A power grid intelligent dispatching system and method based on digital twinning and deep learning

Abstract

The application discloses a kind of power grid intelligent scheduling system and method based on digital twinning and deep learning, it is related to power grid intelligent scheduling technical field, a kind of power grid intelligent scheduling method based on digital twinning and deep learning, specific steps include: step one, by fusing multi-source heterogeneous data and dynamic parameter updating mechanism, construct and physical power grid synchronous mapping power grid digital twinning, step two, adopt federal learning framework aggregation multi-area power grid data, establish equipment degradation model, perfect priority data channel mechanism, step three, based on power grid digital twinning determines real-time power grid state and equipment risk coefficient, designs deep reinforcement learning dynamic decision-making framework, realizes the load scheduling optimization of multi-objective cooperation, the application is based on power grid digital twinning determines real-time power grid state and equipment risk coefficient, designs deep reinforcement learning dynamic decision-making framework, can realize the load scheduling optimization of multi-objective cooperation.

Description

Technical Field

[0001] This invention belongs to the field of intelligent power grid dispatching technology, specifically relating to an intelligent power grid dispatching system and method based on digital twins and deep learning. Background Technology

[0002] With the continuous growth of electricity demand, the power grid is expanding in scale and becoming increasingly complex. The power grid contains a large number of devices, lines, and nodes. The interactions and influences between these elements make it difficult to accurately grasp the grid's operational status. Furthermore, grid operation is affected by various factors such as weather conditions and changes in user load, further increasing the difficulty of grid dispatch and management. To achieve real-time monitoring and management of the power grid, it is necessary to collect a large amount of grid operation data, including electrical parameters, meteorological data, user load data, and equipment status data. However, this data often originates from different systems and devices, exhibiting multi-source and heterogeneous characteristics. How to effectively collect, integrate, and process this data has become a significant challenge for power grid dispatch and management. Power grid load dispatch is a complex multi-objective optimization problem. While meeting user electricity demands, it is also necessary to consider multiple objectives such as grid security, economy, and environmental protection. Traditional load dispatch methods often fail to simultaneously satisfy these objectives, leading to low grid operating efficiency and resource waste.

[0003] The current situation urgently requires a smart power grid scheduling method based on digital twins and deep learning to solve the above problems and achieve multi-objective collaborative optimization. Summary of the Invention

[0004] The purpose of this invention is to provide a smart grid dispatching system and method based on digital twins and deep learning, which solves the technical problem in the prior art that it is difficult to conduct dispatching and management while meeting users' electricity demand and considering multiple objectives such as grid security and economy.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A smart grid dispatching method based on digital twins and deep learning includes:

[0007] Step 1: By integrating multi-source heterogeneous data with a dynamic parameter update mechanism, construct a digital twin of the power grid that is synchronously mapped with the physical power grid;

[0008] Step 2: Use a federated learning framework to aggregate power grid data from multiple regions, establish an equipment degradation model, and improve the priority data channel mechanism;

[0009] Step 3: Based on the digital twin of the power grid, determine the real-time power grid status and equipment risk coefficients, design a deep reinforcement learning dynamic decision-making framework, and realize multi-objective collaborative load scheduling optimization.

[0010] Furthermore, a digital twin of the power grid, synchronously mapped with the physical power grid, is constructed using the following method:

[0011] Identify power grid equipment and lines, using power connection points or energy conversion devices in the physical power grid as nodes and energy transmission channels in the physical power grid as lines;

[0012] Real-time electrical parameter data of the power grid is collected at specific time intervals, meteorological data is obtained by calling the meteorological API at fixed cycles, user load data is aggregated by parsing the smart meter protocol, and equipment status data, including load rate, oil temperature and vibration amplitude, are monitored in real time through IoT sensors.

[0013] Using time window t as the time window, determine the average values ​​of electrical parameter data, meteorological data, user load data and equipment status data within the time window, and establish a four-dimensional data matrix;

[0014] Based on the CIM / E standard, the physical power grid is mapped as a node-edge model, and electrical parameters are embedded for each line.

[0015] Furthermore, a four-dimensional data matrix is ​​constructed, specifically as follows:

[0016] The four-dimensional data matrix includes a power grid operation parameter matrix, a meteorological parameter matrix, a user load matrix, and a fault record matrix. Each matrix has a dimension of T×N, where T is the total number of time windows and N is the total number of physical nodes in the power grid.

[0017] The matrix elements of the power grid operation parameter matrix are current, voltage, and active power;

[0018] The elements of the meteorological parameter matrix are regional temperature and humidity, wind speed, and light intensity;

[0019] The elements of the user load matrix are the real-time user loads;

[0020] The elements of the fault record matrix are encoded with 0 and 1, where 0 indicates that the equipment is normal and 1 indicates that the equipment is abnormal.

[0021] Based on the equipment status data, set the abnormal oil temperature threshold and abnormal vibration amplitude threshold. Within a time window, if either the average abnormal oil temperature or the average abnormal vibration amplitude is greater than or equal to its corresponding threshold, the equipment is judged to be abnormal within that time window; otherwise, the equipment is judged to be normal within that time window.

[0022] Furthermore, a federated learning framework is used to aggregate power grid data from multiple regions and establish an equipment degradation model. The specific method is as follows:

[0023] Deploy a federated learning architecture, set up a central aggregation server and regional grid clients, with the clients only uploading encrypted model gradients;

[0024] The server aggregates the encryption gradients of each region using a weighted average method. The weight of the aggregated encryption gradient is equal to the ratio of the number of local training samples in that region to the number of local training samples in all regions.

[0025] The global model parameters are updated based on the aggregated encryption gradient. The updated global model parameters are equal to the difference between the old global model parameters and the aggregated gradient. The aggregated gradient is the product of the aggregated encryption gradient and the learning rate. The updated global model parameters are then distributed to clients in each region.

[0026] When the equipment load rate is greater than the threshold H for n consecutive time windows, the twin resistance parameter is dynamically updated based on the remaining lifetime prediction value. The updated twin resistance parameter is equal to the combined value of the old resistance parameter and the remaining lifetime decay rate.

[0027] The risk coefficient is obtained by calculating the load intensity of the equipment and the degree of remaining life decay, and then weighting and summing them.

[0028] Furthermore, the risk coefficient is obtained by calculating the equipment load intensity and the degree of remaining lifespan degradation, and then weighting and summing them. The specific method is as follows:

[0029] The thresholds for equipment load rate and remaining lifespan degradation rate are determined. The average equipment load rate of a certain time window is compared with the equipment load rate threshold to determine the equipment load intensity of that time window. The remaining lifespan degradation rate of a certain time window is determined by comparing the predicted remaining lifespan of the equipment in a certain time window with the equipment remaining lifespan of the previous time window. The degradation rate of the equipment is represented by the average remaining lifespan degradation rate of n consecutive time windows. The degradation rate of the equipment is compared with the remaining lifespan degradation rate threshold to determine the degree of remaining lifespan degradation in that time window. Different weight values ​​are assigned to the equipment load intensity and the degree of remaining lifespan degradation, and then the results are summed to obtain the risk coefficient.

[0030] Furthermore, the priority data channel mechanism will be improved, specifically through the following methods:

[0031] A three-tiered transmission assurance mechanism is constructed based on risk coefficients. Higher risk coefficient values ​​indicate a faster risk of equipment health deterioration. Risk coefficient thresholds A1 and A2 are set based on historical equipment failure data. Different equipment categories are then categorized according to these thresholds, and different transmission protocols are matched to each category. The specific transmission protocol matching method is as follows: Where x represents the x-th device, This represents the risk coefficient of device x. Let x represent the transmission protocol of device x. Every n*t time interval, the device attenuation rate is updated through federated learning, and the risk coefficient is recalculated.

[0032] Furthermore, based on the real-time power grid status and equipment risk coefficients determined by the power grid digital twin, a deep reinforcement learning dynamic decision-making framework is designed. The specific method is as follows:

[0033] Constructing the state space of a Markov decision process based on a four-dimensional data matrix of a power grid digital twin;

[0034] The basic action dimensions for dividing the action space are generator output adjustment dimension, energy storage charging and discharging control dimension, and interruptible load management dimension.

[0035] Based on the equipment risk coefficient, dynamic action constraint rules are established. When a certain equipment... Immediately lock the load increase action of the power supply line of the equipment;

[0036] when At that time, the output increment limit will be implemented;

[0037] when At that time, maintain the status quo;

[0038] Establish a power grid topology-action mapping table, assign a unique action index number to each physical node, and automatically expand the action space dimension when the power grid is expanded and new nodes are added;

[0039] Establish an action-effect mapping database and dynamically optimize the action space.

[0040] Furthermore, an action-effect mapping database is established to dynamically optimize the action space. The specific method is as follows:

[0041] Record the actual network loss change rate, equipment load rate fluctuation range, and risk coefficient change rate after the execution of historical actions;

[0042] Statistical analysis of the effectiveness scores of each action dimension is performed to dynamically eliminate inefficient actions, namely interruptible load commands that fail to improve network loss multiple times in a row.

[0043] Based on simulation results, a safe and feasible set of actions is selected to ensure that scheduling decisions meet the constraints of power grid operation.

[0044] This invention also provides a power grid intelligent dispatching system based on digital twins and deep learning, applied to a power grid intelligent dispatching method based on digital twins and deep learning, comprising:

[0045] The power grid digital twin construction module constructs a power grid digital twin that is synchronously mapped to the physical power grid by integrating multi-source heterogeneous data and a dynamic parameter update mechanism.

[0046] The equipment degradation model building module uses a federated learning framework to aggregate power grid data from multiple regions, build an equipment degradation model, and improve the priority data channel mechanism.

[0047] The load scheduling optimization module determines the real-time power grid status and equipment risk coefficients based on the power grid digital twin, and designs a deep reinforcement learning dynamic decision-making framework to achieve multi-objective collaborative load scheduling optimization.

[0048] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0049] 1. This invention achieves real-time assessment of equipment health status by setting abnormal oil temperature thresholds and abnormal vibration amplitude thresholds, and making judgments based on the average data within a time window. This helps to detect potential equipment faults in a timely manner, perform maintenance or replacement in advance, and avoid power outages and losses caused by equipment failures. The established four-dimensional data matrix contains multi-dimensional information such as power grid operating parameters, meteorological parameters, user loads, and fault records, providing rich data support for complex power grid analysis.

[0050] 2. This invention achieves accurate prediction of equipment remaining life by constructing a prediction model for equipment remaining life and using equipment status data (such as load rate, oil temperature, and vibration amplitude) as input features. This helps power grid operators to formulate maintenance plans in advance and avoid power outages and losses caused by equipment failures. The twin resistance parameters are dynamically updated based on the equipment load rate and the predicted remaining life. By determining the threshold between the equipment load rate and the remaining life decay rate and calculating the risk coefficient, a quantitative assessment of the equipment health status is achieved.

[0051] 3. This invention achieves comprehensive control over power grid load scheduling by dividing the action space into basic action dimensions, including generator output adjustment, energy storage charging and discharging control, and interruptible load management. This multi-dimensional action space design enables the decision-making framework to flexibly respond to various changes in power grid operation, improve scheduling efficiency, and ensure that dynamic action constraint rules based on equipment risk coefficients can ensure timely measures are taken when equipment malfunctions, preventing the fault from escalating and guaranteeing the safe and stable operation of the power grid. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, 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.

[0053] Figure 1 A flowchart illustrating the steps of a smart grid dispatching method based on digital twins and deep learning is shown.

[0054] Figure 2A flowchart of a smart grid scheduling method based on digital twins and deep learning is shown.

[0055] Figure 3 A block diagram of a smart grid dispatching system based on digital twins and deep learning is shown. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Example 1, such as Figure 1 , Figure 2 The illustrated intelligent power grid dispatching method based on digital twins and deep learning specifically includes the following steps:

[0058] Step 1: By integrating multi-source heterogeneous data with a dynamic parameter update mechanism, construct a digital twin of the power grid that is synchronously mapped with the physical power grid.

[0059] The system identifies power grid equipment and lines, using power connection points or energy conversion devices in the physical power grid as nodes and energy transmission channels in the physical power grid as lines. It collects multi-source heterogeneous data at specific time intervals (e.g., 10 seconds) through the SCADA system, including real-time electrical parameter data of the power grid, calls meteorological APIs to obtain meteorological data at fixed cycles (e.g., 5 minutes), parses smart meter protocols to aggregate user load data, and monitors equipment status data (load rate, oil temperature, vibration amplitude) in real time through IoT sensors.

[0060] The timestamp is calibrated using the NTP protocol (error <1ms). Spatial interpolation or LSTM prediction is performed to fill missing data. Min-Max normalization and Z-Score standardization are performed on electrical parameters, meteorological data, user load data and equipment status data respectively.

[0061] Using time window t as the time window, the abnormal oil temperature threshold and abnormal vibration amplitude threshold are set according to the equipment status data. The average values ​​of electrical parameter data, meteorological data, user load data and equipment status data within the time window are determined, and a four-dimensional data matrix is ​​established. The four-dimensional data matrix includes a power grid operation parameter matrix (matrix elements are current, voltage and active power), a meteorological parameter matrix (matrix elements are regional temperature and humidity, wind speed and light intensity), a user load matrix (matrix elements are real-time user load), and a fault record matrix (matrix elements are 0 and 1 encoded, 0 indicates normal equipment and 1 indicates abnormal). The dimension of each matrix is ​​T×N, where T is the total number of time windows and N is the total number of physical nodes in the power grid.

[0062] If, within a certain time window, either the average abnormal oil temperature or the average abnormal vibration amplitude of the equipment is greater than or equal to its corresponding threshold, the equipment is judged to be abnormal within that time window.

[0063] Based on the CIM / E standard, the physical power grid is mapped as a node-edge model, and electrical parameters are embedded for each line.

[0064] Step 2: Use a federated learning framework to aggregate power grid data from multiple regions, establish an equipment degradation model, and improve the priority data channel mechanism.

[0065] A federated learning architecture is deployed, with a central aggregation server and regional power grid clients. Clients only upload encrypted model gradients and are prohibited from transmitting raw data. Each client trains using local historical data. The input feature vector is the average value of equipment status data for each time window (average equipment load rate, average oil temperature, and average vibration amplitude). A three-layer fully connected neural network is used, with the input layer being a 3D feature vector (average value of equipment status data) and the output layer being the remaining life prediction value. The activation function is LeakyReLU, resulting in an equipment remaining life prediction model, denoted as Lr.

[0066] Each regional power grid client retains local equipment status data and calculates the local model gradient. The server aggregates the encrypted gradient using a weighted average method. The encrypted gradient is obtained by adding Gaussian noise to the original gradient vector of the local model. The aggregated encrypted gradient is obtained by weighted summation of the encrypted gradients of each region. The gradient direction points to the steepest growth direction of the loss function. The weight of the encrypted gradient of a certain region is equal to the ratio of the number of local training samples in that region to the number of local training samples in all regions. The global model parameters are updated by aggregating the encrypted gradient. That is, the updated global model parameters are equal to the difference between the old global model parameters and the aggregated gradient (indicating parameter adjustment in the opposite direction of the gradient, since the optimization objective is to minimize the prediction error). The aggregated gradient is equal to the product of the aggregated encrypted gradient and the learning rate (a scalar value used to control the parameter update step size, determining the convergence speed and stability). The updated global model parameters are then distributed to each regional client to replace their local models.

[0067] When the equipment load rate remains above H for n consecutive time windows, the twin resistance parameter is dynamically updated based on the predicted remaining equipment lifetime. The remaining lifetime degradation rate is obtained based on the predicted remaining lifetime values ​​from two consecutive time windows. The updated twin resistance parameter is equal to the combined value of the old resistance parameter and the remaining lifetime degradation rate, denoted as […]. ,in This indicates the updated twin resistance parameters. This indicates the old resistor parameters. Indicates the weight of lifetime decay rate. This indicates the remaining lifetime decay rate.

[0068] The thresholds for equipment load rate and remaining lifespan degradation rate are determined. The average equipment load rate of a certain time window is compared with the equipment load rate threshold to determine the equipment load intensity of that time window. The remaining lifespan degradation rate of a certain time window is determined by comparing the equipment remaining lifespan prediction value of a certain time window with that of the previous time window. The degradation rate of the equipment is represented by the average remaining lifespan degradation rate of n consecutive time windows. The degradation rate of the equipment is compared with the remaining lifespan degradation rate threshold to determine the degree of remaining lifespan degradation of that time window. Different weight values ​​are assigned to equipment load intensity and degree of remaining lifespan degradation, and then the results are summed to obtain the risk coefficient Rd that reflects the health status of the equipment.

[0069] A three-tiered transmission assurance mechanism is constructed based on risk coefficients. Higher risk coefficient values ​​indicate a faster risk of equipment health deterioration. Risk coefficient thresholds A1 and A2 are set based on historical equipment failure data. Different equipment categories are then categorized according to these thresholds, and different transmission protocols are matched to each category. The specific transmission protocol matching method is as follows: Where x represents the x-th device, This represents the risk coefficient of device x. Let x represent the transmission protocol of device x. Every n*t time interval, the device attenuation rate is updated through federated learning, and the risk coefficient is recalculated.

[0070] Step 3: Based on the digital twin of the power grid, determine the real-time power grid status and equipment risk coefficients, design a deep reinforcement learning dynamic decision-making framework, and realize multi-objective collaborative load scheduling optimization.

[0071] Constructing the state space of a Markov decision process based on a four-dimensional data matrix of a power grid digital twin;

[0072] The basic action dimensions for dividing the action space include: generator output adjustment dimension, energy storage charging and discharging control dimension, and interruptible load management dimension;

[0073] The generator output adjustment dimension includes setting continuous action variables, allowing for fine-tuning within the generator's maximum output range. In practice, an action range [-Pmax, +Pmax] is defined for each generator, where negative values ​​indicate reduced output and positive values ​​indicate increased output.

[0074] The energy storage charging and discharging control dimension includes the use of three-valued discrete action variables, defining "-1" as the discharging mode (releasing stored energy), "0" as the idle mode (maintaining the current state), and "+1" as the charging mode (absorbing excess energy).

[0075] Interruptible load management includes the use of binary action variables, defining "0" as maintaining power supply and "1" as executing load interruption, with each variable corresponding to a specific industrial user group;

[0076] Based on the equipment risk coefficient, dynamic action constraint rules are established. When a certain equipment... When this happens, immediately lock the load increase action of the power supply line of the equipment. The specific implementation method is to forcibly set the upper limit of the generator output adjustment action associated with the line to 0, and only allow the output to be reduced or the status quo to be maintained.

[0077] when When this is implemented, the output increment limit is applied. In operation, the output adjustment range of the generator on this line is limited to Q% of the rated output. For example, for a unit with a rated output of 100MW, the output adjustment range of the generator on this line is limited to 30% of the rated output, and the operating space is compressed to [-30MW, +30MW].

[0078] when At that time, maintain the status quo;

[0079] A power grid topology-action mapping table is established, and a unique action index number is assigned to each physical node. When the power grid is expanded and new nodes are added, the action space dimension is automatically expanded. In practice, this is achieved through a registration mechanism. When a new device is connected, it sends a registration request to the reinforcement learning agent, triggering an increase of 3 in the action space dimension (corresponding to the addition of three types of actions).

[0080] Before issuing an action, the action to be executed is input into a digital twin for safety simulation. The simulation simulates the distribution of line load rate after execution, calculates the extreme value of node voltage deviation, detects the pass rate of N-1 faults, establishes an action-effect mapping database, records the actual network loss change rate after historical action execution, the fluctuation range of equipment load rate and the change rate of risk coefficient, statistically analyzes the benefit score of each action dimension, and dynamically eliminates inefficient actions (such as disabling a certain interruptible load instruction if it fails to improve network loss after 10 consecutive times).

[0081] Example 2, as Figure 3 The power grid intelligent dispatching system shown includes the following components: (The system is based on digital twins and deep learning.)

[0082] The power grid digital twin construction module constructs a power grid digital twin that is synchronously mapped to the physical power grid by integrating multi-source heterogeneous data and a dynamic parameter update mechanism.

[0083] Furthermore, a digital twin of the power grid, synchronously mapped with the physical power grid, is constructed using the following method:

[0084] Identify power grid equipment and lines, using power connection points or energy conversion devices in the physical power grid as nodes and energy transmission channels in the physical power grid as lines;

[0085] Real-time electrical parameter data of the power grid is collected at specific time intervals, meteorological data is obtained by calling the meteorological API at fixed cycles, user load data is aggregated by parsing the smart meter protocol, and equipment status data, including load rate, oil temperature and vibration amplitude, are monitored in real time through IoT sensors.

[0086] Using time window t as the time window, determine the average values ​​of electrical parameter data, meteorological data, user load data and equipment status data within the time window, and establish a four-dimensional data matrix;

[0087] Based on the CIM / E standard, the physical power grid is mapped as a node-edge model, and electrical parameters are embedded for each line.

[0088] Furthermore, a four-dimensional data matrix is ​​constructed, specifically as follows:

[0089] The four-dimensional data matrix includes a power grid operation parameter matrix, a meteorological parameter matrix, a user load matrix, and a fault record matrix. Each matrix has a dimension of T×N, where T is the total number of time windows and N is the total number of physical nodes in the power grid.

[0090] The matrix elements of the power grid operation parameter matrix are current, voltage, and active power;

[0091] The elements of the meteorological parameter matrix are regional temperature and humidity, wind speed, and light intensity;

[0092] The elements of the user load matrix are the real-time user loads;

[0093] The elements of the fault record matrix are encoded with 0 and 1, where 0 indicates that the equipment is normal and 1 indicates that the equipment is abnormal.

[0094] Based on the equipment status data, set the abnormal oil temperature threshold and abnormal vibration amplitude threshold. Within a time window, if either the average abnormal oil temperature or the average abnormal vibration amplitude is greater than or equal to its corresponding threshold, the equipment is judged to be abnormal within that time window; otherwise, the equipment is judged to be normal within that time window.

[0095] The equipment degradation model building module uses a federated learning framework to aggregate power grid data from multiple regions, build an equipment degradation model, and improve the priority data channel mechanism.

[0096] Furthermore, an equipment degradation model is established, specifically using the following method:

[0097] Deploy a federated learning architecture, set up a central aggregation server and regional grid clients, with the clients only uploading encrypted model gradients;

[0098] The server aggregates the encryption gradients of each region using a weighted average method. The weight of the aggregated encryption gradient is equal to the ratio of the number of local training samples in that region to the number of local training samples in all regions.

[0099] The global model parameters are updated based on the aggregated encryption gradient. The updated global model parameters are equal to the difference between the old global model parameters and the aggregated gradient. The aggregated gradient is the product of the aggregated encryption gradient and the learning rate. The updated global model parameters are then distributed to clients in each region.

[0100] When the equipment load rate is greater than the threshold H for n consecutive time windows, the twin resistance parameter is dynamically updated based on the remaining lifetime prediction value. The updated twin resistance parameter is equal to the combined value of the old resistance parameter and the remaining lifetime decay rate.

[0101] The risk coefficient is obtained by calculating the load intensity of the equipment and the degree of remaining life decay, and then weighting and summing them.

[0102] Furthermore, by determining the thresholds for equipment load rate and remaining lifetime degradation rate, the average equipment load rate of a certain time window is compared with the equipment load rate threshold to determine the equipment load intensity of that time window. By comparing the predicted remaining lifetime of the equipment in a certain time window with the equipment remaining lifetime in the previous time window, the remaining lifetime degradation rate of that time window is determined. The average remaining lifetime degradation rate of n consecutive time windows is used to represent the degradation rate of the equipment. The degradation rate of the equipment is compared with the remaining lifetime degradation rate threshold to determine the degree of remaining lifetime degradation in that time window. Different weight values ​​are assigned to the equipment load intensity and the degree of remaining lifetime degradation, and then the results are summed to obtain the risk coefficient.

[0103] Furthermore, the priority data channel mechanism will be improved, specifically through the following methods:

[0104] A three-tiered transmission assurance mechanism is constructed based on risk coefficients. Higher risk coefficient values ​​indicate a faster risk of equipment health deterioration. Risk coefficient thresholds A1 and A2 are set based on historical equipment failure data. Different equipment categories are then categorized according to these thresholds, and different transmission protocols are matched to each category. The specific transmission protocol matching method is as follows: Where x represents the x-th device, This represents the risk coefficient of device x. Let x represent the transmission protocol of device x. Every n*t time interval, the device attenuation rate is updated through federated learning, and the risk coefficient is recalculated.

[0105] The load scheduling optimization module determines the real-time power grid status and equipment risk coefficients based on the power grid digital twin, and designs a deep reinforcement learning dynamic decision-making framework to achieve multi-objective collaborative load scheduling optimization.

[0106] The design of a deep reinforcement learning dynamic decision-making framework is as follows:

[0107] Constructing the state space of a Markov decision process based on a four-dimensional data matrix of a power grid digital twin;

[0108] The basic action dimensions for dividing the action space are generator output adjustment dimension, energy storage charging and discharging control dimension, and interruptible load management dimension.

[0109] Based on the equipment risk coefficient, dynamic action constraint rules are established. When a certain equipment... Immediately lock the load increase action of the power supply line of the equipment;

[0110] when At that time, the output increment limit will be implemented;

[0111] when At that time, maintain the status quo;

[0112] Establish a power grid topology-action mapping table, assign a unique action index number to each physical node, and automatically expand the action space dimension when the power grid is expanded and new nodes are added;

[0113] Establish an action-effect mapping database and dynamically optimize the action space.

[0114] Furthermore, an action-effect mapping database is established to dynamically optimize the action space. The specific method is as follows:

[0115] Record the actual network loss change rate, equipment load rate fluctuation range, and risk coefficient change rate after the execution of historical actions;

[0116] Statistical analysis of the effectiveness scores of each action dimension is performed to dynamically eliminate inefficient actions, namely interruptible load commands that fail to improve network loss multiple times in a row.

[0117] Based on simulation results, a safe and feasible set of actions is selected to ensure that scheduling decisions meet the constraints of power grid operation.

[0118] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0119] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A smart power grid dispatching method based on digital twins and deep learning, characterized in that, include: Step 1: By integrating multi-source heterogeneous data with a dynamic parameter update mechanism, construct a digital twin of the power grid that is synchronously mapped with the physical power grid; Step 2: Use a federated learning framework to aggregate power grid data from multiple regions, establish an equipment degradation model, and improve the priority data channel mechanism; Deploy a federated learning architecture, set up a central aggregation server and regional grid clients, with the clients only uploading encrypted model gradients; The server aggregates the encryption gradients of each region using a weighted average method. The weight of the aggregated encryption gradient is equal to the ratio of the number of local training samples in that region to the number of local training samples in all regions. The global model parameters are updated based on the aggregated encryption gradient. The updated global model parameters are equal to the difference between the old global model parameters and the aggregated gradient. The aggregated gradient is the product of the aggregated encryption gradient and the learning rate. The updated global model parameters are then distributed to clients in each region. When the equipment load rate is greater than the threshold H for n consecutive time windows, the twin resistance parameter is dynamically updated based on the remaining lifetime prediction value. The updated twin resistance parameter is equal to the combined value of the old resistance parameter and the remaining lifetime decay rate. The risk coefficient is obtained by calculating the equipment load intensity and the degree of remaining life degradation, and then weighting and summing them. A three-tiered transmission assurance mechanism is constructed based on risk coefficients. Higher risk coefficient values ​​indicate a faster risk of equipment health deterioration. Risk coefficient thresholds A1 and A2 are set based on historical equipment failure data. Different equipment categories are then classified according to these thresholds, and different transmission protocols are matched to each category. The specific transmission protocol matching method is as follows: Where x represents the x-th device, This represents the risk coefficient of device x. The transmission protocol of device x is represented, and the device attenuation rate is updated through federated learning every n*t time intervals, and the risk coefficient is recalculated. Step 3: Based on the digital twin of the power grid, determine the real-time power grid status and equipment risk coefficients, design a deep reinforcement learning dynamic decision-making framework, and realize multi-objective collaborative load scheduling optimization.

2. The intelligent power grid dispatching method based on digital twins and deep learning according to claim 1, characterized in that, The specific method for constructing a digital twin of the power grid that is synchronously mapped with the physical power grid is as follows: Identify power grid equipment and lines, using power connection points or energy conversion devices in the physical power grid as nodes and energy transmission channels in the physical power grid as lines; Real-time electrical parameter data of the power grid is collected at specific time intervals, meteorological data is obtained by calling the meteorological API at fixed cycles, user load data is aggregated by parsing the smart meter protocol, and equipment status data, including load rate, oil temperature and vibration amplitude, is monitored in real time through IoT sensors. Using time window t as the time window, determine the average values ​​of electrical parameter data, meteorological data, user load data and equipment status data within the time window, and establish a four-dimensional data matrix; Based on the CIM / E standard, the physical power grid is mapped as a node-edge model, and electrical parameters are embedded for each line.

3. The intelligent power grid dispatching method based on digital twins and deep learning according to claim 2, characterized in that, The specific method for constructing a four-dimensional data matrix is ​​as follows: The four-dimensional data matrix includes a power grid operation parameter matrix, a meteorological parameter matrix, a user load matrix, and a fault record matrix. Each matrix has a dimension of T×N, where T is the total number of time windows and N is the total number of physical nodes in the power grid. The matrix elements of the power grid operation parameter matrix are current, voltage, and active power; The elements of the meteorological parameter matrix are regional temperature and humidity, wind speed, and light intensity; The elements of the user load matrix are the real-time user loads; The elements of the fault record matrix are encoded with 0 and 1, where 0 indicates that the equipment is normal and 1 indicates that the equipment is abnormal. Based on the equipment status data, set the abnormal oil temperature threshold and abnormal vibration amplitude threshold. Within a time window, if either the average abnormal oil temperature or the average abnormal vibration amplitude is greater than or equal to its corresponding threshold, the equipment is judged to be abnormal within that time window; otherwise, the equipment is judged to be normal within that time window.

4. The intelligent power grid dispatching method based on digital twins and deep learning according to claim 1, characterized in that, The risk coefficient is obtained by calculating the equipment load intensity and the degree of remaining life degradation, and then weighting and summing them. The specific method is as follows: The thresholds for equipment load rate and remaining lifespan degradation rate are determined. The average equipment load rate of a certain time window is compared with the equipment load rate threshold to determine the equipment load intensity of that time window. The remaining lifespan degradation rate of a certain time window is determined by comparing the predicted remaining lifespan of the equipment in a certain time window with the equipment remaining lifespan of the previous time window. The degradation rate of the equipment is represented by the average remaining lifespan degradation rate of n consecutive time windows. The degradation rate of the equipment is compared with the remaining lifespan degradation rate threshold to determine the degree of remaining lifespan degradation in that time window. Different weight values ​​are assigned to the equipment load intensity and the degree of remaining lifespan degradation, and then the results are summed to obtain the risk coefficient.

5. The intelligent power grid dispatching method based on digital twins and deep learning according to claim 1, characterized in that, Based on the determination of real-time power grid status and equipment risk coefficients using a digital twin of the power grid, a deep reinforcement learning dynamic decision-making framework is designed. The specific method is as follows: Constructing the state space of a Markov decision process based on a four-dimensional data matrix of a power grid digital twin; The basic action dimensions for dividing the action space are generator output adjustment dimension, energy storage charging and discharging control dimension, and interruptible load management dimension. Based on the equipment risk coefficient, dynamic action constraint rules are established. When a certain equipment... Immediately lock the load increase action of the power supply line of the equipment; when At that time, the output increment limit will be implemented; when At that time, maintain the status quo; Establish a power grid topology-action mapping table, assign a unique action index number to each physical node, and automatically expand the action space dimension when the power grid is expanded and new nodes are added; Establish an action-effect mapping database and dynamically optimize the action space.

6. The intelligent power grid dispatching method based on digital twins and deep learning according to claim 5, characterized in that, Establish an action-effect mapping database and dynamically optimize the action space. The specific method is as follows: Record the actual network loss change rate, equipment load rate fluctuation range, and risk coefficient change rate after the execution of historical actions; Statistical analysis of the effectiveness scores of each action dimension is performed to dynamically eliminate inefficient actions, namely interruptible load commands that fail to improve network loss multiple times in a row. Based on simulation results, a safe and feasible set of actions is selected to ensure that scheduling decisions meet the constraints of power grid operation.

7. A power grid intelligent dispatching system based on digital twins and deep learning, applied to the power grid intelligent dispatching method based on digital twins and deep learning as described in any one of claims 1-6, characterized in that, include: The power grid digital twin construction module constructs a power grid digital twin that is synchronously mapped to the physical power grid by integrating multi-source heterogeneous data and a dynamic parameter update mechanism. The equipment degradation model building module uses a federated learning framework to aggregate power grid data from multiple regions, build an equipment degradation model, and improve the priority data channel mechanism. The load scheduling optimization module determines the real-time power grid status and equipment risk coefficients based on the power grid digital twin, and designs a deep reinforcement learning dynamic decision-making framework to achieve multi-objective collaborative load scheduling optimization.

Images