DMMoEQN-based power system parameter identification method and system

Abstract

The invention relates to the field of power systems, and discloses a DMMoEQN-based power system parameter identification method, which comprises the following steps: acquiring an initial observation value # imgabs0 # and a target observation value # imgabs1 # of a power system, and selecting an optimal parameter adjustment action # imgabs3 # by an intelligent agent according to a current observation value # imgabs2 # through an online network; an action # imgabs4 # is executed, the electric power system updates a parameter state # imgabs6 # of the electric power system according to the action # imgabs5 #, and a new observation value # imgabs7 # is obtained; calculating a step reward according to an error between the new observation value # imgabs8 # and the target observation value # imgabs9 #, and obtaining a total reward according to the step reward; judging whether the observation value # imgabs10 # of the electric power system is within the error range of the target observation value # imgabs11 #, if so, performing initialization, and continuing training; when the epoch number is larger than the preset maximum epoch number, training is ended, a power system parameter set # imgabs12 # is obtained through the trained power system parameter identification model, and the power system is adjusted according to the parameter set # imgabs13 #. According to the invention, efficient and accurate parameter identification and dynamic adjustment of the power system can be realized.

Description

Technical Field

[0001] The present invention relates to the field of power systems, and in particular to a method for identifying power system parameters based on DMMoEQN. Background Art

[0002] As the scale and complexity of modern power systems continue to increase, power system parameter identification (PI) has become particularly important in ensuring system stability and reliability. Its principle is to determine the unknown parameters in the system model by observing the input and output data of the system. Existing power system parameter identification technologies mainly include methods based on traditional statistical methods, numerical optimization methods, and methods based on machine learning and deep learning.

[0003] 1. Traditional statistical methods Traditional statistical methods include least squares, maximum likelihood estimation, and Kalman filtering. These methods estimate system parameters by building mathematical models and combining them with observed data. For example, the least squares method estimates parameters by minimizing the error between observed values ​​and model predicted values. However, these methods often show poor adaptability and computational efficiency when faced with nonlinear, large-scale, and dynamically changing power systems.

[0004] 2. Numerical Optimization Methods Numerical optimization methods use optimization algorithms to find the optimal parameter values ​​for the objective function. Common optimization algorithms include genetic algorithms, particle swarm optimization algorithms, and simulated annealing algorithms. These methods have certain advantages in dealing with complex nonlinear problems, but in practical applications, they often have problems such as slow convergence and easy to fall into local optimal solutions.

[0005] 3. Methods based on machine learning and deep learning In recent years, with the increase in computing power and data volume, machine learning and deep learning methods have been increasingly used in power system parameter identification. Typical methods include support vector machines, random forests, neural networks, etc. These methods can accurately estimate parameters by learning the complex relationships of the system from a large amount of data. However, deep learning methods usually require a large amount of training data and computing resources, and the model training process is complex and consumes a lot of computing resources, making them difficult to apply in resource-constrained environments.

[0006] Problems and shortcomings Although the existing technology has achieved certain results in power system parameter identification, there are still the following main problems and shortcomings: 1. Insufficient ability to cope with dynamic changes: Existing methods have difficulty in obtaining accurate parameter estimates when faced with the dynamic changes and nonlinear characteristics of power systems. Power system parameters may change frequently due to environmental changes, load fluctuations and other factors. Existing methods are difficult to adjust in real time and maintain accurate parameter estimates.

[0007] 2. Poor multi-parameter identification effect: Existing methods often ignore the correlation between different parameters when dealing with multi-parameter identification problems, resulting in high computational complexity, low efficiency, and inaccurate parameter estimation. Faced with the situation where multiple parameters in the power system change simultaneously, existing methods are difficult to deal with effectively, often resulting in unsatisfactory overall identification results.

[0008] 3. High computational complexity: Traditional statistical methods and numerical optimization methods have high computational complexity and slow convergence speed when processing large-scale and high-dimensional data, and it is difficult to meet the requirements of real-time applications. Summary of the invention

[0009] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for identifying power system parameters based on DMMoEQN, comprising the following steps: Step 1: Set the initial parameters of the power system and the target parameters of the power system, and obtain the target observation value , and create a memory pool; Step 2: Obtain the initial observation value of the power system and the target observation , the agent uses the online network to calculate the current observation value Select the best parameter tuning action ; Execute action , the power system operates according to Update the parameter status of the power system , and get new observations ; Step 3: Based on the new observations and the target observation Calculate the step reward based on the error between them, and get the total reward based on the step reward; Step 4: Record the current parameter status ,action ,award and the next parameter status , and stored in the memory pool; Step 5: Determine the observed value of the power system Is it in the target observation value? If the error range is within the error range, update the target value and target observation value , and continue to iterate until the predetermined target value n is reached, and then go to step eight; otherwise, go to step six; Step 6: Sample from the memory pool, update the parameters of the online network by maximizing the total reward based on the collected samples, and copy the parameters of the online network to the target network within the set time interval; Step 7, repeating steps 2 to 6. When the predetermined maximum number of iterations or the predetermined target value n is reached, the epoch ends and proceeds to step 8. Step eight, perform initialization and continue training; when the epoch number is greater than the preset maximum epoch number, the training ends and a trained power system parameter identification model is obtained. The power system parameters are obtained through the trained power system parameter identification model, and the power system is adjusted according to the power system parameters.

[0010] Furthermore, the agent uses an online network to Select the best parameter tuning action ; Execute action , the power system operates according to Update the parameter status of the power system , and get new observations ,include: The intelligent agent includes a target network and an online network based on MMoE; the intelligent agent is based on the current power system state observation value Select and execute the best parameter tuning action ; In each iteration, the power system updates the parameter state of the power system according to the actions of the intelligent agent , and get new observations .

[0011] Furthermore, the new observations and the target observation The step reward is calculated based on the error between , and the total reward is obtained based on the step reward, including: Calculate the two-norm error between the current observation and the target observation; convert the two-norm error into an immediate reward, the immediate reward is: in is the current observation value, is the target observation value; accumulate all immediate rewards to get the total reward.

[0012] Furthermore, the sampling from the memory pool and updating the parameters of the online network by maximizing the total reward according to the collected samples include: For each sample, the current Q value and the target Q value are calculated; the timing difference error is calculated, and the timing difference error is: Where r is the immediate reward, is the discount factor, Q' is the Q value of the target network; The mean square error loss function is used to minimize the temporal difference error. The mean square error loss function is: in is the TD error of the i-th sample; back-propagates the error and updates the network weights; periodically copies the parameters of the online network to the target network.

[0013] Furthermore, the observation value of the power system is determined Is it in the target observation value? If the error range is within the error range, update the target value and target observation value ,include: At the beginning of training, N target values ​​are preset. After reaching the target observation value corresponding to the current target value, the next target value is switched. A new target observation value is obtained through simulation, and the new target observation value is set as the current target. The record of the target value is updated until all the preset N target values ​​are reached.

[0014] A DMMoEQN-based power system parameter identification system, applying the DMMoEQN-based power system parameter identification method, including a power system parameter identification model module, a power system parameter acquisition module, a data processing module and a data storage module; The power system parameter identification model module, the power system parameter acquisition module and the data storage module are respectively connected to the data processing module.

[0015] The beneficial effects of the present invention are: 1. Improve the ability to identify frequently changing parameters: Existing technologies are difficult to cope with the situation where parameters in power systems frequently change due to environmental changes and load fluctuations. The DMMoEQN method introduces a continuous target value tracking mechanism for intelligent agents, which enables the system to quickly respond to changes in environmental and operating conditions, dynamically adjust parameters, and achieve real-time tracking and efficient identification of frequently changing parameters. Experiments show that in a simulation environment, this method can quickly adjust after the target parameters change, and the identification accuracy is significantly improved.

[0016] 2. Enhance the ability of simultaneous identification of multiple parameters: The simultaneous change of multiple parameters is a common phenomenon in power systems. Existing methods often ignore the correlation between parameters when dealing with multi-parameter identification, resulting in poor identification results. By introducing the multi-task learning capability of the MMoE model into DDQN, the present invention can simultaneously process the identification of multiple parameters and improve the overall identification accuracy and efficiency. Experimental results show that in complex scenarios with multiple parameter changes, this method can effectively reduce parameter identification errors.

[0017] 3. Reduce computational complexity and improve response speed: Traditional statistical methods and numerical optimization methods have high computational complexity and slow convergence speed when processing large-scale and high-dimensional data, and it is difficult to meet the requirements of real-time applications. The present invention uses the DDQN reinforcement learning framework based on MMoE, and through the dual structure of online network and target network, it effectively reduces the computational complexity, speeds up the parameter identification process, and improves the real-time response capability of the system. According to experimental data, the total computation time of the DMMoEQN method in large-scale data processing is significantly reduced, and the response speed is significantly improved.

[0018] 4. Improve the accuracy and stability of identification: By updating the target network regularly, the present invention ensures the stability of the training process, reduces the fluctuation and instability in the parameter adjustment process, and thus improves the accuracy and stability of parameter identification. Experimental data show that the root mean square error (RMSE) of this method remains at a low level under different test environments, and the identification results are more reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of a power system parameter identification method based on DMMoEQN. DETAILED DESCRIPTION

[0020] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0021] The features and performance of the present invention are further described in detail below in conjunction with the embodiments.

[0022] like Figure 1 As shown, a method for identifying power system parameters based on DMMoEQN includes the following steps: Step 1: Set the initial parameters of the power system and the target parameters of the power system, and obtain the target observation value , and create a memory pool; Step 2: Obtain the initial observation value of the power system and the target observation , the agent uses the online network to calculate the current observation value Select the best parameter tuning action ; Execute action , the power system operates according to Update the parameter status of the power system , and get new observations ; Step 3: Based on the new observations and the target observation Calculate the step reward based on the error between them, and get the total reward based on the step reward; Step 4: Record the current parameter status ,action ,award and the next parameter status , and stored in the memory pool; Step 5: Determine the observed value of the power system Is it in the target observation value? If the error range is within the error range, update the target value and target observation value , and continue to iterate until the predetermined target value n is reached, and then go to step eight; otherwise, go to step six; Step 6: Sample from the memory pool, update the parameters of the online network by maximizing the total reward based on the collected samples, and copy the parameters of the online network to the target network within the set time interval; Step 7, repeating steps 2 to 6. When the predetermined maximum number of iterations or the predetermined target value n is reached, the epoch ends and proceeds to step 8. Step eight, perform initialization and continue training; when the epoch number is greater than the preset maximum epoch number, the training ends and a trained power system parameter identification model is obtained. The power system parameters are obtained through the trained power system parameter identification model, and the power system is adjusted according to the power system parameters.

[0023] The agent uses the online network to calculate the current observation value. Select the best parameter tuning action ; Execute action , the power system operates according to Update the parameter status of the power system , and get new observations ,include: The intelligent agent includes a target network and an online network based on MMoE; the intelligent agent is based on the current power system state observation value Select and execute the best parameter tuning action ; In each iteration, the power system updates the parameter state of the power system according to the actions of the intelligent agent , and get new observations .

[0024] According to the new observations and the target observation The step reward is calculated based on the error between , and the total reward is obtained based on the step reward, including: Calculate the two-norm error between the current observation and the target observation; convert the two-norm error into an immediate reward, the immediate reward is: in is the current observation value, is the target observation value; accumulate all immediate rewards to get the total reward.

[0025] The sampling from the memory pool and updating the parameters of the online network by maximizing the total reward based on the collected samples include: For each sample, the current Q value and the target Q value are calculated; the timing difference error is calculated, and the timing difference error is: Where r is the immediate reward, is the discount factor, Q' is the Q value of the target network; The mean square error loss function is used to minimize the temporal difference error. The mean square error loss function is: in is the TD error of the i-th sample; back-propagates the error and updates the network weights; periodically copies the parameters of the online network to the target network.

[0026] The observation value of the power system is determined Is it in the target observation value? If the error range is within the error range, update the target value and target observation value ,include: At the beginning of training, N target values ​​are preset. After reaching the target observation value corresponding to the current target value, the next target value is switched. A new target observation value is obtained through simulation, and the new target observation value is set as the current target. The record of the target value is updated until all the preset N target values ​​are reached.

[0027] Specifically, the agent interacts with the environment Agent: It includes the target network and online network based on MMoE. The agent is responsible for selecting the best parameters to adjust the action according to the current environment state at each moment, and improving the stability and performance of the model by regularly updating the parameters of the target network.

[0028] Environment: Represents the power system. In each iteration, the environment updates the power system state according to the agent's actions and provides corresponding reward signals and new system states.

[0029] 2. Parameter identification process Initial environment setting: In the initial environment, the initial parameters of the power system are set and the initial state of the system is observed through simulation. This step ensures that the agent starts learning under known initial conditions.

[0030] Action selection and execution: The agent selects an optimal action based on the current environmental state through the online network. The action represents the adjustment of a certain power system parameter. After the action is selected, the agent executes the action, causing the power system state to change.

[0031] Reward calculation: The environment calculates the corresponding reward value based on the action performed by the agent through the current observation value (Current Observation) and the target observation value (Target Observation). The reward value includes the step reward (StepReward) and the total reward (Total Reward), which are used to evaluate the quality of the action.

[0032] 3. Target update and loop iteration Target update: When the power system state reaches the current target value, the new target value is updated. The agent needs to continuously track and adjust the power system state to gradually approach and reach the preset target value each time until the preset n target values ​​are completed.

[0033] Iteration: The whole process is an iterative process. In each cycle, the agent interacts with the environment, selects actions, executes actions, obtains rewards, and updates network parameters through the memory pool. Through multiple iterations, the agent continuously optimizes its strategy and gradually improves the accuracy of power system parameter identification.

[0034] Memory pool storage: The results of each iteration (including state, action, reward, and next state) are stored in the database. These data are not only used for current parameter optimization, but also for subsequent analysis and improvement.

[0035] 2. Model and Implementation Principle 1. Input representation The state s of the power system is represented as a vector , where n is the dimension of the state space, and each dimension represents a parameter to be identified.

[0036] 2. Expert Network and Gating Layer In this model, expert networks and gating layers are introduced to handle different types of parameter adjustment relationships. The underlying representation of each parameter to be identified is connected to the expert network, and the gating network is used to determine how each parameter adjustment amount utilizes these expert networks.

[0037] (1) Expert Network The expert network consists of a set of underlying neural networks, represented as ,in , m is the number of experts. Each expert network is responsible for learning a part of the task space, and its output is expressed as : (2) Gating layer The gating layer consists of a set of gating networks, represented as ,in , n is the number of parameters to be identified. Each gating network receives input features of parameter states s and outputs a set of softmax gating , used to weight the output of the expert network to generate each parameter adjustment Representation: in, and are the weight matrix and bias vector of the i-th gating network, is the gating weight of the jth expert network corresponding to the i-th parameter.

[0038] 3. Tower Each parameter i corresponds to a tower layer, and the input of each tower layer The weighted sum of the output of the expert network is obtained: In this formula, The tower layer input representing parameter i is obtained by weighting the gate layer Output to the expert network The weighted sum is obtained. Here, The softmax function output of the gating network represents the weight of the expert network j on the parameter i. Therefore, the tower layer integrates the outputs of different expert networks to generate the final representation of the action, that is, the adjustment amount to the target parameter. 4. Action selection, environment interaction, and memory pool update For each parameter i to be identified, the output from the tower is Select the best action Forming an array : where n is the number of parameters, Choose the action space that maximizes the tower output Each is the optimal action selected for parameter i, i.e., the parameter adjustment amount.

[0039] After selecting actions, the agent interacts with the power system simulation model by applying these actions to obtain rewards. For the rewards obtained.

[0040] State s, optimal action array , reward r and next state s' are stored in the memory pool: 5. Parameter Update The parameters of the online network are updated by sampling a mini-batch of samples from the memory pool. For each sample, its target Q value y is calculated. Since y is an array, it needs to be converted to a single value to calculate the loss function. Here, we use the Bellman equation to calculate the target Q value: in is the i-th element of the target Q value, is the reward for the corresponding sample, is the discount factor, is the next state and Q' is the output of the target network.

[0041] Compare the target Q value y and the output of the online network , use mean square error (MSE) as the loss function to update the parameters. The loss function is: Where n is the length of the action array, is the i-th element in the target array, is the corresponding action, is the corresponding value of the online network.

[0042] 6. Target network update The target network is updated periodically by copying the parameters of the online network.

[0043] 3. Implementation steps 1. Initialization Step 1: Set up the initial environment: Setting the initial parameters of the power system (Parameters to be identified).

[0044] Set the target parameters of the power system and obtain the target observation values . Step 2: Initialize the memory pool: Create a Memory Pool.

[0045] 2. Iteration process Step 1: Environmental status observation Obtaining initial observations of the power system in the initial environment and the target observation .

[0046] Step 2: Action Selection The agent uses the online network to calculate the current observation value. Select the best parameter tuning action , which indicates the adjustment amount of a certain power system parameter.

[0047] Step 3: Perform the action Execute an action , the environment updates the parameter state of the power system according to the action , and get new observations .

[0048] Step 4: Reward Calculation The environment is based on the new observations and the target observation The error between the current observation and the target observation is used to calculate the step reward and the total reward. The reward is calculated by measuring the difference between the current observation and the target observation through the loss function.

[0049] Step 5: Data Logging Record current parameter status ,action ,award and the next parameter status , and store this data in the memory pool.

[0050] Step 6: Check target value Checking power system observations Is it in the target observation value? If it is within the error range, update the new target value and update the new target observation value , and continue iterating until the predetermined number of target values ​​n is reached; otherwise, proceed to the next iteration.

[0051] Step 7: Update network parameters Sample from the memory pool and use these samples to update the parameters of the online network by maximizing the total reward.

[0052] At timed intervals, the parameters of the online network are copied to the target network to ensure the stability of the training process.

[0053] Step 8: Loop Iteration Repeat steps 2 to 7. When the maximum number of iterations or the target value n is reached, the epoch ends, initialization is performed, and training continues. When the number of epochs is greater than the maximum number of epochs, the entire training ends.

[0054] Experimental design and results analysis Experiment 1: Multi-parameter recognition ability test Experimental comparison method: Traditional statistical methods (TP) Genetic Algorithm (GA) Particle Swarm Optimization (PSO) Differential Evolution (DE) Deep Reinforcement Learning Method (DDQN) Evaluation Metrics: Root mean square error (RMSE): Evaluates the error between the identified parameters and the actual parameters, which is divided into maximum value (Max), average value (Mean) and minimum value (Min).

[0055] Convergence time: The time required to reach a predetermined error range.

[0056] Experimental results: analyze: Identification accuracy: The DMMoEQN method has the lowest RMSE with an average value of 0.0024, indicating that its identification accuracy is higher than other methods.

[0057] Convergence time: The convergence time of the DMMoEQN method is 510 minutes, which is significantly better than other methods, indicating that it converges faster.

[0058] Experiment 2: Test of responsiveness when target parameter value changes Purpose: Verify the responsiveness and identification performance of the DMMoEQN method when the target parameter values ​​change.

[0059] Experimental comparison method: Traditional statistical methods (TP) Genetic Algorithm (GA) Particle Swarm Optimization (PSO) Differential Evolution (DE) Deep Reinforcement Learning Method (DDQN) Evaluation Metrics: RMSE (maximum, average, minimum): Evaluates the error between the identified parameters and the actual parameters.

[0060] Response time: The time it takes for a system to adjust after a target parameter value changes.

[0061] Number of adjustments: The number of adjustments required by the system after the target parameter value changes.

[0062] Experimental results: analyze: Identification accuracy: When the target parameter value changes, the average RMSE of the DMMoEQN method is 0.0026, indicating that its identification accuracy is better than other methods.

[0063] Response time: The response time of the DMMoEQN method after the target parameter value changes is 8.7 seconds, which is significantly lower than other methods, indicating that it has significant advantages in coping with dynamic changes.

[0064] Number of adjustments: The number of adjustments required by the DMMoEQN method is 7, which is significantly lower than other methods, indicating that it is more efficient.

[0065] From the above comparison, it can be seen that the DMMoEQN method proposed in the present invention is significantly superior to the traditional method in terms of parameter identification accuracy, computational efficiency and response speed. Specifically, the DMMoEQN method shows a lower error in RMSE, and the total calculation time and single calculation time are also greatly reduced, which can better meet the requirements of real-time and accuracy in practical applications. Therefore, the present invention has obvious technical advantages and broad application prospects.

Claims

1. A method for identifying power system parameters based on DMMoEQN, characterized in that: The steps include: Step 1: Set the initial parameters of the power system and the target parameters of the power system, and obtain the target observation value , and create a memory pool; Step 2: Obtain the initial observation value of the power system and the target observation , the agent uses the online network to calculate the current observation value Selecting the best parameter tuning action ; Execute action , the power system operates according to Update the parameter status of the power system , and get new observations ; Step 3: Based on the new observations and the target observation Calculate the step reward based on the error between them, and get the total reward based on the step reward; Step 4: Record the current parameter status ,action ,award and the next parameter status , and stored in the memory pool; Step 5: Determine the observed value of the power system Is it in the target observation value? If the error range is within the error range, update the target value and target observation value , and continue to iterate until the predetermined target value n is reached, and then go to step eight; Otherwise, go to step 6; Step 6: Sample from the memory pool, update the parameters of the online network by maximizing the total reward based on the collected samples, and copy the parameters of the online network to the target network within the set time interval; Step 7, repeating steps 2 to 6. When the predetermined maximum number of iterations or the predetermined target value n is reached, the epoch ends and proceeds to step 8. Step eight, perform initialization and continue training; when the epoch number is greater than the preset maximum epoch number, the training ends and a trained power system parameter identification model is obtained. The power system parameters are obtained through the trained power system parameter identification model, and the power system is adjusted according to the power system parameters.

2. According to a DMMoEQN-based power system parameter identification method according to claim 1, it is characterized in that: The agent uses the online network to calculate the current observation value. Selecting the best parameter tuning action ; Execute action , the power system operates according to Update the parameter status of the power system , and get new observations ,include: The intelligent agent includes a target network and an online network based on MMoE; the intelligent agent is based on the current power system state observation value Select and execute the best parameter tuning action ; In each iteration, the power system updates the parameter state of the power system according to the actions of the intelligent agent , and get new observations .

3. The method for identifying power system parameters based on DMMoEQN according to claim 1, characterized in that: According to the new observations and the target observation The step reward is calculated based on the error between , and the total reward is obtained based on the step reward, including: Calculate the two-norm error between the current observation and the target observation; convert the two-norm error into an immediate reward, the immediate reward is: in is the current observation value, is the target observation value; accumulate all immediate rewards to get the total reward.

4. The method for identifying power system parameters based on DMMoEQN according to claim 1, characterized in that: The sampling from the memory pool and updating the parameters of the online network by maximizing the total reward based on the collected samples include: For each sample, the current Q value and the target Q value are calculated; the timing difference error is calculated, and the timing difference error is: Where r is the immediate reward, is the discount factor, Q' is the Q value of the target network; The mean square error loss function is used to minimize the temporal difference error. The mean square error loss function is: in is the TD error of the i-th sample; back-propagates the error and updates the network weights; periodically copies the parameters of the online network to the target network.

5. The method for identifying power system parameters based on DMMoEQN according to claim 1, characterized in that: The observation value of the power system is determined Is it in the target observation value? If the error range is within the error range, update the target value and target observation value ,include: At the beginning of training, N target values ​​are preset. After reaching the target observation value corresponding to the current target value, the next target value is switched. A new target observation value is obtained through simulation, and the new target observation value is set as the current target. The record of the target value is updated until all the preset N target values ​​are reached.

6. A power system parameter identification system based on DMMoEQN, characterized in that: A method for identifying power system parameters based on DMMoEQN as described in any one of claims 1 to 5 is applied, comprising a power system parameter identification model module, a power system parameter acquisition module, a data processing module and a data storage module; The power system parameter identification model module, the power system parameter acquisition module and the data storage module are respectively connected to the data processing module.

Images