Power grid intelligent Internet of Things platform management method and system
By decomposing the grid optimization problem into sub-problems and using the multi-layer perceptual DQN network model and ADMM algorithm, the problem of inflexible power grid management strategies in the existing technology is solved, and the global optimal scheduling and operation efficiency of the distribution network are improved. .
Patent Information
- Application Number
- CN202411260696.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-10
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-09-10
AI Technical Summary
When facing large-scale distribution network systems, the existing power grid management method has high computational complexity and is difficult to deal with the optimization problems of multiple constraints and multiple goals. The management strategy is not flexible enough, making it difficult to deal with the dynamic changes in grid load, power generation and faults. , resulting in waste of resources and unreasonable load allocation.
Decompose the distribution network optimization problem into multiple sub-problems, use multi-layer perceptual DQN network model and ADMM algorithm to learn and adjust in a dynamic environment, and coordinate the various sub-problems through ε-greedy strategy and Lagrangian multiplier updates. Scheduling plan to achieve global optimal scheduling.
It improves the operating efficiency and reliability of the distribution network, provides more flexible and intelligent decision-making solutions, coordinates the scheduling of each sub-problem, and enables the overall management to achieve global optimality.
Smart Images

Figure CN118798579B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power grid management, and in particular to a management method and system for a power grid intelligent Internet of Things platform. Background Art
[0002] In the power industry, the rapid development of smart grids has led to the rapid integration of information and communication technologies with power grid production and enterprise management at an unprecedented breadth and depth. As a key node in the distribution network system, the distribution room is responsible for distributing electricity to users. Traditional distribution rooms have problems such as daily inspections that are difficult to fully cover all areas and points, insufficient personnel and equipment inspections, incomplete environmental monitoring information records that cannot complete centralized monitoring and management, difficulty in mastering all equipment working conditions, and incomplete monitoring of system real-time operation conditions. From functional response to operation and maintenance management, they cannot match the construction requirements of smart grids.
[0003] Driven by the development of Internet technology, Internet of Things technology has also achieved vigorous development in the power system. Compared with traditional power grid management, the power grid management system based on Internet of Things technology has irreplaceable advantages in information acquisition, data integration, state perception and differentiated operation and maintenance.
[0004] The prior art has the following deficiencies:
[0005] 1. Existing management methods often have high computational complexity when facing large-scale distribution network systems. They are difficult to effectively handle multi-constraint and multi-objective optimization problems and are prone to falling into local optimal solutions. In addition, management strategies are usually based on preset rules and static models, which are difficult to cope with dynamic changes in grid load, power generation and faults, resulting in inflexible strategies and slow response speeds.
[0006] 2. Management methods usually focus on global optimization or local optimization. It is difficult to take into account both global and local coordination at the same time, which may lead to resource waste or unreasonable load distribution in some areas.
[0007] Based on this, the present invention proposes a management method and system for a power grid intelligent Internet of Things platform, which first decomposes the large-scale optimization problem into sub-problems that can be solved in parallel, and then learns and adjusts in a dynamic environment, adapts to changes in the power grid in real time, provides more flexible and intelligent decision-making solutions, and coordinates the scheduling solutions for each sub-problem, so that the overall management reaches the global optimum, thereby improving the operating efficiency and reliability of the distribution network. Summary of the invention
[0008] The purpose of the present invention is to provide a power grid intelligent Internet of Things platform management method and system to solve the deficiencies in the background technology.
[0009] In order to achieve the above object, the present invention provides the following technical solution: a management method for a power grid intelligent IoT platform, the management method comprising the following steps:
[0010] The management system decomposes the distribution network optimization problem into multiple sub-problems, inputs the current state information of each sub-problem into the multi-layer perception DQN network model, and the multi-layer perception DQN network model outputs a Q value for each sub-problem. Based on the Q value, the ε-greedy strategy is used to explore and find the initial scheduling strategy for the sub-problem.
[0011] The Lagrange multiplier of each subproblem is updated by the ADMM algorithm, and the initial scheduling strategy of each subproblem is adjusted according to the Lagrange multiplier to obtain the final scheduling strategy of each subproblem;
[0012] According to the final dispatching strategy of each sub-problem, the operation of the distribution network is simulated in the simulation environment, and the final dispatching strategy is adjusted and corrected through multiple iterations. After the convergence conditions are met, the global optimal dispatching strategy is output for each sub-problem and applied in the distribution network.
[0013] In a preferred embodiment, the current state information of each sub-problem is input into the multi-layer perception DQN network model, and the multi-layer perception DQN network model outputs a Q value for each sub-problem, including the following steps:
[0014] Define the current state information for each subproblem, including voltage level, load demand, generation capacity, and renewable energy output:
[0015] The input state information is normalized and then input into the multi-layer perception DQN network model;
[0016] The multilayer perceptron DQN network model calculates the Q value vector through forward propagation. Each Q value in the Q value vector represents the corresponding scheduling strategy selected in the current state, and the scheduling strategy with the largest Q value is selected as the scheduling strategy for the sub-problem.
[0017] In a preferred embodiment, the multi-layer perception DQN network model includes an input layer, a hidden layer and an output layer;
[0018] Input layer: receives the normalized state information vector;
[0019] Hidden layer: set to several, each hidden layer contains neurons, and ReLU activation function is used between hidden layers;
[0020] Output layer: Output a Q value vector, where each element corresponds to the Q value of a scheduling strategy;
[0021] The state information of the sub-problem is represented by a vector s with a dimension of n. The state vector received by the input layer is: ;
[0022] Assume that there is a multi-layer perceptron DQN network model hidden layers, the calculation of each hidden layer is expressed as: , where It is The output of the layer, Indicates layer hidden layer, It is The output of the layer, It is The weight matrix of the layer has the dimension ,in It is The number of neurons in the layer, It is The bias vector of the layer, dimension is , is the activation function;
[0023] In passing After hidden layers, the output layer will produce Value vector , the dimension is ,in is the number of scheduling strategies, expressed as: , is the weight matrix of the output layer, with dimension , is the bias vector of the output layer, with dimension , yes A vector of values containing the value;
[0024] For a certain scheduling strategy, The value calculation expression is: , where for value, is the corresponding scheduling strategy in the output layer No. The weights of the input neurons, It is Layer The activation value of a neuron, The corresponding scheduling strategy The bias.
[0025] In a preferred embodiment, based on The ε-greedy strategy is used to explore and find the initial scheduling strategy for the subproblem, which includes the following steps:
[0026] The probability ε is initially set to 0.9, and is reduced by 0.1 each iteration. A random number r between 0 and 1 is generated, and a scheduling strategy is selected based on the generated random number r. If the random number r is less than the probability ε, a scheduling strategy is randomly selected. If the random number r is greater than or equal to the probability ε, a scheduling strategy is selected. The scheduling strategy with the largest value;
[0027] Apply the selected scheduling strategy in the simulation environment, obtain the new state information of the sub-problem, and use the reward function to combine the selected scheduling strategy with the new state information. The value is updated, the expression is: , where For the updated value, Before update value, is the learning rate, and its value range is , is the reward function, is the discount factor, and its value range is , Indicates new status information The maximum of all scheduling strategies value;
[0028] Repeated iterations to select multiple scheduling strategies Value update, when probability ε=0.1, select The scheduling strategy with the largest value is used as the initial scheduling strategy.
[0029] In a preferred embodiment, the steps of obtaining the reward function are:
[0030] When applying the scheduling strategy in the simulation environment, the utilization rate of renewable energy, energy loss and constraint violation index are obtained, and the utilization rate of renewable energy, energy loss and constraint violation index are comprehensively calculated to obtain the reward function, which is expressed as: , where is the reward function, is the utilization rate of renewable energy, is the energy loss, To constrain the violation index, , , are the adjustment coefficients of renewable energy utilization rate, energy loss and constraint violation index, respectively, and , , Both are greater than 0.
[0031] In a preferred embodiment, the Lagrange multiplier of each subproblem is updated by the ADMM algorithm, and the initial scheduling strategy of each subproblem is adjusted according to the Lagrange multiplier to obtain the final scheduling strategy of each subproblem, including the following steps:
[0032] After obtaining the initial scheduling strategy for each subproblem, initialize the Lagrange multiplier of the subproblem ;
[0033] For each subproblem, according to the current Lagrange multiplier and the initial scheduling strategy of other subproblems, the updated scheduling strategy vector is obtained after solving the local optimization problem of the subproblem;
[0034] updating a global consistency variable based on the updated scheduling strategy vector, and updating a Lagrange multiplier based on the updated global consistency variable and the updated scheduling strategy vector;
[0035] Repeat the scheduling strategy vector update step. When the number of iterations is equal to the number threshold, it is determined that the convergence condition is met, and the scheduling strategy vector obtained by the last update of each sub-problem is output as the final scheduling strategy of the sub-problem.
[0036] In a preferred embodiment, after solving the local optimization problem of the subproblem, an updated scheduling strategy vector is obtained, and the function expression is: , where For the The sub-question is The scheduling strategy vector in the iteration, that is, the updated scheduling strategy vector, For the The sub-question is The scheduling strategy vector in the iteration, For the The sub-question is The Lagrange multiplier in the iteration, is the step size parameter, is a global consistency variable, which is the average value of the scheduling strategies of each sub-problem. For sub-problems The objective function is For sub-problems The distance between the scheduling strategy and the global consistency variable;
[0037] Update the global consistency variable based on the updated scheduling strategy vector. The expression is: , where is the updated global consistency variable, is the number of subproblems, is the updated scheduling strategy vector. The Lagrange multiplier is updated according to the updated global consistency variable and the updated scheduling strategy vector. The expression is: , where is the updated Lagrange multiplier, is the updated global consistency variable, is the updated scheduling strategy vector, is the step size parameter.
[0038] In a preferred embodiment, the constraint violation index is calculated as follows: , where For the state deviation value, is the number of states, and the calculation expression of the state deviation value is: , where is the state deviation value, Get the value of real-time status. The state constraint range.
[0039] In a preferred embodiment, the calculation expression of the renewable energy utilization rate is: , where is the total energy provided by renewable energy sources, is the total energy supply in the distribution network;
[0040] The calculation expression of the energy loss is: , where For the The power loss of the line, For Line The transmission time, is the number of lines.
[0041] The power grid intelligent IoT platform management system includes sub-problem division module, strategy output module and simulation module:
[0042] Sub-problem division module: decompose the distribution network optimization problem into multiple sub-problems;
[0043] Strategy output module: The current state information of each sub-problem is input into the multi-layer perception DQN network model. The multi-layer perception DQN network model outputs a Q value for each sub-problem. Based on the Q value, the ε-greedy strategy is used to explore and find the initial scheduling strategy for the sub-problem. The Lagrange multiplier of each sub-problem is updated through the ADMM algorithm. After adjusting the initial scheduling strategy of each sub-problem according to the Lagrange multiplier, the final scheduling strategy of each sub-problem is obtained.
[0044] Simulation module: Based on the final dispatching strategy of each sub-problem, the operation of the distribution network is simulated in the simulation environment. The final dispatching strategy is adjusted and corrected through multiple iterations. After the convergence conditions are met, the global optimal dispatching strategy is output for each sub-problem and applied in the distribution network.
[0045] In the above technical solution, the technical effects and advantages provided by the present invention are:
[0046] The present invention decomposes the distribution network optimization problem into multiple sub-problems, inputs the current state information of each sub-problem into the multi-layer perception DQN network model, and the multi-layer perception DQN network model outputs a Q value for each sub-problem, uses the ε-greedy strategy to explore based on the Q value to find the initial scheduling strategy for the sub-problem, updates the Lagrange multiplier of each sub-problem through the ADMM algorithm, and adjusts the initial scheduling strategy of each sub-problem according to the Lagrange multiplier to obtain the final scheduling strategy of each sub-problem. The management system first decomposes the large-scale optimization problem into sub-problems that can be solved in parallel, and then learns and adjusts in a dynamic environment, adapts to the changes of the power grid in real time, provides a more flexible and intelligent decision-making solution, coordinates the scheduling solution of each sub-problem, so that the overall management reaches the global optimum, and improves the operation efficiency and reliability of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.
[0048] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION
[0049] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0050] Example 1: Please refer to Figure 1 As shown, the power grid smart IoT platform management method described in this embodiment includes the following steps:
[0051] The management system decomposes the distribution network optimization problem into multiple sub-problems, each sub-problem corresponds to an area, and each sub-problem independently optimizes its own scheduling strategy. The scheduling strategy includes load distribution and energy management strategy. The current state information of each sub-problem is input into the multi-layer perception DQN network model. The multi-layer perception DQN network model outputs a Q value for each sub-problem. The ε-greedy strategy is used to explore based on the Q value to find the initial scheduling strategy for the sub-problem. The Lagrange multiplier of each sub-problem is updated by the ADMM algorithm. After adjusting the initial scheduling strategy of each sub-problem according to the Lagrange multiplier, the final scheduling strategy of each sub-problem is obtained, so that the initial scheduling strategy of each sub-problem gradually tends to a coordinated global optimal solution. According to the final scheduling strategy of each sub-problem, the operation of the distribution network is simulated in the simulation environment, and the final scheduling strategy is adjusted and corrected for multiple iterations. After meeting the convergence conditions, the global optimal scheduling strategy is output for each sub-problem and applied in the distribution network.
[0052] This application decomposes the distribution network optimization problem into multiple sub-problems, inputs the current state information of each sub-problem into the multi-layer perception DQN network model, and the multi-layer perception DQN network model outputs a Q value for each sub-problem. Based on the Q value, the ε-greedy strategy is used to explore and find the initial scheduling strategy for the sub-problem. The Lagrange multiplier of each sub-problem is updated by the ADMM algorithm, and the initial scheduling strategy of each sub-problem is adjusted according to the Lagrange multiplier to obtain the final scheduling strategy for each sub-problem. The management system first decomposes the large-scale optimization problem into sub-problems that can be solved in parallel, and then learns and adjusts in a dynamic environment, adapts to the changes of the power grid in real time, provides more flexible and intelligent decision-making solutions, and coordinates the scheduling solutions of each sub-problem, so that the overall management reaches the global optimal, and improves the operating efficiency and reliability of the distribution network.
[0053] Embodiment 2: The management system decomposes the distribution network optimization problem into multiple sub-problems, each sub-problem corresponds to a region, and each sub-problem independently optimizes its own dispatching strategy, which includes load distribution and energy management strategy, including the following steps:
[0054] The entire distribution network is divided into several regions, each of which is determined based on factors such as geographical location, load characteristics, and renewable energy distribution. Each region is optimized independently, and independent sub-problems are defined for each region, mainly including:
[0055] Load distribution strategy: Determines how the load in the area is distributed among various nodes to minimize energy loss and meet demand.
[0056] Energy management strategy: Manage energy resources (such as solar energy, wind energy, etc.) in the region to maximize the use of renewable energy and ensure voltage stability while meeting load demand.
[0057] Define a state vector for each region, which includes:
[0058] Voltage level: The voltage value of each node.
[0059] Load demand: the demand at each load node.
[0060] Power generation capacity: the real-time power generation of each power generation source in the region.
[0061] Renewable energy output: The output of renewable energy such as solar energy and wind energy in the region.
[0062] The current state information of each sub-problem is input into the multi-layer perception DQN network model, and the multi-layer perception DQN network model outputs a Q value for each sub-problem, including the following steps:
[0063] Define the current state information for each subproblem, which usually includes voltage level, load demand, generation capacity, and renewable energy output:
[0064] Voltage level: The voltage value of each node in the area.
[0065] Load demand: the real-time demand of each load node.
[0066] Power generation capacity: the real-time power generation of each power generation source in the region.
[0067] Renewable energy output: The current output of renewable energy sources such as solar and wind.
[0068] Normalize the input state information to make it suitable for input into the neural network. This step usually includes:
[0069] Data normalization: converting state information of different dimensions into a unified range (e.g. [0, 1]).
[0070] Data preprocessing: Smoothing or filtering noisy data to improve the input quality of the model.
[0071] The multilayer perceptron DQN network model calculates the Q value vector through forward propagation. Each Q value in the Q value vector represents the expected long-term return after selecting the corresponding scheduling strategy under the current state. For each sub-problem, the scheduling strategy with the largest Q value is selected as the scheduling decision under the current state.
[0072] The multi-layer perception DQN network model includes an input layer, a hidden layer, and an output layer;
[0073] Input layer: receives the normalized state information vector.
[0074] Hidden layer: set to several, each hidden layer contains neurons, and ReLU (Rectified-Linear-Unit) activation function can be used between hidden layers. The number of hidden layers and the number of neurons can be adjusted according to the complexity of the problem.
[0075] Output layer: Outputs a Q value vector, where each element corresponds to the Q value of a possible scheduling strategy.
[0076] The state information of the sub-problem is represented by a vector s with a dimension of n. The state vector received by the input layer is: ;
[0077] Assume that there is a multi-layer perceptron DQN network model hidden layers, the calculation of each hidden layer is expressed as: , where It is The output of the layer, Indicates layer hidden layer, It is The output of the layer (for the first layer ), It is The weight matrix of the layer has the dimension ,in It is The number of neurons in the layer, It is The bias vector of the layer, dimension is , is the activation function (usually using ).
[0078] In passing After hidden layers, the output layer will produce Value vector , the dimension is ,in is the number of scheduling strategies, expressed as: ,in, is the weight matrix of the output layer, with dimension , is the bias vector of the output layer, with dimension , yes A vector of values containing the values for each possible scheduling strategy value.
[0079] For a certain scheduling strategy, The value calculation expression is: , where for value, is the corresponding scheduling strategy in the output layer No. The weights of the input neurons, It is Layer The activation value of a neuron, The corresponding scheduling strategy The bias.
[0080] based on The ε-greedy strategy is used to explore and find the initial scheduling strategy for the subproblem, which includes the following steps:
[0081] In the ε-greedy strategy, the multi-layer perceptron DQN network model selects a random scheduling strategy (exploration) with probability ε each time it selects a scheduling strategy, and selects the scheduling strategy with the highest estimated Q value (utilization) with probability. This helps to avoid falling into the local optimal solution while finding the optimal scheduling strategy.
[0082] The probability ε is initially set to 0.9, and is reduced by 0.1 each iteration. A random number r between 0 and 1 is generated to determine whether this decision is to explore or exploit. A scheduling strategy is selected based on the generated random number r. If the random number r is less than the probability ε, a scheduling strategy is randomly selected. If the random number r is greater than or equal to the probability ε, a scheduling strategy is selected. The scheduling strategy with the largest value;
[0083] Apply the selected scheduling strategy in the simulation environment to obtain the new state information of the sub-problem. For example, the execution of the scheduling strategy may cause the voltage, load demand, power generation and other parameters of each node to change. The changed voltage, load demand, power generation and other parameters are used as the new state information, and the selected scheduling strategy is combined with the new state information using the reward function. The value is updated, the expression is: , where For the updated value, Before update value, is the learning rate, and its value range is , is the reward function, is the discount factor, and its value range is , Indicates new status information The maximum of all scheduling strategies value;
[0084] Repeated iterations to select multiple scheduling strategies Value update, when probability ε=0.1, select The scheduling strategy with the largest value is used as the initial scheduling strategy.
[0085] In this application, the steps for obtaining the reward function are:
[0086] When applying the scheduling strategy in the simulation environment, the utilization rate of renewable energy, energy loss and constraint violation index are obtained, and the utilization rate of renewable energy, energy loss and constraint violation index are comprehensively calculated to obtain the reward function, which is expressed as: , where is the reward function, is the utilization rate of renewable energy, is the energy loss, To constrain the violation index, , , are the adjustment coefficients of renewable energy utilization rate, energy loss and constraint violation index, respectively, and , , Both are greater than 0.
[0087] The larger the value of the reward function, the better the overall performance of the scheduling strategy.
[0088] The calculation expression of renewable energy utilization rate is: , where is the total energy provided by renewable energy sources (kWh), It is the total energy supply in the distribution network (including energy from all energy sources, kilowatt-hours). The greater the utilization rate of renewable energy, the more efficient the scheduling strategy is in utilizing renewable energy, and the more it should be used.
[0089] The calculation expression of energy loss is: , where For the The power loss of the line, For Line The transmission time, is the number of lines. The greater the energy loss, the greater the energy consumption caused by the scheduling strategy, and the less likely it should be used.
[0090] The calculation expression of constraint violation index is: , where For the state deviation value, is the number of states, and the calculation expression of the state deviation value is: , where is the state deviation value, Get the value of real-time status. is the state constraint range. The larger the constraint violation index is, the greater the deviation of the state information from the constraint condition is caused by the scheduling strategy in the simulation environment, and the less likely it is to be used.
[0091] During initial training, the management system may randomly select different scheduling strategies. Even though these scheduling strategies may not seem optimal in the short term (such as resulting in slightly higher energy consumption), these explorations can help the network better understand the relationship between the state space and actions.
[0092] The Lagrange multiplier of each subproblem is updated by the ADMM algorithm, and the initial scheduling strategy of each subproblem is adjusted according to the Lagrange multiplier to obtain the final scheduling strategy of each subproblem, so that the initial scheduling strategy of each subproblem gradually tends to a coordinated global optimal solution, including the following steps:
[0093] After obtaining the initial scheduling strategy for each subproblem, initialize the Lagrange multiplier of the subproblem
[0094] For each subproblem, according to the current Lagrange multiplier and the initial scheduling strategy of other subproblems, the updated scheduling strategy vector is obtained by independently solving the local optimization problem of the subproblem. The function expression is: , where For the The sub-question is The scheduling strategy vector in the iteration, that is, the updated scheduling strategy vector, For the The sub-question is The scheduling strategy vector in the iteration, For the The sub-question is The Lagrange multiplier in the iteration, is the step size parameter, is a global consistency variable, which is the average value of the scheduling strategies of each sub-problem. For sub-problems The objective function is For sub-problems The distance between the scheduling strategy and the global consistency variable.
[0095] Update the global consistency variable based on the updated scheduling strategy vector. The expression is: , where is the updated global consistency variable, is the number of subproblems, is the updated scheduling strategy vector.
[0096] The Lagrange multiplier is updated according to the updated global consistency variable and the updated scheduling strategy vector, and the expression is: , where is the updated Lagrange multiplier, is the updated global consistency variable, is the updated scheduling strategy vector, is the step size parameter;
[0097] Repeat the scheduling strategy vector update step. When the number of iterations is equal to the number threshold, it is determined that the convergence condition is met, and the scheduling strategy vector obtained by the last update of each sub-problem is output as the final scheduling strategy of the sub-problem.
[0098] According to the final dispatching strategy of each sub-problem, the operation of the distribution network is simulated in the simulation environment, and the final dispatching strategy is adjusted and corrected for multiple iterations. After the convergence conditions are met, the global optimal dispatching strategy is output for each sub-problem and applied in the distribution network, including the following steps:
[0099] Import the final dispatching strategy of each sub-problem into the simulation environment as the starting point of the simulation, run the distribution network in the simulation environment, simulate the operation of the distribution network within a certain time range according to the imported dispatching strategy, and adjust and correct the dispatching strategy according to the simulation results and data analysis. For example, if the energy loss in a certain area is high, its load distribution strategy can be adjusted, and the corrected dispatching strategy can be re-imported into the simulation environment, and the next simulation iteration can be prepared. Repeat the process of simulation operation and dispatching strategy adjustment. After each iteration, the strategy will be gradually optimized to get closer to the global optimal solution. When the Q value of the final dispatching strategy is greater than or equal to the Q value threshold, it is judged that the convergence condition is met, and the global optimal dispatching strategy is output for each sub-problem and applied in the distribution network.
[0100] Embodiment 3: The power grid intelligent IoT platform management system described in this embodiment includes a sub-problem division module, a strategy output module, and a simulation module:
[0101] Sub-problem division module: decomposes the distribution network optimization problem into multiple sub-problems. Each sub-problem corresponds to a region. Each sub-problem independently optimizes its own dispatching strategy, which includes load distribution and energy management strategy. The sub-problem division results are sent to the strategy output module.
[0102] Strategy output module: The current state information of each sub-problem is input into the multi-layer perception DQN network model. The multi-layer perception DQN network model outputs the Q value for each sub-problem. Based on the Q value, the ε-greedy strategy is used to explore and find the initial scheduling strategy for the sub-problem. The Lagrange multiplier of each sub-problem is updated through the ADMM algorithm. After adjusting the initial scheduling strategy of each sub-problem according to the Lagrange multiplier, the final scheduling strategy of each sub-problem is obtained, so that the initial scheduling strategy of each sub-problem gradually tends to a coordinated global optimal solution. The final scheduling strategy information of the sub-problem is sent to the simulation module.
[0103] Simulation module: Based on the final dispatching strategy of each sub-problem, the operation of the distribution network is simulated in the simulation environment. The final dispatching strategy is adjusted and corrected through multiple iterations. After the convergence conditions are met, the global optimal dispatching strategy is output for each sub-problem and applied in the distribution network.
[0104] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.
[0105] It should be understood that the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. A and B can be singular or plural. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship, but it may also indicate an "and / or" relationship. Please refer to the context for specific understanding.
[0106] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.
[0107] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.
[0108] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
Claims
1. A method for managing a power grid intelligent IoT platform, characterized in that: The management method comprises the following steps: The management system decomposes the distribution network optimization problem into multiple sub-problems, inputs the current state information of each sub-problem into the multi-layer perception DQN network model, and the multi-layer perception DQN network model outputs a Q value for each sub-problem. Based on the Q value, the ε-greedy strategy is used to explore and find the initial scheduling strategy for the sub-problem. The Lagrange multiplier of each subproblem is updated by the ADMM algorithm, and the initial scheduling strategy of each subproblem is adjusted according to the Lagrange multiplier to obtain the final scheduling strategy of each subproblem; According to the final dispatching strategy of each sub-problem, the operation of the distribution network is simulated in the simulation environment, and the final dispatching strategy is adjusted and corrected through multiple iterations. After the convergence conditions are met, the global optimal dispatching strategy is output for each sub-problem and applied in the distribution network. based on The ε-greedy strategy is used to explore and find the initial scheduling strategy for the subproblem, including the following steps: The probability ε is initially set to 0.9, and is reduced by 0.1 each iteration. A random number r between 0 and 1 is generated, and a scheduling strategy is selected based on the generated random number r. If the random number r is less than the probability ε, a scheduling strategy is randomly selected. If the random number r is greater than or equal to the probability ε, a scheduling strategy is selected. The scheduling strategy with the largest value; Apply the selected scheduling strategy in the simulation environment, obtain the new state information of the sub-problem, and use the reward function to combine the selected scheduling strategy with the new state information. The value is updated, the expression is: , where For the updated value, Before update value, is the learning rate, and its value range is , is the reward function, is the discount factor, and its value range is , Indicates new status information The maximum of all scheduling strategies value; Repeated iterations to select multiple scheduling strategies Value update, when probability ε=0.1, select The scheduling strategy with the largest value is used as the initial scheduling strategy; The current state information of each sub-problem is input into the multi-layer perception DQN network model, and the multi-layer perception DQN network model outputs a Q value for each sub-problem, including the following steps: Define the current state information for each subproblem, including voltage level, load demand, generation capacity, and renewable energy output: The input state information is normalized and then input into the multi-layer perception DQN network model; The multilayer perceptron DQN network model calculates the Q value vector through forward propagation. Each Q value in the Q value vector represents the corresponding scheduling strategy selected in the current state. The scheduling strategy with the largest Q value is selected as the scheduling strategy for the sub-problem. The multi-layer perception DQN network model includes an input layer, a hidden layer and an output layer; Input layer: receives the normalized state information vector; Hidden layer: set to several, each hidden layer contains neurons, and ReLU activation function is used between hidden layers; Output layer: Output a Q value vector, where each element corresponds to the Q value of a scheduling strategy; The state information of the sub-problem is represented by a vector s with a dimension of n. The state vector received by the input layer is: ; Assume that there is a multi-layer perceptron DQN network model hidden layers, the calculation of each hidden layer is expressed as: , where It is The output of the layer, Indicates layer hidden layer, It is The output of the layer, It is The weight matrix of the layer has the dimension ,in It is The number of neurons in the layer, It is The bias vector of the layer, dimension is , is the activation function; In passing After hidden layers, the output layer will produce Value vector , the dimension is ,in is the number of scheduling strategies, expressed as: , is the weight matrix of the output layer, with dimension , is the bias vector of the output layer, with dimension , yes A vector of values containing the value; For a certain scheduling strategy, The value calculation expression is: , where for value, is the corresponding scheduling strategy in the output layer No. The weights of the input neurons, It is Layer The activation value of a neuron, The corresponding scheduling strategy Bias of The steps for obtaining the reward function are: When applying the scheduling strategy in the simulation environment, the utilization rate of renewable energy, energy loss and constraint violation index are obtained, and the utilization rate of renewable energy, energy loss and constraint violation index are comprehensively calculated to obtain the reward function, which is expressed as: , where is the reward function, is the utilization rate of renewable energy, is the energy loss, To constrain the violation index, , , are the adjustment coefficients of renewable energy utilization rate, energy loss and constraint violation index, respectively, and , , Both are greater than 0.
2. The method for managing a power grid intelligent IoT platform according to claim 1, characterized in that: The Lagrange multiplier of each subproblem is updated by the ADMM algorithm, and the initial scheduling strategy of each subproblem is adjusted according to the Lagrange multiplier to obtain the final scheduling strategy of each subproblem, including the following steps: After obtaining the initial scheduling strategy for each subproblem, initialize the Lagrange multiplier of the subproblem ; For each subproblem, according to the current Lagrange multiplier and the initial scheduling strategy of other subproblems, the updated scheduling strategy vector is obtained after solving the local optimization problem of the subproblem; updating a global consistency variable based on the updated scheduling strategy vector, and updating a Lagrange multiplier based on the updated global consistency variable and the updated scheduling strategy vector; Repeat the scheduling strategy vector update step. When the number of iterations is equal to the number threshold, it is determined that the convergence condition is met, and the scheduling strategy vector obtained by the last update of each sub-problem is output as the final scheduling strategy of the sub-problem.
3. The method for managing a power grid intelligent IoT platform according to claim 2, characterized in that: After solving the local optimization problem of the subproblem, the updated scheduling strategy vector is obtained. The function expression is: , where For the The sub-question is The scheduling strategy vector in the iteration, that is, the updated scheduling strategy vector, For the The sub-question is The scheduling strategy vector in the iteration, For the The sub-question is The Lagrange multiplier in the iteration, is the step size parameter, is a global consistency variable, which is the average value of the scheduling strategies of each sub-problem. For sub-problems The objective function is For sub-problems The distance between the scheduling strategy and the global consistency variable; Update the global consistency variable based on the updated scheduling strategy vector. The expression is: , where is the updated global consistency variable, is the number of subproblems, is the updated scheduling strategy vector. The Lagrange multiplier is updated according to the updated global consistency variable and the updated scheduling strategy vector. The expression is: , where is the updated Lagrange multiplier, is the updated global consistency variable, is the updated scheduling strategy vector, is the step size parameter.
4. The method for managing a power grid intelligent IoT platform according to claim 3, characterized in that: The calculation expression of the constraint violation index is: , where For the state deviation value, is the number of states, and the calculation expression of the state deviation value is: , where is the state deviation value, Get the value of real-time status. The state constraint range.
5. The method for managing a power grid intelligent IoT platform according to claim 4, characterized in that: The calculation expression of the renewable energy utilization rate is: , where is the total energy provided by renewable energy sources, is the total energy supply in the distribution network; The calculation expression of the energy loss is: , where For the The power loss of the line, For Line The transmission time, is the number of lines.
6. A power grid intelligent IoT platform management system, used to implement the management method according to any one of claims 1 to 5, characterized in that: Including sub-problem division module, strategy output module, simulation module: Sub-problem division module: decompose the distribution network optimization problem into multiple sub-problems; Strategy output module: The current state information of each sub-problem is input into the multi-layer perception DQN network model. The multi-layer perception DQN network model outputs a Q value for each sub-problem. Based on the Q value, the ε-greedy strategy is used to explore and find the initial scheduling strategy for the sub-problem. The Lagrange multiplier of each sub-problem is updated through the ADMM algorithm. After adjusting the initial scheduling strategy of each sub-problem according to the Lagrange multiplier, the final scheduling strategy of each sub-problem is obtained. Simulation module: Based on the final dispatching strategy of each sub-problem, the operation of the distribution network is simulated in the simulation environment. The final dispatching strategy is adjusted and corrected through multiple iterations. After the convergence conditions are met, the global optimal dispatching strategy is output for each sub-problem and applied in the distribution network.
Citation Information
Patent Citations
Micro-grid energy optimization scheduling method
CN118174355A
Microgrid space-time perception energy management method based on secure deep reinforcement learning
WO2024108817A1
