generation surplus or by the passive influence of abnormal conduction at a remote location, thereby effectively eliminating interference from the external environment. By introducing network branch impedance and source node activation mechanisms, the single-point state is extended to the entire network, accurately capturing the real propagation trajectory of excess power spreading outward along low-impedance physical channels. Finally, relying on a deep learning model to identify risk levels, and sorting according to the actual ability of each node to absorb excess power, nodes with strong absorption capacity are given priority to execute control commands, realizing precise allocation and closed-loop control of distribution network resources, and significantly improving the safety of grid operation in high photovoltaic access scenarios. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for collaborative management of distributed energy resources in a power distribution network based on deep learning, provided in Embodiment 1 of the present invention. Detailed Implementation

[0019] The embodiments of this disclosure are described in detail below. The embodiments described below are exemplary and intended to explain this disclosure, and should not be construed as limiting this disclosure.

[0020] It should be noted that the terms "first," "second," etc., used in this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The implementations described in the following exemplary embodiments do not represent all implementations consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure.

[0021] To illustrate the technical solution of the present invention, specific embodiments are described below.

[0022] This invention provides a deep learning-based distributed energy collaborative management system for power distribution networks, including a processor and a memory. The processor executes a computer program stored in the memory to implement a deep learning-based distributed energy collaborative management method for power distribution networks. The method includes the following steps:

[0023] Step S101: Obtain multi-source measurement data and network topology parameters, and perform data cleaning and standard alignment processing.

[0024] It should be noted that with the increase in distributed photovoltaic (PV) grid connections, the fluctuating nature of PV power causes frequent changes in power flow within the grid. Existing technologies directly utilize basic voltage and power measurement data for statistical analysis, neglecting the actual physical circuit principles. This leads to an inability to accurately distinguish whether a voltage rise is directly caused by excessive local PV power injection or by voltage anomalies at distant nodes propagated through branch impedance. Furthermore, considering the difficulty of comprehensively installing detection sensors on the low-voltage side of the distribution network in practical industrial applications, this invention performs standardized processing and time alignment on the acquired multi-source measurement data and the electrical topology of the distribution network at the front end to accurately determine the true source of overvoltage risk and its propagation path in the network topology. This ensures data integrity and compliance with electrical constraints, thereby providing accurate basic data for subsequent calculations of characteristic parameters reflecting physical causal relationships.

[0025] Specifically, the busbars and grid connection points of the distribution network that are connected to photovoltaic equipment and simultaneously connected to at least one of the following: energy storage equipment, electric vehicle charging pile equipment, basic electrical equipment, and flexible electrical equipment, are designated as nodes. The voltage amplitude, basic rigid load, maximum energy storage charging power, dispatchable charging load, flexible load adjustment margin, and actual photovoltaic power injection of the nodes are synchronously collected through a Supervisory Control and Data Acquisition (SCADA) system, a phasor measurement unit (PAU), and user-side smart meters. The data sampling period is empirically set to 1 minute by the implementers. The specific data collection method is as follows: (1) Extract the amplitude component of the AC voltage signal monitored by the synchronous phasor measuring device in the polar coordinate system as the voltage amplitude of the node, which is used to reflect the actual voltage level at the node bus at the current moment.

[0026] (2) Collect the actual active power injected into the distribution network by the photovoltaic equipment at the current moment, use it as the actual photovoltaic power injected into the node, and write it in the form of a positive value.

[0027] (3) Extract the power consumed by the basic electrical equipment that is currently running and cannot be interrupted inside the node, and use it as the basic rigid load of the node. When the node does not contain basic electrical equipment, the basic rigid load of the node is assigned a value of 0, and the basic rigid load of the node is negative. In the calculation of the total absorption capacity boundary of the subsequent node, the absolute value is taken and written as a positive value.

[0028] (4) Read the maximum active power that the energy storage device can absorb at the current moment in real time, use it as the maximum charging power of the node's energy storage, and write it in positive form. When the node does not contain an energy storage device, the maximum charging power of the node's energy storage is assigned a value of 0.

[0029] (5) Retrieve the maximum absorption power of various electric vehicle charging pile devices at the current moment of the access node as the schedulable charging load of the node. When the node does not contain electric vehicle charging pile devices, the schedulable charging load of the node is assigned a value of 0.

[0030] (6) Extract the upper limit of active power consumption that can be instantaneously increased at the current operating point for flexible electrical equipment such as temperature control load and industrial transferable processes, and use it as the flexible load adjustment margin of the node. When the node does not contain flexible electrical equipment, the flexible load adjustment margin of the node is assigned a value of 0.

[0031] Furthermore, a data cleaning window is constructed using 60 consecutive data sampling periods. The ratio of the number of missing sampling points within the data cleaning window to 60 is used as the missing rate. When the missing rate is below 15% and any consecutive missing segment contains no more than 10 sampling points, linear interpolation is performed using the nearest normal sampling values ​​before and after the missing segment. When the missing rate reaches 15% or any consecutive missing segment contains more than 10 sampling points, the corresponding data cleaning window is removed. For each sample value to be judged, five normal sampling values ​​before and after it are extracted, and the local mean and local standard deviation are calculated. When the absolute value of the difference between the sample value to be judged and the local mean exceeds three times the local standard deviation, the sample value to be judged is marked... Data is recorded as abnormal and replaced by the algebraic mean of the nearest normal sampled values. If 10 normal sampled values ​​cannot be obtained, the corresponding data cleaning window is removed. 15%, three times the local standard deviation, and 10 sampling points are determined through historical complete data occlusion verification. Specifically, simulated missing data and abnormal perturbations are applied to the candidate missing rate threshold, standard deviation multiple, and continuous missing length, respectively. The selection conditions are that the interpolation root mean square error does not exceed the allowable error of the measurement device and the abnormal data deletion rate is the lowest. In this embodiment, the corresponding selection results are 15%, 3, and 10, respectively. When the data sampling period or the measurement device model changes, the occlusion verification is re-executed and the corresponding parameters are updated.

[0032] The adjacency relationships, branch resistances, and branch reactances between nodes in the current operating mode of the distribution network are derived from the combined energy management system and geographic information system. The system baseline capacity is assumed to be... ,node With nodes The reference voltage for the voltage level of the branch is The corresponding reference impedance is The branch resistance expressed in ohms and branch circuit reactor Divide by the reference impedance respectively Obtain the per-unit values ​​of branch resistance and branch reactance, and take the square root of the sum of their squares as the dimensionless branch impedance magnitude. When the original branch parameters use per-unit values ​​under other references, they are first restored to ohmic values ​​according to the original reference capacity and voltage, and then converted according to the aforementioned system reference. Only branches with a branch impedance magnitude greater than 0 and complete resistance, reactance, and reference parameters are included in the adjacency relationship. Branches with a branch impedance magnitude of 0 or missing parameters are marked as abnormal branches and are stopped from participating in the calculation of spatial propagation imbalance characteristics. Let the number of nodes that pass the verification be... Build according to a fixed node order The topology routing matrix is ​​set such that the diagonal elements and elements without direct branches are set to 0, and the elements with valid branches are set to the corresponding branch impedance magnitudes. The matrix is ​​then set to the first element... The nodes corresponding to the elements with values ​​greater than 0 in a row constitute a node. The set of adjacent nodes ensures that the branch impedance magnitudes in the calculation of spatial propagation imbalance characteristics are all positive.

[0033] Step S102: Construct the total absorption capacity boundary by superimposing each absorption resource, and calculate the energy imbalance driving quantity that characterizes the degree of power overflow.

[0034] Existing technologies, when quantitatively assessing the grid integration level of distributed photovoltaic (PV) systems, typically limit themselves to algebraically subtracting the actual PV output from the local load level. This conventional quantitative relationship fails to consider the equipment-level adjustment limits of the various active and flexible controllable resources connected to the node over a dynamic time axis, making it impossible to accurately determine whether the power surplus has exceeded the physical constraints of the node's grid integration. To provide a physical premise with clear constraints for subsequent causal attribution analysis, this invention superimposes all independent and electrically parallel grid integration resources within the node according to their respective grid connection constraints, constructing a physical boundary that describes the node's maximum absorption capacity at the current moment. Based on this boundary, the energy imbalance driving force of the node is calculated, serving as the core basis for determining whether the grid integration capacity has reached its upper limit.

[0035] Based on Kirchhoff's current law, a linear summation is performed on the node's basic rigid load, maximum energy storage charging power, dispatchable charging load, and flexible load adjustment margin on the node's grid-connected bus. The total absorption capacity boundary of the node is constructed using the superposition results of these various absorption resources. The formula for calculating the total absorption capacity boundary of the node is as follows: ; In the formula, This represents the boundary of the node's total absorption capacity. For the basic rigid load of the node; Indicates taking the absolute value; The maximum charging power of the node's energy storage; For the schedulable charging load of the node; The flexibility of the node's load can be adjusted upwards.

[0036] The total absorption capacity boundary of a node essentially represents the absolute limit of active power absorption capacity that the node possesses at the current moment. This limit capacity not only covers the fixed power consumption necessary for the uninterrupted basic electrical equipment within the node to maintain normal operation, but also fully quantifies the maximum additional power absorption potential that all controllable and adjustable devices connected to the node can provide under the current operating state. By calculating the total absorption capacity boundary of the node, this invention provides a baseline for subsequently judging whether the active power generated by photovoltaic equipment will experience rigid overflow locally. This enables the system to accurately identify extreme states where even if all local resources are mobilized, excess power cannot be completely absorbed locally, thus laying the foundation for the physical source tracing of overvoltage risks.

[0037] Finally, an algebraic difference is performed between the actual photovoltaic injected power of the synchronously acquired nodes and the calculated total absorption capacity boundary of the nodes to calculate the energy imbalance driving force of the nodes, which reflects the degree of power overflow. The formula for calculating the energy imbalance driving force of the nodes is as follows: ; In the formula, The driving force for energy imbalance at the node; The actual photovoltaic power injected into the node; This represents the boundary of the node's total absorption capacity.

[0038] The above calculation formula achieves dynamic quantitative capture of the physical absorption capacity of the node by algebraically summing the active power absorption forms of each independent port of the node's grid-connected bus. This results in a positive increasing trend in the energy imbalance driving force of the node caused by either an increase in the actual photovoltaic power injected into the node or a decrease in the maximum charging power of the node's energy storage.

[0039] When the value of the energy imbalance driving force of a node is a real number greater than 0, it indicates that the actual photovoltaic power injected into the node has exceeded the node's total absorption capacity boundary. At this time, even if all the local controllable resources connected to the node are operating at full capacity, there is still a residual power flow that cannot be absorbed locally within the node, and the system is thus determined to enter a constrained imbalance state. When the value of the energy imbalance driving force of a node is less than or equal to 0, it indicates that the actual photovoltaic power injected into the node is still within the node's total absorption capacity boundary, and the node still has sufficient adjustment flexibility to respond to the current energy fluctuations.

[0040] Step S103: Combine voltage amplitude deviation and energy imbalance driving quantity to construct gating weights and calculate voltage response coupling characteristics.

[0041] Considering that the nodes in a distribution network are physically connected to each other through branch impedance, when a voltage rise is observed at a node, this phenomenon may be directly caused by the node's own energy imbalance driving force being greater than zero, or it may be a passive influence caused by severe power overflow from adjacent nodes being conducted along the branch impedance, or it may even be caused by voltage fluctuations in the main distribution network itself. Therefore, in order to accurately determine the true physical cause of the voltage rise, this invention combines the node's energy imbalance driving force, which reflects the degree of power overflow, with the node's voltage amplitude deviation, which reflects the voltage change. A specific mathematical function is used to construct a gating weight, and the sign of the node's voltage amplitude deviation is determined based on the adequacy of the node's current absorption capacity. In this way, the single power over-limit judgment is transformed into a characteristic parameter that can clearly indicate whether the voltage anomaly is directly caused by local power imbalance.

[0042] Specifically, the voltage amplitude of the node collected at the current moment is retrieved, and the voltage amplitude deviation of the node is calculated by subtracting the rated voltage of the node from the voltage amplitude of the node; the rated capacity parameters of the distribution transformers connected to the node are retrieved from the pre-stored system ledger, and the rated capacity parameters of the distribution transformers connected to the node are multiplied by a fixed proportional coefficient to determine the capacity conversion benchmark value of the node; wherein, the fixed proportional coefficient is calibrated to one-third by the implementers based on the historical operating data of the distribution transformer, so as to map the energy imbalance driving quantity of the node to the approximate linear interval of the subsequent hyperbolic tangent function.

[0043] Furthermore, the calculated energy imbalance driving quantity of the node is divided by the node's capacity conversion reference value, and the hyperbolic tangent function is used for calculation to obtain the energy imbalance gating weight of the node in the [-1,1] interval; the voltage amplitude deviation of the node is divided by the node's rated voltage, and multiplied by the node's energy imbalance gating weight to calculate the node's voltage response coupling characteristics, wherein the formula for calculating the node's voltage response coupling characteristics is: ; In the formula, The voltage response coupling characteristics of the node; This refers to the voltage amplitude deviation at the node; The rated voltage of the node; It is the hyperbolic tangent function; The driving force for energy imbalance at the node; This serves as the baseline value for calculating the node's capacity.

[0044] The above calculation formula constructs an asymmetric bidirectional gating response law by introducing the hyperbolic tangent function. This means that an increase in the energy imbalance driving force of a node or an amplification of the voltage amplitude deviation of a node will cause the voltage response coupling characteristic of the node to increase monotonically. When a node has a rigid surplus, since the energy imbalance driving force of the node is positive, the energy imbalance gating weight of the node calculated by the hyperbolic tangent function approaches 1. At this time, if the voltage amplitude deviation of the node is positive, the product term of the two will be positively amplified and retained, thus outputting a voltage response coupling characteristic of the node greater than 0. This physically determines that there is a direct causal relationship between the overvoltage offset phenomenon and the local power imbalance. When the node has sufficient absorption elasticity, the energy imbalance driving force of the node is negative, and the value of the hyperbolic tangent function turns negative, causing the algebraic sign of the voltage amplitude deviation of the node to reverse after the product operation, thus outputting a negative value opposite to the aforementioned state direction, achieving effective elimination of local passively conducted overvoltage interference.

[0045] Step S104: Introduce the source node hard threshold activation mechanism and electrical coupling strength weight to calculate the spatial propagation imbalance characteristics that characterize the spatial disturbance propagation strength of the network.

[0046] The voltage response coupling characteristic of a node is essentially an isolated characteristic quantity that depends on the local measurement of the node. When facing the scenario of large-scale grid access of distributed photovoltaics, the characteristics of a single node cannot effectively capture the dynamic propagation path of electrical disturbances spreading outward along the network topology, which can easily lead to false forwarding disturbances in passive nodes during feature calculation. In order to lock the core source path that drives the overvoltage risk to propagate outward at the spatial topology level, this invention extends the isolated node-level causal score to the network topology space. By designing a source node hard threshold activation mechanism to remove interference data without causal correlation, and introducing physical coupling weights based on the real branch impedance, a spatial propagation imbalance characteristic of the node that can quantitatively characterize the intensity of spatial disturbance propagation in the network is constructed.

[0047] For any node Real-time monitoring of the calculated nodes Energy imbalance driving force And perform conditional threshold determination: if the determination node Energy imbalance driving force If the value is greater than 0, the activation mechanism is triggered and the node is activated. The source node's hard threshold activation function is set to 1; if the decision node... Energy imbalance driving force If the value is less than or equal to 0, then the node will be... The value of the source node hard threshold activation function is set to 0.

[0048] Locate nodes by retrieving a pre-defined topology routing matrix. The set of all directly connected adjacent nodes Traversal and nodes For each directly connected neighboring node in the set, retrieve the node calculated at the current time. Voltage response coupling characteristics and retrieve the first Voltage response coupling characteristics of adjacent nodes , will node voltage response coupling characteristics minus the first The voltage response coupling characteristics of adjacent nodes are used to obtain the node. The voltage response coupling characteristics and the first The difference in voltage response coupling characteristics of adjacent nodes and for nodes The voltage response coupling characteristics and the first The difference in voltage response coupling characteristics of adjacent nodes is squared to calculate the node. The voltage response coupling characteristics and the first The square of the difference in voltage response coupling characteristics of adjacent nodes .

[0049] Furthermore, nodes are retrieved from the network parameter ledger. With the The branch impedance magnitude between adjacent nodes , will node With the The reciprocal of the branch impedance magnitude between adjacent nodes is used to calculate the electrical coupling strength weight, which characterizes the tightness of the grid structure between the two nodes.

[0050] Finally, the calculated electrical coupling strength weights and node... The voltage response coupling characteristics and the first The square of the difference in voltage response coupling characteristics of adjacent nodes and the node The source node hard threshold activation function is continuously multiplied, and then multiplied with the node... The node is generated by summing the product results of all directly connected adjacent nodes in the set of all adjacent nodes. The spatial propagation imbalance characteristic, wherein the node The formula for calculating the spatial propagation imbalance characteristics is: ; In the formula, For nodes Spatial propagation unevenness characteristics; For nodes An operator that sums the values ​​of all directly connected adjacent nodes in the set of all adjacent nodes; For nodes The sequence number of the directly connected adjacent nodes; For nodes The set of all directly connected adjacent nodes; For nodes With the The branch impedance magnitude between adjacent nodes; For nodes With the The electrical coupling strength weight between adjacent nodes; For nodes Voltage response coupling characteristics; For the first Voltage response coupling characteristics of adjacent nodes; For nodes The voltage response coupling characteristics and the first The square of the difference in voltage response coupling characteristics between adjacent nodes; For nodes The source node hard threshold activation function.

[0051] The above calculation formula is passed through nodes. The control effect of the source node hard threshold activation function removes the paths for passive nodes with energy imbalance driving forces less than or equal to 0 to participate in network space perturbation calculations, thus enabling the nodes to... With the The decrease in the branch impedance magnitude between adjacent nodes or the node The voltage response coupling characteristics and the first The amplification of the difference in voltage response coupling characteristics of adjacent nodes will drive the node. The spatial propagation imbalance characteristics show a significant increase in numerical values; at the spatial topology level, branches with smaller branch impedance moduli indicate closer physical electrical distances and stronger electrical coupling. When the two end nodes generate a large difference in voltage response coupling characteristics due to differences in local photovoltaic sufficiency, the nodes... The voltage response coupling characteristics and the first The square of the difference in voltage response coupling characteristics of adjacent nodes, amplified by the inverse weight, causes a step in the accumulated result, thus enabling precise capture of the evolution of the energy imbalance driving force of the node to transmit disturbances outward along a specific low-impedance electrical path.

[0052] Step S105: Use a deep neural network model to assess the overvoltage risk level and execute coordinated peak-shaving closed-loop control based on schedulable capacity adequacy.

[0053] The spatial propagation imbalance characteristics of nodes obtained from the aforementioned calculations fully reflect the actual evolution process of the node's energy imbalance driving force exceeding the absorption limit, the node's voltage amplitude deviation increasing in the same direction due to the energy imbalance driving force, and the overvoltage risk spreading to the outside of the distribution network along the low-impedance branch. This solves the defects of traditional data processing models, such as unclear physical mechanisms and easy over-reliance on historical data. In order to apply the calculated spatial propagation imbalance characteristics of nodes to the real-time control of the distribution network, this invention uses them as input parameters of a deep neural network model and establishes a control process that includes overvoltage risk level identification, allocation of available absorption resources, and control result feedback verification.

[0054] During the offline training phase, nodes are arranged according to a fixed order in the topology routing matrix. Each valid node represents a historical sampling time. The spatial propagation imbalance characteristics of each node at that moment are determined according to... Arranged in order of length The input feature vector will be continuous The input feature vectors at each historical sampling time are arranged in chronological order as follows: The training input matrix is ​​obtained; the voltage thresholds and overvoltage action thresholds corresponding to the voltage levels of each node are extracted from the power grid dispatch system. A normal level label is generated when the node voltage amplitude is not greater than the voltage threshold; a high-risk level label is generated when the node voltage amplitude is greater than the voltage threshold but not greater than the overvoltage action threshold; and a high-risk level label is generated when the node voltage amplitude is greater than the overvoltage action threshold. Each level label is converted into a one-hot code containing three elements, thus constructing a dimension-based system. The training label matrix; each input feature vector corresponds one-to-one with the node level label at the same sampling time, and the training dataset is divided into training subset, validation subset and test subset according to the time sequence.

[0055] The preset deep neural network model adopts a multilayer perceptron architecture, and the input layer includes... The three hidden layers contain 256, 128, and 64 neurons respectively, all using a linear rectified activation function. The output layer contains... The neurons are divided according to their node order. There are output groups, with nodes set. The corresponding output group The output value of each neuron is Then the node belongs to the first... The probability of each risk level is ,in and All values ​​are risk level numbers and can be 1, 2, or 3, thus clarifying that the normalization exponent operation is only performed between the three output values ​​corresponding to the same node.

[0056] During training, the average multi-class cross-entropy of all training samples and all nodes is used as the loss function, and the formula for calculating the loss function is: ; In the formula, The average cross-entropy loss for multi-class classification; This represents the number of training samples; The number of valid nodes; The training sample number; The valid node sequence number; This is the risk level number, with a value of 1, 2, or 3, corresponding to normal level, attention level, and high-risk level, respectively. For the first Nodes in training samples Corresponding to the Each risk level has a uniquely encoded label, and the node... Belongs to the The value is 1 when the risk level is specified, and 0 otherwise. Nodes output by a deep neural network model Belongs to the The probability of each risk level; This is a natural logarithm operation; the cross-entropy of all risk levels, all valid nodes, and all training samples is summed and divided by . The average loss is obtained to measure the overall deviation between the model's output risk probability and the risk level label.

[0057] An adaptive moment estimation optimization algorithm is adopted. The model parameters are updated according to the gradient of the average multi-class cross-entropy loss relative to the network weights and biases. The initial learning rate is set to 0.001, the batch size is set to 64, and the maximum number of iterations is set to 200. After each iteration, the average multi-class cross-entropy loss is calculated using the validation subset and the corresponding model parameters are saved. After 200 iterations, the model parameters saved when the validation subset loss is the lowest are used as the trained deep learning risk assessment model.

[0058] Furthermore, during the online monitoring and operation of the distribution network, the current time is recorded according to the fixed node sequence used during model training. The spatial propagation imbalance characteristics of the effective nodes are arranged in a length of... The online input feature vector is then input into the trained deep learning risk assessment model, and the model output dimension is... The risk probability matrix is ​​such that each row corresponds to a node and the three matrix elements represent the probabilities of normal level, attention level and high risk level respectively. The level corresponding to the highest probability in each row is taken as the current overvoltage risk level of the node. When the number of effective nodes or the node order changes, the original deep learning risk assessment model is stopped and the training input matrix and training label matrix are reconstructed according to the changed topology routing matrix.

[0059] Furthermore, for nodes identified as high-risk and their directly connected adjacent nodes, the collaborative management system immediately retrieves and collects the maximum energy storage charging power, the node's dispatchable charging load, and the node's flexible load adjustment margin of the matched node. It then performs an algebraic summation operation on the maximum energy storage charging power, dispatchable charging load, and flexible load adjustment margin of the node identified as high-risk to calculate the dispatchable capacity adequacy of the node. Simultaneously, it performs an algebraic summation operation on the maximum energy storage charging power, dispatchable charging load, and flexible load adjustment margin of the adjacent nodes to calculate the dispatchable capacity adequacy of the adjacent nodes.

[0060] For nodes identified as high-risk, the larger of their energy imbalance driving force and 0 is used as the power to be allocated. A resource schedulable list is generated for high-risk nodes and their adjacent nodes in descending order of schedulable capacity adequacy. The initial remaining power to be allocated is set to equal the power to be allocated. The resource schedulable list is traversed sequentially, and the smaller of the current remaining power to be allocated and the current node's schedulable capacity adequacy is used as the power to be allocated for the current node. The power to be allocated for the current node is deducted from the current remaining power to be allocated. The traversal stops when the remaining power to be allocated is 0. If the remaining power to be allocated is still greater than 0 after the resource schedulable list has been traversed, the remaining power to be allocated is written into the expanded collaborative management scope prompt message and uploaded to the scheduling center.

[0061] For nodes that receive allocated power, the node's allocated power is decomposed according to the proportion of the maximum energy storage charging power, the dispatchable charging load, and the flexible load adjustment margin in the node's dispatchable capacity adequacy, thereby obtaining the energy storage charging power allocation, the charging pile power allocation, and the flexible load power allocation respectively. When the power upper limit corresponding to a certain resource is 0, the power allocation of that resource is set to 0. When the node's dispatchable capacity adequacy is 0, the node is skipped, thereby ensuring that the denominator in the power decomposition process is greater than 0.

[0062] The collaborative peak shaving power allocation instruction includes an instruction identifier, generation time, target node address, target device address, resource type, target power allocation amount, start execution time, effective duration, and checksum. It determines the target device address and control register based on a pre-stored node device mapping table. Specifically, the energy storage charging power allocation amount is mapped to the incremental active power of the energy storage controller, the charging pile power allocation amount is mapped to the incremental upper limit of the group charging power of the charging pile group control terminal, and the flexible load power allocation amount is mapped to the target incremental active power consumption of the flexible load controller. After receiving and verifying the collaborative peak shaving power allocation instruction, each hardware execution terminal limits the target power allocation amount according to the device's rated power, ramp rate, and current operating status, and returns execution confirmation information.

[0063] In the next one-minute data sampling cycle after the hardware execution terminal returns the execution confirmation information, multi-source data acquisition is re-triggered and the spatial propagation imbalance characteristics of the node are calculated. The trained deep learning risk assessment model then re-outputs the current overvoltage risk level of the node. When the current overvoltage risk level of the node does not fall back to the attention level or normal level, a risk warning message containing unabsorbed power, the address of the unresponsive device, and an expanded collaborative management scope identifier is generated and uploaded to the dispatch center.

[0064] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A deep learning-based distributed energy collaborative management system for power distribution networks, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it performs the following steps: The busbar and grid connection point of the distribution network that are connected to photovoltaic equipment and at least one of the following: energy storage equipment, electric vehicle charging pile equipment, basic electrical equipment and flexible electrical equipment are used as nodes. The rigid load of the linearly superimposed node, the maximum charging power of the energy storage, the dispatchable charging load, and the adjustable margin of the flexible load are used to construct the total absorption capacity boundary describing the maximum absorption capacity of the node at the current moment. The energy imbalance driving quantity reflecting the degree of power overflow is calculated by subtracting the actual photovoltaic injected power from the total absorption capacity boundary. The voltage amplitude deviation of the node relative to the rated voltage is calculated, and an energy imbalance gating weight is generated based on the energy imbalance driving quantity. The energy imbalance gating weight is multiplied by the voltage amplitude deviation to construct a voltage response coupling feature that indicates whether the voltage anomaly is directly caused by the local power imbalance. When the energy imbalance driving force of a node is greater than 0, the activation mechanism is triggered. Based on the branch impedance magnitude between the node and its neighboring nodes, the difference in voltage response coupling characteristics between the node and its neighboring nodes is weighted and summed to generate spatial propagation imbalance characteristics. The spatial propagation imbalance characteristics of all nodes are input into a preset deep neural network model to identify and output the current overvoltage risk level of each node. For nodes with high risk levels, a collaborative peak-shaving power allocation instruction is generated based on the adequacy of local schedulable capacity and sent to the hardware execution terminal to perform closed-loop control.

2. The deep learning-based distributed energy collaborative management system for power distribution networks according to claim 1, characterized in that, The generation energy imbalance gating weights include: The rated capacity parameter of the distribution transformer connected to the node is multiplied by a fixed proportional coefficient to determine the capacity conversion benchmark value of the node; the energy imbalance driving quantity is divided by the capacity conversion benchmark value of the node, and the energy imbalance gating weight is calculated by using the hyperbolic tangent function.

3. The deep learning-based distributed energy collaborative management system for power distribution networks according to claim 1, characterized in that, The formula for calculating the voltage response coupling characteristics is: ； In the formula, The voltage response coupling characteristics of the node; This refers to the voltage amplitude deviation at the node; The rated voltage of the node; It is the hyperbolic tangent function; The driving force for energy imbalance at the node; The baseline value for calculating the capacity of a node; The energy imbalance gating weight for the node.

4. The deep learning-based distributed energy collaborative management system for power distribution networks according to claim 1, characterized in that, The activation mechanism is triggered when the energy imbalance driving force of the determination node is greater than 0, including: If the value of the energy imbalance driving force of a node is greater than 0, the value of the source node hard threshold activation function of the node is set to 1. If the value of the energy imbalance driving force of a node is less than or equal to 0, the value of the source node hard threshold activation function of the node is set to 0.

5. The deep learning-based distributed energy collaborative management system for power distribution networks according to claim 4, characterized in that, The formula for calculating the spatial propagation imbalance feature is: ； In the formula, for Spatial propagation unevenness characteristics; For the purpose of and An operator that sums the values ​​of all directly connected adjacent nodes in the set of all adjacent nodes; The sequence number of the adjacent nodes directly connected to the node; To and The set of all directly connected adjacent nodes; for With the The branch impedance magnitude between adjacent nodes; for With the The electrical coupling strength weight between adjacent nodes; for Voltage response coupling characteristics; For the first Voltage response coupling characteristics of adjacent nodes; for The voltage response coupling characteristics and the first The difference in voltage response coupling characteristics of adjacent nodes; for The source node hard threshold activation function.

6. The deep learning-based distributed energy collaborative management system for power distribution networks according to claim 1, characterized in that, The data collected includes the node's basic rigid load, maximum energy storage charging power, dispatchable charging load and flexible load adjustment margin, actual photovoltaic injection power, and voltage amplitude. Extract the amplitude component of the AC voltage signal monitored by the synchronous phasor measuring device in the polar coordinate system as the voltage amplitude of the node; collect the actual active power injected into the distribution network by the photovoltaic equipment at the current moment as the actual photovoltaic injected power of the node; Extract the power consumed by the basic electrical equipment currently in operation and in an uninterruptible state within the node as the node's basic rigid load; read in real time the maximum active power limit that the energy storage device can absorb at the current moment as the node's maximum energy storage charging power; retrieve the maximum power absorbed by various electric vehicle charging pile devices connected to the node at the current moment as the node's dispatchable charging load; extract the upper limit of active power consumption that flexible electrical equipment such as temperature-controlled loads and industrial transferable processes can instantaneously increase at the current operating point as the node's flexible load adjustment margin.

7. The deep learning-based distributed energy collaborative management system for power distribution networks according to claim 1, characterized in that, The generation of collaborative peak-shaving power allocation instructions based on the sufficiency of locally schedulable capacity includes: The maximum energy storage charging power, dispatchable charging load, and flexible load adjustment margin of high-risk nodes and their adjacent nodes are summed to obtain the dispatchable capacity adequacy of each node, and a resource dispatchable list is generated in descending order of dispatchable capacity adequacy. The larger of the energy imbalance driving force of high-risk level nodes and 0 is taken as the power to be allocated. According to the resource schedulable list, the smaller of the current remaining power to be allocated and the current node's schedulable capacity sufficiency is taken as the power to be allocated to the current node. The power to be allocated to the current node is deducted from the current remaining power to be allocated until the remaining power to be allocated is 0 or the resource schedulable list is traversed to the end. Generate a collaborative peak shaving power allocation instruction that includes at least an instruction identifier, generation time, target node address, target device address, resource type, allocated power, start execution time, effective duration, and checksum, and send the collaborative peak shaving power allocation instruction to the energy storage controller, charging pile group control terminal, or flexible load controller of the corresponding target node.

8. The deep learning-based distributed energy collaborative management system for power distribution networks according to claim 1, characterized in that, After the closed-loop control is executed by sending the command to the hardware execution terminal, the following steps are also included: In the next data sampling cycle after the hardware execution terminal responds to the collaborative peak-shaving power allocation instruction, the multi-source data acquisition is re-triggered and the spatial propagation imbalance characteristics of the node are calculated again in real time. The trained deep neural network model is used to perform online verification of the current overvoltage risk level of the updated nodes; If the current overvoltage risk level of a node fails to fall back to the level of concern or normal, a risk warning signal containing a prompt to expand the scope of collaborative management will be issued and uploaded to the dispatch center.

9. The deep learning-based distributed energy collaborative management system for power distribution networks according to claim 1, characterized in that, The preset deep neural network model adopts a multilayer perceptron architecture, including having The system consists of an input layer with 10 neurons, three hidden layers with 256, 128, and 64 neurons respectively, and a layer with 3... The output layer has 10 neurons, and the three hidden layers all use linear rectified activation functions. The output layer is divided according to the node order in the topology routing matrix. There are three output groups, each containing three neurons. A normalized exponent operation is performed independently on each output group to obtain the probability that the corresponding node belongs to the normal level, the attention level, or the high-risk level. This represents the number of valid nodes in the topology routing matrix.

Images