A regional power internet coordination autonomous method for large-scale grid connection of new energy

Abstract

The present application relates to the technical field of power grid, in particular to a regional power internet coordination autonomous method for large-scale grid connection of new energy. First, the influence of historical reactive current in each line is eliminated to obtain the active transmission upper limit of each line; based on the active transmission upper limit, a local power scheduling model is constructed to obtain a scheduling cost marginal variable and determine the optimal power of each line; combined with the optimal power of each line, the historical average power factor and the line parameters, the end voltage drop of each line is determined, and the predicted voltage drop ratio is determined; the predicted node net power of the current node, the scheduling cost marginal variable and the predicted voltage drop ratio are input into the power scheduling strategy network to obtain the preliminary charging and discharging power setting value; the transformer residual capacity margin is used to limit the preliminary charging and discharging power setting value to obtain the final power setting value. The present application realizes the economic mutual aid optimization of regional electric energy and the absolute operation safety of the AC bottom layer physical hardware.

Description

Technical Field

[0001] This invention relates to the field of power grid technology, and specifically to a regional power internet coordination and autonomy method for large-scale grid connection of new energy sources. Background Technology

[0002] With the large-scale grid connection of distributed photovoltaic and energy storage devices, regional AC distribution networks frequently face local power supply and demand imbalances caused by random fluctuations in renewable energy output. In order to maintain local power balance, existing edge computing terminals typically rely on local area network interactions with surrounding nodes to autonomously perform cross-node power allocation and coordination, in the absence of a global communication master station.

[0003] Currently, when handling cross-node power dispatch at the edge, the main approach relies on simple unit price ranking and the static rated current of cables for active power allocation. However, the actual operation of AC power grids is constrained by multiple underlying electrical laws. First, conventional dispatch schemes ignore the implicit occupation of the actual current carrying capacity of cables by reactive power flow. Under low power factor conditions, directly issuing commands based on rated current can easily lead to cable overcurrent and overheating. Second, when multiple interconnecting lines experience concurrent congestion, the simple price ranking method cannot handle resource allocation conflicts constrained by many-to-many inequalities. Furthermore, high-power cross-node flows are significantly affected by AC line impedance. Issuing dispatch commands directly without considering impedance can lead to severe voltage drops at the end nodes of the lines.

[0004] Existing technologies have attempted to introduce data-driven deep reinforcement learning to replace traditional rule-based scheduling. However, a purely data-driven network is a black-box system lacking the concept of physical boundaries, and it cannot directly understand the impedance coupling patterns and capacity limits of the underlying AC power grid. When faced with load surges caused by extreme weather, the probabilistic power output commands have generalization errors, easily issuing overload commands that exceed the thermal stability tolerance limits of distribution transformers, ultimately leading to serious accidents such as transformer tripping and power outages. Summary of the Invention

[0005] To address the technical problem of generalization error in power commands output by neural networks under extreme conditions, the present invention aims to provide a coordinated and autonomous regional power internet method for large-scale grid connection of new energy sources. The specific technical solution adopted is as follows:

[0006] Obtain the predicted net power of the distribution transformer corresponding to the current node; connect the current node and its corresponding adjacent nodes to construct multiple lines;

[0007] For each line, after removing the influence of historical reactive current in each line, the upper limit of active power transmission for each line is obtained.

[0008] A local power scheduling model is constructed based on the active power transmission upper limit. The dual variables at the optimal poles corresponding to the equal balance constraints in the local power scheduling model are extracted as marginal variables of scheduling cost, and the optimal power of each line is determined.

[0009] By combining the optimal power, historical average power factor, and line parameters of each line, the voltage drop at the end of each line is determined; based on the voltage drop at the end of all lines, the predicted voltage drop ratio is determined.

[0010] The node scheduling state feature vector is composed of the predicted net power of the current node, the marginal variable of the scheduling cost, and the predicted voltage drop ratio; the node scheduling state feature vector is input into the power scheduling strategy network to obtain the preliminary charging and discharging power setting value.

[0011] By utilizing the remaining capacity margin of the transformer, the initial charging and discharging power setting value is limited to obtain the final power setting value.

[0012] The embodiments of the present invention have at least the following beneficial effects:

[0013] This invention first eliminates the reactive current consumption in each line, obtaining the active power transmission limit for each line, which represents the true transmission boundary of the line. Based on the active power transmission limit, a local power dispatch model is constructed, ultimately outputting a dispatch cost marginal variable representing the degree of congestion, and obtaining the optimal power for each line. The dispatch cost marginal variable reflects the consumption of low-cost electricity within the current local area network connectivity range. The larger the dispatch cost marginal variable, the more it indicates that the low-cost electricity within the current local area network connectivity range has been exhausted, or the only available tie line has reached its transmission limit. Then, the optimal power for each line is converted into a predicted voltage drop ratio to assess the voltage drop at the end of the AC grid. Based on the predicted net power of the nodes, the dispatch cost marginal variable, and the predicted voltage drop ratio, a preliminary charging and discharging power setting value is obtained, effectively avoiding the safety hazard of end-point voltage collapse caused by blindly issuing dispatch commands. Finally, based on the remaining capacity margin of the transformer, a mandatory threshold truncation is applied to the output preliminary charging and discharging power setting value, establishing a robust hardware protection bottom line and preventing the risk of equipment overload caused by algorithm overreach. This invention balances the economic optimization of regional power supply with the absolute operational security of the underlying physical hardware of the AC system at the edge side without global communication intervention. Attached Figure Description

[0014] Figure 1 A flowchart illustrating a method for coordinated and autonomous regional power internet integration of new energy sources, provided in one embodiment of the present invention;

[0015] Figure 2This is a flowchart illustrating a method for obtaining the active power transmission limit for each line, as provided in one embodiment of the present invention. Detailed Implementation

[0016] This invention provides a specific implementation method for a regional power internet coordinated autonomy approach for large-scale grid connection of new energy sources. This method is applicable to power internet coordinated autonomy scenarios. In this scenario, the initial operation of the entire regional power coordinated autonomy system is performed first. Device parameters and historical data from local static configuration files and smart meters are received. Through parameter presetting, net power prediction, and financial cost conversion operations, the node supply and demand status data packets of the current node and adjacent nodes are finally obtained. These packets provide static constraint benchmarks and real-time supply and demand environment data for downstream cross-node power scheduling and feature extraction. Specifically, this includes: reading the static configuration file, setting each distribution transformer to correspond to one node, and extracting the total rated capacity of the transformer recorded on the factory nameplate of the distribution transformer of the current node to be analyzed. The rated total capacity of the distribution transformer serves as the hardware upper limit for subsequent judgments on whether the distribution transformer is overloaded. Simultaneously, the battery life loss curve specified at the factory for the energy storage battery associated with the node is read. Specifically, the battery life loss curve is represented in memory as a mapping data table, recording the corresponding numerical relationship between the battery's lifespan loss at different depths of discharge and the lifespan lost in one charge-discharge cycle; that is, recording the corresponding numerical relationship between different remaining battery capacity percentages and the marginal lifespan loss rate in one charge-discharge cycle. Subsequently, nodes with direct electrical connections and local area network communication to the current node are identified as adjacent nodes, and the total number of adjacent nodes is defined as follows: ,in A positive integer. Initialize the index variables of adjacent nodes. Read the cable nameplate parameters connecting the current node and the first adjacent node, and record the maximum rated current carrying capacity of the first line. Next, the index variables of adjacent nodes. Incrementing sequentially, recording routes from the second to the third. The rated current carrying capacity of the line ,in, The total rated capacity of all the transformers read above. The battery life degradation curve and the line rated current carrying capacity are both stored as static constants in the terminal's working memory for direct access in all subsequent scheduling cycles. Finally, a pre-trained power scheduling policy network is loaded into the microprocessor's runtime memory to complete system initialization.

[0017] The following description, in conjunction with the accompanying drawings, details the specific scheme of the regional power internet coordination and autonomy method for large-scale grid connection of new energy provided by this invention.

[0018] Please see Figure 1 The diagram illustrates a flowchart of a regional power internet coordinated autonomous method for large-scale grid connection of new energy sources, provided by an embodiment of the present invention. The method includes the following steps:

[0019] Step S100: Using each distribution transformer as a node, connect the current node and its corresponding adjacent nodes to construct multiple lines.

[0020] First, the duration of a single scheduling cycle that runs continuously is set to... In this embodiment of the invention, the following is set: The value is 15 minutes.

[0021] Each distribution transformer is assigned a node. Taking any node to be analyzed as the current node as an example, for the current time and the current node, the actual load power sequence and actual photovoltaic output sequence of the current node in the previous scheduling cycle are collected; that is, the load power sequence and photovoltaic output sequence of the current node in the previous 15 minutes are obtained as the actual load power sequence and actual photovoltaic output sequence.

[0022] The actual load power sequence and the actual photovoltaic output sequence are input into a pre-set prediction model. After forward calculation, the prediction model outputs the predicted net power of the current node at the next future time. The unit for predicting the net electricity generation at each node is kilowatt-hours (kWh). Prediction models can employ time series forecasting algorithms such as Long Short-Term Memory networks or autoregressive moving average models.

[0023] Calculate the predicted node net power Then, based on the positive or negative polarity of this value, the boundary determination of supply and demand roles is performed:

[0024] when When the local photovoltaic power generation is insufficient to cover the load electricity consumption, the current node is determined to be a power shortage node.

[0025] when When the local power generation has met its own load, there is still surplus power, and the current node is determined to be a surplus node.

[0026] Furthermore, if the current node is determined to be a surplus node, the predicted net power of the node is extracted. The absolute value of the current node's available surplus power is defined separately. ,Right now This available surplus power This represents the physical upper limit of the maximum amount of electrical energy that the current node can transmit to the external network. It should be noted that the available surplus power of a node experiencing a power shortage is set to 0.

[0027] For surplus nodes, the remaining State of Charge (SOC) of the energy storage batteries at the current moment is read in real time by local smart meters. The remaining SOC ranges from 0% to 100%. The pre-stored battery life degradation curve in the working memory is retrieved. Using the current SOC as the independent variable, the corresponding marginal life degradation rate is calculated by looking up and interpolating in the battery life degradation curve mapping data table. This marginal life degradation rate represents the lifespan lost during one charge-discharge cycle. Subsequently, the local battery discharge cost is obtained by multiplying the marginal life degradation rate by the total purchase and replacement cost of the energy storage battery system. Local battery discharge cost The unit is yuan per kilowatt-hour.

[0028] The positive and negative trends reflected in this calculation mechanism are as follows: the lower the current remaining capacity percentage of the energy storage battery, the higher the marginal lifetime loss rate obtained from the table, and thus the higher the calculated local battery discharge cost. The larger the value, the higher the local battery discharge cost. The larger the value, the more serious the damage to the overall lifespan of the equipment caused by the current external discharge action, and the lower the willingness to supply power to the outside.

[0029] For nodes experiencing power shortages, directly reducing the cost of local battery discharge is crucial. The value is forcibly assigned to a preset maximum penalty constant. In this embodiment of the invention, a preset maximum penalty constant is used. It can be preset to a very large value, such as 9999 as a specific example. This assignment operation can directly cut off the logical feasibility of the power shortage node to supply power to the outside at the numerical level of the subsequent mathematical model. In other embodiments, the implementer can adjust the value according to the actual situation.

[0030] After all calculations are completed, the predicted net power of the current node will be calculated. Available surplus power and local battery discharge costs Numerical data is extracted and combined, and then packaged into a single node supply and demand status data packet.

[0031] Subsequently, a node data reception time window with a preset duration of 2 seconds is opened. Within this node data reception time window, the current node's supply and demand status data packet is broadcast to all connected neighboring nodes through the local area network communication interface; at the same time, the receiving port continuously listens for and collects node supply and demand status data packets sent by other neighboring nodes.

[0032] Once the 2-second node data reception time window countdown ends and closes, external data reception stops. The node supply and demand status data packets received by the current node from all neighboring nodes are stored in memory as external parameter benchmarks for subsequent cross-node power allocation model calculations.

[0033] The method for obtaining neighboring nodes is as follows: Nodes that have a direct electrical connection and local area network communication with the current node are identified as the neighboring nodes of the current node, and the total number of neighboring nodes is defined as follows: ,in It is a positive integer.

[0034] Then, connect the current node with its corresponding neighboring nodes to construct multiple lines. Initialize the index variables of the neighboring nodes. Read the cable nameplate parameters connecting the current node and the first adjacent node, and record the rated current carrying capacity of the first line. Next, the index variables of adjacent nodes. Incrementing sequentially, recording the second to the third subsequent routes. The rated current carrying capacity of the line ,in, Step S100 obtains the node supply and demand status data packets of the current node and its neighboring nodes, which are used to provide static constraint benchmarks and real-time supply and demand environment data for downstream cross-node power scheduling and feature extraction.

[0035] Step S200: For each line, after removing the influence of historical reactive current in each line, the upper limit of active power transmission for each line is obtained.

[0036] After obtaining the node supply and demand status data packets of each node, as well as the historical data recorded by the local smart meters, the actual transmission boundary of each line is analyzed by eliminating the occupation of reactive current, and the active power transmission upper limit of each line is obtained.

[0037] In AC power distribution networks, reactive power flow directly occupies the actual current-carrying capacity of cables. Therefore, this invention combines historical voltage data and power factor, and uses mathematical calculations to remove the influence of reactive power components, outputting the true upper limit of active power transmission for each connection.

[0038] Please see Figure 2 , Figure 2 A flowchart illustrating a method for obtaining the active power transmission limit for each line is provided. The method includes the following steps:

[0039] Step S210: For each line, obtain the line rated current limit, historical average line voltage, and historical average power factor.

[0040] It should be noted that the method for obtaining the rated current carrying capacity of each line has already been given in step S100, and will not be repeated here.

[0041] First, the node supply and demand status data packets of each node acquired within the node data reception time window are aligned and cleaned.

[0042] For the total number of Iterate through the neighboring nodes and check the reception status of the node supply and demand status data packets of each neighboring node. If it is determined that the data packet of the first node has not been received... If a data packet from a neighboring node is received, it indicates that that neighboring node has lost communication or experienced a failure; in other words, that neighboring node is determined to be a communication-disconnected node. At this time, the available surplus power corresponding to that neighboring node is... The value is assigned to 0, and the local battery discharge cost corresponding to the adjacent node is set to 0. The value is forcibly assigned to the preset maximum penalty constant. By assigning values ​​to nodes that have lost communication, the optimizer can be completely prevented from endlessly demanding power from these nodes in subsequent steps, thus avoiding the algorithm from falling into an infinite loop.

[0043] Next, the average line voltage of the current node in the previous scheduling cycle is read from the local smart meter and recorded as the historical average line voltage. It reads the average power factor of the lines connecting the current node and each adjacent node during the previous scheduling period, and records it as the historical average power factor. Taking the j-th adjacent node as an example, the historical average power factor of the line corresponding to the current node and the j-th adjacent node is recorded as... .

[0044] Then, the auto-incrementing loop calculation process is started, and the index variables of adjacent nodes are initialized. To calculate the active power transmission limit for each line, the rated current carrying capacity of each line is retrieved from the static configuration file. Taking the j-th adjacent node as an example, the rated current carrying capacity of the line corresponding to the current node and the j-th adjacent node is denoted as... .

[0045] Step S220: Combining the historical average power factor, the line rated current carrying limit, and the historical average line voltage, calculate the active power transmission limit for each line after removing the historical reactive current occupation.

[0046] Among them, the historical average power factor, the upper limit of the line rated current carrying capacity, and the historical average line voltage are all positively correlated with the upper limit of active power transmission.

[0047] As an example of the present invention, taking the j-th adjacent node as an example, the active power transmission limit of the line corresponding to the current node and the j-th adjacent node is... The calculation formula is:

[0048] ;

[0049] Among them, the upper limit of active transmission After removing historical reactive current occupancy, the current node and its j-th neighboring node are represented as follows: The upper limit of the maximum active power that the line can actually transmit.

[0050] The formula for calculating the upper limit of active power transmission ensures that, when the operating conditions of the power grid change, if the historical average power factor of the line... The lower the value, the greater the capacity occupied by reactive current, and the higher the calculated upper limit of active power transmission. It will decrease proportionally accordingly.

[0051] This yields the active power transmission limit for the current node and each adjacent node. Since there are N adjacent nodes, the active power transmission limit for N lines is obtained. This active power transmission limit reflects the safe power transmission space of each line.

[0052] Step S300: Construct a local power scheduling model based on the active power transmission upper limit, extract the dual variables at the optimal poles corresponding to the isoform balance constraints in the local power scheduling model as marginal variables of scheduling cost, and determine the optimal power of each line.

[0053] Since the power costs of adjacent nodes differ and the line capacity is limited, a local power scheduling model is constructed to solve the problem after obtaining the active power transmission limit of each line and the supply and demand status data packets of adjacent nodes. Finally, the dual variables at the optimal pole corresponding to the isotropic balance constraint in the local power scheduling model are extracted to quantify the economic penalty when multiple lines are congested concurrently, and the optimal power of each line is calculated simultaneously.

[0054] First, regardless of whether the current node is a power-deficient node or a power-surplus node, a local power scheduling model is uniformly constructed in the local working memory. This local power scheduling model is a linear programming model with the objective function of minimizing the total scheduling cost, the equality constraint of supply and demand balance, and the inequality constraints of the active power transmission limit of each line and the available surplus power of adjacent nodes.

[0055] The predicted node net power generated from the previous scheduling cycle corresponding to the current moment. Divide by the duration of the scheduling cycle The obtained value is used as the benchmark value for average demand power. If the current node is a surplus node, predict the node's net power consumption. In the model, this is represented by a negative demand, which means that the current node supplies electrical energy to the outside.

[0056] To calculate the power allocation under the cost-minimizing state, the objective function of this local power dispatching model is set as minimizing the total dispatching cost. The formula for the objective function is: ;in, The objective function value; The sequence of decision variables to be solved in the model represents the active power to be retrieved from adjacent nodes. For the first Local battery discharge cost of each adjacent node; This is a slack variable, representing the power shortfall that cannot be compensated for externally due to insufficient surrounding power supply or line congestion. The absolute value of the slack variable ensures that when a node has excess power that it cannot absorb, i.e. when the slack variable is negative, the objective function still imposes a positively increasing penalty cost on it, prompting the system to prioritize meeting supply and demand balance through cross-node scheduling. The preset penalty constant is used to set the financial weight of the relaxation term. As a specific example of the present invention, The value is taken as 10 times the replacement cost of a brand new battery. It should be noted that the replacement cost of a brand new battery refers to the cost per kilowatt-hour or unit of electricity obtained by purchasing and installing a brand new energy storage battery system under current conditions and dividing its total cost by the total capacity of the battery system.

[0057] It should be noted that the summation combination term This represents the total cost of normal cross-node power access. This objective function ensures that the optimizer will prioritize the cost of discharging local batteries. Power scheduling is performed on neighboring nodes with lower prices, and the penalty constant is only invoked when all low-priced nodes run out of power or the line becomes congested. The corresponding slack variables.

[0058] At the same time, an equality balance constraint is added to the above objective function to force supply and demand balance:

[0059] ;

[0060] Then, based on the obtained active power transmission limit for each line, inequality constraints are added to limit the transmission limits of the physical devices. For each line, the following constraints are added:

[0061] ;

[0062] ;

[0063] in, The available surplus power of the j-th adjacent node; This represents the upper limit of active power transmission.

[0064] The objective function and all constraints are input into a pre-defined simplex method calculation module for matrix iteration, which is equivalent to performing matrix iteration on the local power dispatch model using the simplex method. After the calculation, the simplex method calculation module outputs a sequence of decision variables that minimizes the objective function, denoted as the preliminary optimal power sequence. Each element in the preliminary optimal power sequence represents the optimal power for each line. Taking the j-th line as an example, the optimal power corresponding to the j-th line is denoted as... .

[0065] It should be noted that, since the local model solved by the simplex method is a static ideal state, it does not take into account the nonlinear decay of the energy storage battery discharge depth and the transient disturbances of the AC power grid. Therefore, this preliminary optimal power sequence is only used as a virtual simulation benchmark to characterize the current static supply and demand environment of the local area network, and is not used as the actual execution command at the system bottom layer.

[0066] Based on the above calculation process, the simplex method calculation module, using the Lagrange multiplier method principle, solves and outputs the dual variable values ​​corresponding to the equilibrium constraints in the above equation at the optimal pole. These dual variable values ​​are extracted and defined as the scheduling cost marginal variable. .

[0067] Among them, the marginal variable of scheduling cost Numerically, it reflects the marginal rate of change of total cost caused by changes in the equality constraints. (Schedule cost marginal variable) The larger the value, the more objectively it indicates that the low-cost electricity within the current local area network range has been exhausted, or that the only available connection has reached its transmission limit. If the power dispatch demand is to be further met, a high penalty will be incurred.

[0068] Step S400: Combine the optimal power, historical average power factor and line parameters of each line to determine the voltage drop at the end of each line; based on the voltage drop at the end of all lines, determine the predicted voltage drop ratio.

[0069] After obtaining the preliminary optimal power sequence containing the optimal power of each line and the marginal variable of scheduling cost, the line parameters of the local lines are also obtained. By performing a single-step power flow simulation, the linear preliminary optimal power sequence is transformed into an assessment of the voltage drop at the end of the AC grid; then, based on the voltage drop at the end of all lines, the predicted voltage drop ratio is determined.

[0070] Since the simplex method in the previous stage only assessed the current-carrying heat capacity of the line, it completely ignored the voltage drop caused by line impedance and current coupling in the AC power grid. In order to predict this electrical risk in advance, the obtained preliminary optimal power sequence was extrapolated and calculated to determine the predicted voltage drop ratio.

[0071] First, obtain the AC resistance and AC reactance of each line. Specifically, read the AC resistance and AC reactance of each line from the local static configuration file, and extract the optimal power of each line from the preliminary optimal power sequence output in the previous stage. Taking the j-th line as an example, obtain its corresponding AC resistance. AC reactance and optimal power .

[0072] Second, based on the optimal power and historical average power factor of each line, the estimated value of the reactive power component for each line is determined. Specifically, in AC circuits, the transmission of active power is inevitably accompanied by the flow of reactive power. To reconstruct the accompanying reactive power, for each line, the historical average power factor recorded in the data reception time window of the preceding node is retrieved. ;

[0073] In this embodiment of the invention, taking the j-th line as an example, the formula for calculating the estimated reactive power component of the j-th line is as follows:

[0074] ;

[0075] in, This represents the estimated reactive power component of the j-th line generated by the initial optimal power flow; Indicates assignment to the first The optimal power of the line; The j-th line represents the historical average power factor; max is the function that takes the maximum value. This represents a preset, extremely small positive number, as an example of the present invention. Can be preset as .

[0076] Since the average power factor tends to be as close as the current flowing through the cable approaches pure reactive power. Use the maximum value function With the preset minimum positive number The combination of these can effectively prevent floating-point arithmetic overflow crashes caused by the denominator of a division expression being zero.

[0077] The formula for calculating the reactive power component estimate ensures that, under the same optimal power conditions, a lower historical average power factor of the line results in a higher estimated reactive power component value. It will increase significantly in proportion.

[0078] Third, combining the estimated values ​​of AC resistance, AC reactance, and reactive power components, calculate the voltage drop at the end of each line. Specifically:

[0079] Using the principle of the single-step forward-backward substitution algorithm, the forward voltage drop caused by the transmission of the above power combination on each line is analyzed, and the terminal voltage drop of each line is determined.

[0080] Taking any line as the target line, multiply the optimal power of the target line by the AC resistance to obtain the active voltage drop data; multiply the estimated reactive power component of the target line by the AC reactance to obtain the reactive voltage drop data.

[0081] Using the sum of active voltage drop data and reactive voltage drop data as the numerator and the historical average line voltage of the current node as the denominator, the voltage drop at the end of the target line is obtained.

[0082] Taking the j-th line as an example, the formula for calculating the voltage drop at the end of the j-th line is:

[0083] ;

[0084] in, Indicates the first The voltage drop at the end of the line; Represents the AC resistance of the j-th line; Represents the AC reactance of the j-th line; This represents the historical average line voltage of the current node; This is the active power voltage drop data; This is the reactive power voltage drop data.

[0085] In the formula for calculating the voltage drop at the end, the numerator term... This represents the sum of the voltage drops across the impedance caused by active and reactive currents, respectively. The formula for calculating the terminal voltage drop reflects the increase in terminal voltage drop when the dispatched power value is larger or the line impedance is larger. The more serious it becomes, the better.

[0086] The calculations for N lines are completed sequentially to obtain N voltage drop values ​​at the ends.

[0087] Fourth, select the maximum value among the voltage drops at the ends of all lines as the maximum predicted voltage drop, denoted as . .

[0088] Fifth, calculate the proportion of the maximum predicted voltage drop in the rated operating line voltage of the power grid to obtain the predicted voltage drop ratio, more specifically:

[0089] Retrieve the rated operating line voltage of the power grid from the static configuration file. The predicted voltage drop ratio is obtained by division. The calculation formula is: ;

[0090] Among them, the predicted voltage drop ratio The larger the value, the more severe the voltage drop caused by the impedance of the heavy-load cable is when adjusting the power supply according to the current preliminary optimal power projection, which has approached or even exceeded the safety bottom line of power quality.

[0091] Step S500: The node scheduling state feature vector is constructed from the predicted net power of the current node, the marginal variable of the scheduling cost, and the predicted voltage drop ratio; the node scheduling state feature vector is input into the power scheduling strategy network to obtain the preliminary charging and discharging power setting value.

[0092] Since the obtained net power of the predicted node, the marginal variable of the scheduling cost, and the predicted voltage drop ratio are isolated indicators, a unified data set is output by executing vector combination actions, providing a complete observation environment for the neural network.

[0093] The node scheduling state feature vector is composed of the predicted net energy of the current node, the marginal variable of scheduling cost, and the predicted voltage drop ratio. Specifically, the predicted net energy of the node, the marginal variable of scheduling cost, and the predicted voltage drop ratio are extracted sequentially. Following the fixed extraction order, these three variables are arrayed and concatenated to form a one-dimensional vector containing three floating-point values. The system defines this one-dimensional vector as the node scheduling state feature vector. .

[0094] The constructed node scheduling state feature vector The specified data interaction address is stored in memory. This operation signifies that the local environmental constraint information and electrical risk information have been reconstructed into standardized environmental characteristic data, and the system is ready to send the node scheduling state feature vector as a trigger condition to the next stage of the algorithm network for forward propagation calculation.

[0095] Because the internal parameters of the node scheduling state feature vector exhibit significant differences in dimensions and orders of magnitude, directly inputting them into a neural network can lead to numerical overflow or computational failure. Therefore, a linear mapping is performed on the variables in the node scheduling state feature vector through data standardization and neural network operations.

[0096] To eliminate the interference of differences in physical dimensions and orders of magnitude on the algorithm, a maximum-minimum normalization method is used to extract the... , , These three variables are linearly mapped to... Within a standardized numerical range. This standardization operation effectively prevents large, unprocessed values ​​from directly entering the neural network, which could lead to activation function saturation or gradient vanishing errors. Specifically, during maximum and minimum value normalization, the maximum value of each variable is the maximum value in the historical database, and the minimum value is the minimum value in the historical database. If the current real-time value of a variable is greater than the maximum value, it is directly set to 1; if the current real-time value is less than the minimum value, it is directly set to 0.

[0097] Subsequently, the normalized node scheduling state feature vector is input into the input layer of the power scheduling policy network, which has been loaded into the runtime memory in the initial stage. This power scheduling policy network adopts a multi-layer fully connected neuron structure with fixed parameters, such as the action output network in the deep deterministic policy gradient algorithm, and uses a microprocessor to perform forward propagation matrix multiplication operations.

[0098] The power scheduling strategy network is a fully connected feedforward neural network, with the following structure example: The input layer has 3 neurons, corresponding to the normalized... , , The network has two hidden layers: the first hidden layer has 64 neurons and the second hidden layer has 32 neurons, both using the ReLU activation function; the output layer has one neuron; this power scheduling policy network is trained offline using either the Deep Deterministic Policy Gradient (DDPG) algorithm or the Proximal Policy Optimization (PPO) algorithm, and the training dataset consists of historical running data and the simulation environment.

[0099] During the internal matrix operations of the power scheduling strategy network, the input environment characteristics and output actions exhibit the following feedback pattern:

[0100] When the input scheduling cost marginal variable When the value increases, it indicates an extreme shortage of low-cost surplus power within the current grid connection range. At this time, the power dispatch strategy network will map an output with an absolute value approaching zero to avoid consuming extremely high battery depreciation costs to supplement a small amount of power.

[0101] When the input predicted voltage drop ratio When the value approaches the alarm lower limit of power quality, it indicates that the voltage drop caused by the AC impedance of the line can no longer maintain system stability. At this time, the power dispatch strategy network will forcibly reduce the amplitude of the output command and terminate the cross-node mutual assistance action in order to maintain the local voltage level.

[0102] The power scheduling strategy network needs to be trained offline before loading. This offline training process uses the aforementioned node scheduling state feature vectors as environmental state inputs and the initial charge / discharge power settings as action outputs.

[0103] The reward function is set as follows during training: ,in, , , The preset positive weighting coefficients are used by the implementer to set corresponding positive weighting coefficients for the predicted node net power, the marginal variable of scheduling cost, and the predicted voltage drop ratio according to the actual situation. As an example, , , The value of can be 1 / 3; This is the initial charge / discharge power setting value; This is the absolute value of the initial charge / discharge power setting.

[0104] The reward function indicates that the penalty imposed on the system increases as the marginal variable of scheduling cost or the predicted voltage drop ratio increases. Through continuous interaction with the simulation environment, the network parameters gradually converge, learning to output initial charge and discharge power setpoints with amplitudes approaching zero under high cost or high voltage drop risk.

[0105] After linear mapping at the output layer of the power scheduling strategy network, a real-valued result is obtained, which is then established as the initial charge and discharge power setpoint for the current scheduling cycle. .

[0106] In step S600, the remaining capacity margin of the transformer is used to limit the initial charging and discharging power setting value to obtain the final power setting value.

[0107] Because neural networks have a blind spot of generalization error when encountering extremely rare grid load disturbances, directly accepting the initial charge / discharge power setting value may cause transformer overload tripping. Therefore, after obtaining the initial charge / discharge power setting value, a safety verification mechanism is established by introducing objective physical equipment rating constants, and a truncated safety command is output.

[0108] First, read the rated total capacity of the distribution transformers of the current node from the local static configuration file. To avoid drastic numerical fluctuations caused by high-frequency harmonics during grid sampling, the arithmetic mean of the load power at all power sampling points recorded by the smart meter within one minute prior to the current command generation is calculated to obtain the average base load power. The average base load power This serves as the real-time baseline load power for the transformer.

[0109] Secondly, to calculate the additional power capacity the transformer can currently handle, the difference between the transformer's rated total capacity and its real-time base load power is used to obtain the transformer's remaining capacity margin. If the remaining capacity margin is negative, it is updated to 0. This remaining capacity margin... The calculation formula is: .

[0110] This remaining capacity margin ensures that under normal conditions, the remaining capacity margin equals the rated total capacity minus the base load; while under extreme and sudden operating conditions, i.e., when the average base load power... Exceeding the rated total capacity Time, that is When the difference is negative, the remaining capacity margin is forcibly output using the maximum value function. This extreme value protection mechanism completely eliminates program crashes and deadlocks caused by negative remaining capacity margins during subsequent absolute value comparisons. (Remaining capacity margin) The closer to This indicates that the local distribution transformer is close to full-load thermal stability, which narrows the adjustment space for additional operations of energy storage devices.

[0111] Then, it is determined whether the absolute value of the initial charge / discharge power setting is greater than the remaining capacity margin of the transformer.

[0112] If the absolute value of the initial charge / discharge power setting value Greater than the transformer's remaining capacity margin ,Right now If the result is deemed as exceeding the limit, it indicates that directly issuing the preliminary command generated by the algorithm will trigger a transformer overload shutdown. In this case, a mandatory safety threshold limiting action is immediately executed: the preliminary charging and discharging power setting value is retained. The sign of the symbol is positive, representing charging, and negative, representing discharging. Its absolute value is then updated to the transformer's remaining capacity margin. The updated value will be used as the final power setting. It should be noted that if there is remaining capacity margin... equal The absolute value of the instruction is then preserved until... This means immediately stopping the charging and discharging of the energy storage device to protect the hardware. More specifically: in At that time, if If the value is negative, the corresponding final power setting value is ;like If the value is positive, the corresponding final power setting value is .

[0113] If the absolute value of the initial charge / discharge power setting is less than or equal to the transformer's remaining capacity margin, i.e. If the result is safe, it indicates that the initial charge / discharge power setting value output by the neural network is within the real-time safe bearing range of the transformer, and the initial charge / discharge power setting value is retained as the final power setting value.

[0114] The safety value obtained after the above comparison and truncation operations is defined as the final power setpoint. Finally, the edge computing terminal transmits the final power setting value via an RS485 industrial communication bus or an industrial Ethernet interface. The data is converted into digital control messages and sent to the microcontroller register of the underlying energy storage converter. Based on these numerical instructions, the energy storage converter executes the corresponding internal inverter or rectifier control actions. This completes the full data and control closed-loop process for regional power coordination and autonomy.

Claims

1. A method for coordinated and autonomous regional power internet integration of new energy sources on a large scale, characterized in that, The method includes the following steps: Each distribution transformer is treated as a node, and multiple lines are constructed by connecting the current node with its corresponding adjacent nodes. For each line, after removing the influence of historical reactive current in each line, the upper limit of active power transmission for each line is obtained. A local power scheduling model is constructed based on the active power transmission upper limit. The dual variables at the optimal poles corresponding to the equal balance constraints in the local power scheduling model are extracted as marginal variables of scheduling cost, and the optimal power of each line is determined. By combining the optimal power, historical average power factor, and line parameters of each line, the voltage drop at the end of each line is determined; based on the voltage drop at the end of all lines, the predicted voltage drop ratio is determined. The node scheduling state feature vector is composed of the predicted net power of the current node, the marginal variable of the scheduling cost, and the predicted voltage drop ratio; the node scheduling state feature vector is input into the power scheduling strategy network to obtain the preliminary charging and discharging power setting value. By utilizing the remaining capacity margin of the transformer, the initial charging and discharging power setting value is limited to obtain the final power setting value.

2. The regional power internet coordination and autonomy method for large-scale grid connection of new energy sources according to claim 1, characterized in that, For each line, after removing the influence of historical reactive current, the upper limit of active power transmission for each line is obtained, including: For each line, obtain the line rated current limit, historical average line voltage, and historical average power factor. By combining the historical average power factor, the line rated current carrying capacity limit, and the historical average line voltage, the active power transmission limit for each line after removing historical reactive current occupancy is calculated; wherein, the historical average power factor, the line rated current carrying capacity limit, and the historical average line voltage are all positively correlated with the active power transmission limit.

3. The regional power internet coordination and autonomy method for large-scale grid connection of new energy sources according to claim 1, characterized in that, The construction of a local power scheduling model based on the active power transmission upper limit includes: A linear programming model is constructed with the objective function of minimizing total scheduling cost, the equality constraint of supply and demand balance, and the inequality constraint of the active power transmission limit of each line and the available surplus power of adjacent nodes, as a local power scheduling model.

4. The regional power internet coordination and autonomy method for large-scale grid connection of new energy sources according to claim 3, characterized in that, The step of extracting the dual variables at the optimal poles corresponding to the equal balance constraints in the local power scheduling model as marginal variables of scheduling cost, and determining the optimal power of each line, includes: The local power dispatching model is subjected to matrix iteration using the simplex method, and the output is a sequence of decision variables that minimizes the objective function, which serves as the preliminary optimal power sequence; each element in the preliminary optimal power sequence represents the optimal power for each line. Based on the principle of the Lagrange multiplier method, the values ​​of the dual variables of the equality balance constraint at the optimal pole are determined as the marginal variables of the scheduling cost.

5. The regional power internet coordination and autonomy method for large-scale grid connection of new energy sources according to claim 1, characterized in that, The determination of the terminal voltage drop for each line, combining the optimal power, historical average power factor, and line parameters, includes: Obtain the AC resistance and AC reactance of each line; Based on the optimal power and historical average power factor of each line, the estimated value of the reactive power component of each line is determined. Based on the estimated values ​​of AC resistance, AC reactance, and reactive power components, the voltage drop at the end of each line is calculated.

6. The regional power internet coordination and autonomy method for large-scale grid connection of new energy sources according to claim 1, characterized in that, The determination of the predicted voltage drop ratio based on the voltage drop at the end of all lines includes: The maximum value among the voltage drops at the ends of all lines is selected as the maximum predicted voltage drop. The proportion of the maximum predicted voltage drop in the rated operating line voltage of the power grid is calculated to obtain the predicted voltage drop ratio.

7. The regional power internet coordination and autonomy method for large-scale grid connection of new energy sources according to claim 1, characterized in that, The step of inputting the node scheduling state feature vector into the power scheduling policy network to obtain the preliminary charging and discharging power setting value includes: Before inputting the power scheduling policy network, the element values ​​in the node scheduling state feature vector are normalized respectively, and the normalized node scheduling state feature vector is input into the input layer of the power scheduling policy network. The power scheduling strategy network is a multi-layer fully connected neuron structure network. After linear mapping through the output layer of the power scheduling strategy network, the initial charging and discharging power settings are obtained.

8. The regional power internet coordination and autonomy method for large-scale grid connection of new energy sources according to claim 1, characterized in that, The method for obtaining the remaining capacity margin of the transformer is as follows: The difference between the rated total capacity of the transformer and the real-time base load power of the transformer is calculated as the remaining capacity margin of the transformer. If the remaining capacity margin is negative, then the remaining capacity margin of the transformer will be updated to 0.

9. The regional power internet coordination and autonomy method for large-scale grid connection of new energy sources according to claim 1, characterized in that, The method of utilizing the remaining capacity margin of the transformer to limit the initial charging and discharging power setting value to obtain the final power setting value includes: If the absolute value of the initial charge / discharge power setting is greater than the remaining capacity margin of the transformer, then the sign of the initial charge / discharge power setting is retained, its absolute value is updated to the remaining capacity margin of the transformer, and the updated value is used as the final power setting value. If the absolute value of the initial charge / discharge power setting is less than or equal to the remaining capacity margin of the transformer, then the initial charge / discharge power setting is retained as the final power setting.

10. The regional power internet coordination and autonomy method for large-scale grid connection of new energy sources according to claim 5, characterized in that, The calculation of the voltage drop at the end of each line, combining the estimated values ​​of AC resistance, AC reactance, and reactive power components, includes: Taking any line as the target line, multiply the optimal power of the target line by the AC resistance to obtain the active voltage drop data; multiply the estimated reactive power component of the target line by the AC reactance to obtain the reactive voltage drop data. Using the sum of active voltage drop data and reactive voltage drop data as the numerator and the historical average line voltage of the current node as the denominator, the voltage drop at the end of the target line is obtained.

Images