Artificial Intelligence-Based Distributed Generation Control System and Method for Microgrids

Abstract

This invention relates to the field of microgrid control technology, and particularly to an artificial intelligence-based distributed generation control system and method for microgrids. The system includes a data acquisition and topology perception module, a multi-agent adaptive decision-making module, a distributed security and coordination module, an enhanced execution control module, and a dynamic strategy optimization module. In this scheme, the system utilizes sensors to collect operational data in real time and dynamically establishes a mathematical model based on an improved topology recognition algorithm to describe the topological relationships between generation units, forming a topology matrix reflecting the real-time structure of the system. In multi-agent decision-making, an attention mechanism is introduced to dynamically weight the state observation vector, automatically focusing on key influencing factors and avoiding information overload and decision bias. This enables the system to quickly generate realistic multi-objective decisions under complex scenarios such as load fluctuations and topology reconfiguration, significantly improving the adaptability and operational economy of the microgrid.

Description

Technical Field

[0001] This invention relates to the field of microgrid control technology, and in particular to a microgrid distributed generation control system and method based on artificial intelligence. Background Technology

[0002] Microgrids, as a novel power supply model integrating various distributed generation resources, play a crucial role in promoting energy structure transformation and improving power supply reliability. With the increasing penetration rate of renewable energy, microgrid operation and control face multiple challenges, including strong generation volatility and complex load changes. In existing technologies, artificial intelligence methods are gradually being introduced into the field of microgrid control, attempting to improve the system's automation level through data-driven modeling and intelligent decision support. Some studies employ centralized optimization or rule-based distributed control strategies to achieve power allocation and operational regulation of generation units. However, these methods often rely on preset models or fixed parameters, making them difficult to adapt to the dynamic and ever-changing operating environment of microgrids. Their control effectiveness is often limited, especially in complex scenarios such as topology changes and external market interference.

[0003] While existing technologies have improved the operational efficiency of microgrids to some extent, significant shortcomings remain. Traditional control systems lack real-time perception of the global operating state and accurate modeling of dynamic topology relationships, leading to incomplete decision-making information and delayed responses. Specifically, existing topology identification methods are mostly based on fixed electrical connections, failing to effectively quantify the dynamic correlation between units as power interaction intensity changes. Furthermore, most distributed control methods focus only on local optimization, failing to effectively coordinate power interactions between multiple units, easily causing safety issues such as voltage exceeding limits and frequency fluctuations. Particularly at the decision-making level, traditional methods struggle to automatically focus on key factors from high-dimensional, multi-source state information and lack effective mechanisms for embedding safety constraints in a differentiable form into the optimization process, making it difficult to balance safety and economy. In addition, control strategy updates often rely on offline training, making it difficult to adapt to changes in operating conditions online, limiting the system's long-term optimization capabilities. Mainstream online learning algorithms (such as the standard PPO) often face a conflict between exploration and convergence stability in dynamic environments like microgrids due to fixed hyperparameters. To address the aforementioned issues, this application proposes an artificial intelligence-based distributed generation control system and method for microgrids. By integrating multi-source data sensing, multi-agent collaborative decision-making, and dynamic strategy evolution, it achieves intelligent, safe, and efficient operation of the microgrid. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a microgrid distributed generation control system and method based on artificial intelligence, thereby solving the technical problems mentioned in the background section.

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

[0006] An AI-based distributed generation control system for microgrids includes:

[0007] The data acquisition and topology sensing module acquires the operating status data and communication connectivity status of distributed generation units in the microgrid through deployed sensors and communication links, and dynamically establishes a mathematical model based on an improved Bayesian network to describe topological association information.

[0008] The multi-agent adaptive decision-making module is connected to the data acquisition and topology perception module. By configuring an independent agent for each power generation unit and integrating state observation vectors and attention mechanisms, it generates dynamically adapted multi-objective decisions.

[0009] The distributed security collaboration module is connected to the multi-agent adaptive decision-making module. It exchanges decision information with neighboring agents through a consensus protocol and performs corrections based on a security optimization model constructed using the obstacle Lyapunov function.

[0010] An enhanced execution control module, connected to the distributed security collaboration module, achieves power command tracking control by introducing a target power reference value with an adaptive ramp rate limit and a feedforward-feedback composite control structure.

[0011] The dynamic policy optimization module is connected to the multi-agent adaptive decision-making module and the distributed security collaboration module, respectively. It updates network parameters online by filtering high-information-gain empirical data and using an improved proximal policy optimization algorithm to achieve continuous evolution and performance optimization of the agent's decision-making strategy.

[0012] In one possible implementation, the data acquisition and topology sensing module includes:

[0013] The state sensing unit is used to collect the instantaneous values ​​of the three-phase voltage, the instantaneous values ​​of the three-phase current, the output active power, the reactive power, the effective value of the voltage at the microgrid's point of common coupling, and the system frequency of the distributed generation unit.

[0014] The dynamic topology sensing unit is used to detect the communication connectivity status and power interaction strength of adjacent power generation units in real time, and to generate a dynamic topology association matrix based on an improved Bayesian network.

[0015] In one possible implementation, the multi-agent adaptive decision-making module includes:

[0016] The attention mechanism processing unit is used to fuse state-aware data, topological correlation matrix and external environment parameters from multiple sources to construct a state observation vector, and dynamically weight each dimension of the state observation vector through the attention mechanism.

[0017] The adaptive policy network unit is used to adaptively adjust the structure of the deep neural network and the parameters of the dynamic activation function based on the information entropy of the state vector, and output the initial control action.

[0018] In one possible implementation, the distributed security collaboration module includes:

[0019] The constraint parsing unit is used to store the hard constraints for microgrid operation;

[0020] The collaborative verification unit is used to exchange decision information with neighboring agents through a consensus protocol and to calculate the global collaborative error.

[0021] The safety correction unit is used to construct a safety optimization model based on the obstacle Lyapunov function and correct the initial decision.

[0022] In one possible implementation, the enhanced execution control module includes:

[0023] The reference value generation unit is used to calculate the target power reference value based on the final control action and the current operating power, and to add an adaptive ramp rate limit;

[0024] The composite control unit adopts a feedforward-feedback composite control structure to achieve zero steady-state error tracking of the fundamental component through a proportional-resonant regulator, and introduces a feedforward compensation term to generate a pulse width modulation drive signal through a space vector modulation algorithm.

[0025] In one possible implementation, the AI-based microgrid distributed generation control method includes:

[0026] S1: Through the data acquisition and topology sensing module, the operating status data of distributed generation units in the microgrid are collected, and a topology correlation matrix is ​​dynamically established based on the physical distance and power interaction intensity between generation units to complete the comprehensive perception and topology modeling of the microgrid's operating status.

[0027] S2: Based on the operating status data and topological correlation matrix obtained in step S1, a state observation vector is constructed for each power generation unit through the multi-agent adaptive decision-making module and dynamically weighted using the attention mechanism, and then preliminary control actions are generated through the adaptive policy network.

[0028] S3: Based on the preliminary control action generated in step S2, the distributed safety collaboration module interacts with neighboring agents to calculate the global collaboration error, and performs collaborative correction and safety constraint verification on the preliminary action based on the safety optimization model constructed based on the obstacle Lyapunov function, generating the final control action that satisfies global optimization and safe operation.

[0029] S4: Based on the final control action obtained in step S3, the target power reference value with adaptive ramp rate limit is calculated through the enhanced execution control module, and the power electronic converter is driven by the feedforward-feedback composite control strategy to achieve high-precision and fast tracking of power commands.

[0030] S5: Through the dynamic policy optimization module, the system operation experience data is collected and effective experiences with high information gain are screened. The improved proximal policy optimization algorithm is used to dynamically update the policy network parameters, so as to realize the online learning and continuous optimization of the agent's decision-making ability.

[0031] In one possible implementation, in step S1, the dynamic topology sensing unit calculates the topology correlation degree using the following formula. To generate dynamic topological correlation matrix :

[0032]

[0033] in, For the first , The physical distance between units This is the distance attenuation coefficient. For the first , Power interaction between the two units For the first , Power interaction between the two units Power weighting coefficients, dynamic topological correlation matrix By topological correlation It is formed by filling in matrix elements.

[0034] In one possible implementation, in step S2, the adaptive policy network unit uses a dynamic ELU function as the activation function, with a negative interval decay coefficient. Information entropy based on the state observation vector Dynamic adjustment, the adjustment method is as follows:

[0035] .

[0036] In one possible implementation, in step S3, the safety correction unit constructs a safety optimization model based on the barrier Lyapunov function, the model being required to satisfy the following constraints:

[0037] , ,

[0038] in, , , , , , The attenuation coefficient of the corresponding function. , , These are the first-order derivatives of the voltage safety barrier function, frequency safety barrier function, and power safety barrier function with respect to time, respectively.

[0039] In one possible implementation, in step S4, the reference value generation unit calculates the target power reference value after incorporating the ramp rate limit in the following manner. :

[0040]

[0041] in, Current operating power, , The maximum permissible rate of change of power per unit time. For maximum permissible power, To control the cycle, This is the final active power adjustment.

[0042] Beneficial effects compared to existing technologies:

[0043] 1. In this scheme, by constructing a data acquisition and topology sensing module and combining it with a multi-agent adaptive decision-making mechanism, comprehensive perception of the microgrid's operating status and dynamic adaptation of intelligent decision-making are achieved. The system utilizes high-precision sensors to collect multi-dimensional operating data such as voltage, current, and power in real time, and dynamically establishes a mathematical model based on an improved topology recognition algorithm to describe the topological associations between power generation units, forming a topology matrix reflecting the real-time structure of the system. In multi-agent decision-making, an attention mechanism is introduced to dynamically weight the state observation vector, automatically focusing on key influencing factors and avoiding information overload and decision bias. Simultaneously, the adaptive strategy network flexibly adjusts the network structure and activation function characteristics according to the complexity of the operating conditions, outputting preliminary control actions that balance economy and stability. This enables the system to quickly generate realistic multi-objective decisions under complex scenarios such as load fluctuations and topology reconfiguration, significantly improving the microgrid's adaptability and operational economy.

[0044] 2. In this scheme, the synergistic effect of the distributed security collaboration module and the enhanced execution control module effectively solves the local optimization problem and the challenge of satisfying security constraints in single-agent decision-making, achieving global power collaboration and operational safety assurance. The system exchanges decision information with neighboring agents based on a consensus protocol, calculates the global collaboration error, and dynamically allocates correction amounts through topological correlation to ensure that the power adjustment of each power generation unit meets the overall system requirements. In the security correction stage, a safety optimization model is constructed using a barrier Lyapunov function to verify and correct the constraint satisfaction of the initial decision, ensuring that voltage, frequency, and power output always operate within the safety boundaries. The execution control stage uses a feedforward-feedback composite structure, calculates the target power reference value through a reference value generation unit, and employs space vector modulation technology to achieve high-precision and fast tracking of power commands, effectively suppressing the impact of external disturbances. These measures enable the system to maintain global collaboration and safe and stable operation without the need for central node intervention, significantly improving the system's flexibility and stability.

[0045] 3. In this scheme, the dynamic policy optimization module enables online adaptive updating and continuous evolution of the agent's decision-making strategy, ensuring optimal performance throughout long-term operation. The system employs an experience-based selection mechanism, quantifying the value of each experience based on information gain, and storing only high-value experiences in the replay library to avoid data redundancy and reduced optimization efficiency. During the policy update phase, an improved near-end policy optimization algorithm is introduced, combined with dynamic parameter pruning and an adaptive learning rate mechanism to balance policy exploration and convergence stability. A multi-objective reward function guides the policy towards the optimal direction for economic, stability, and safety. This optimization process is executed online in short cycles, enabling the agent to respond in real-time to changes in microgrid conditions, continuously accumulate operational experience, and improve itself. Attached Figure Description

[0046] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0047] Figure 1 This is a schematic diagram of the operational framework of the microgrid distributed generation control system of the present invention;

[0048] Figure 2 This is a schematic diagram of the microgrid distributed generation control method of the present invention. Detailed Implementation

[0049] Preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, the present invention can also be implemented in various different forms, and therefore the present invention is not limited to the embodiments described below. In addition, for the purpose of more clearly describing the present invention, parts not connected to the invention will be omitted from the drawings.

[0050] The technical solutions in this application are designed to address the problems described in the background, and are generally as follows:

[0051] Example:

[0052] Please refer to Figure 1 This embodiment introduces an artificial intelligence-based microgrid distributed generation control system, including a data acquisition and topology sensing module, a multi-agent adaptive decision-making module, a distributed security collaboration module, an enhanced execution control module, and a dynamic strategy optimization module.

[0053] The data acquisition and topology sensing module is used to comprehensively acquire microgrid operating status data and topology correlation information, including a status sensing unit and a dynamic topology sensing unit.

[0054] The state sensing unit deploys high-precision voltage sensors, current sensors, and power sensors to collect instantaneous three-phase voltage values ​​for different types of distributed generation units, such as photovoltaic, wind power, and energy storage. , , Instantaneous value of three-phase current , , Output active power reactive power Simultaneously, the effective voltage value at the microgrid's point of common coupling (PCC) is collected. With system frequency ;

[0055] The dynamic topology sensing unit designs a topology identification algorithm based on an improved Bayesian network. This improved Bayesian network integrates physical distance attenuation and power interaction strength to detect the communication connectivity and power interaction strength of adjacent power generation units in real time, generating a dynamic topology association matrix. ,in, The total number of distributed generation units within the microgrid, matrix elements Indicates the first The and the first The topological correlation degree of each power generation unit, with a value range of... , =0 indicates no direct topological association. =1 indicates that the strongest topological association exists.

[0056] The multi-agent adaptive decision-making module configures an independent agent for each distributed generation unit and generates dynamically adapted multi-objective decisions based on operating conditions, including an attention mechanism processing unit and an adaptive policy network unit.

[0057] The attention mechanism processing unit integrates state-aware data, topological correlation matrix, and external environmental parameters (real-time electricity price). Load forecast Multi-source fusion to construct a high-dimensional state observation vector ,in, For the topological incidence matrix, the first Row vectors, through the attention weight matrix Dynamically weight each dimension of the state observation vector to automatically focus on key influencing factors of the current operating condition;

[0058] The adaptive policy network unit incorporates a deep neural network with a dynamic activation function. It adaptively adjusts the network structure and connection weights based on the complexity of the operating conditions (quantization of state vector information entropy), and outputs preliminary control actions that balance economic benefits and system stability. (i.e., active power adjustment amount) ).

[0059] The distributed security collaboration module is used to solve the local optimization problem and the security constraint satisfaction problem of single agent decision-making, and to achieve global collaborative optimization and operational security assurance. It includes a constraint parsing unit, a collaborative verification unit and a security correction unit.

[0060] The constraint parsing unit stores the hard constraints for microgrid operation: power capacity constraints of generation units. PCC voltage constraint System frequency constraints and distributed collaborative constraints ,in, To meet the system's total power adjustment requirements, Generally refers to power adjustment, specifically referring to active power adjustment. Or final active power adjustment ;

[0061] The collaborative verification unit exchanges decision information with neighboring agents through a consensus protocol and calculates the global collaborative error. ;

[0062] The safety correction unit constructs a safety optimization model based on the obstacle Lyapunov function, corrects the initial decision, and generates the final control action that meets safety constraints and coordination requirements. (i.e., final active power adjustment) ).

[0063] The enhanced execution control module receives the final control action command and realizes high-precision and fast power tracking control, including a reference value generation unit and a composite control unit;

[0064] The reference value generation unit is based on With current operating power Calculate the target power reference value At the same time, a gradient limit is added. ,in, This is the maximum permissible rate of power change per unit time, to avoid system fluctuations caused by sudden power changes;

[0065] The composite control unit adopts a feedforward-feedback composite control structure. The feedback channel is based on a proportional-resonant (PR) regulator to achieve zero steady-state error tracking of the fundamental component. The feedforward channel introduces grid voltage and load disturbance compensation terms and generates PWM drive signals through a space vector modulation (SVM) algorithm to precisely control the switching action of the power electronic converter.

[0066] The dynamic policy optimization module is used to realize the online adaptive update of the agent's decision-making policy and continuously optimize the decision-making performance. It includes an experience screening unit and a gradient descent unit.

[0067] The experience screening unit uses experience data for each control cycle. Filtering, among which, This is the state observation vector for the next time step. As a reward value, only information gain greater than a threshold is stored. Effective experience;

[0068] The gradient descent unit employs an improved proximal policy optimization (PPO) algorithm, which dynamically adjusts the learning rate based on the complexity of the operating conditions. With cutting parameters By minimizing the constrained policy loss function, the parameters of the adaptive policy network are updated, thereby enabling the continuous evolution of decision-making capabilities.

[0069] Please refer to Figure 2 This embodiment also introduces an artificial intelligence-based microgrid distributed generation control method, which is executed collaboratively by the various modules of the above system. The method achieves intelligent, safe, and collaboratively optimized control of microgrid distributed generation through the following steps, each step in a single control cycle. =Executes in a loop every 50ms:

[0070] S1: Data Acquisition and Dynamic Topology Modeling

[0071] This step is performed by the data acquisition and topology sensing module. First, multi-dimensional operational data acquisition is completed through the state sensing unit. The sampling frequency of the voltage and current sensors is set to 10kHz to acquire the instantaneous values ​​of the three-phase voltage. , , Instantaneous value of three-phase current , , After filtering out high-frequency noise using a moving average filter (filter window length of 20 sampling points), the voltage and current on the abc axis are converted to the dq axis using the Park transform, resulting in... , and , Calculate instantaneous active power With instantaneous reactive power The active power is obtained by performing a moving average of the instantaneous power over 20 power frequency cycles (400ms). With reactive power Simultaneously, the phase-locked loop (PLL) tracks the phase and frequency of the PCC voltage to output the system frequency. And calculate the effective value of the PCC voltage. ,in, This represents the instantaneous voltage value at the microgrid's point of common coupling (PCC). =20ms is the power frequency cycle.

[0072] The dynamic topology sensing unit establishes communication with adjacent power generation units via a CAN bus (communication period 10ms). Each unit periodically broadcasts its own device ID, operating status, and real-time output power. Topology correlation is designed based on an improved Bayesian network. The computational model quantifies topological relationships by integrating physical distance and power interaction intensity.

[0073]

[0074] in, For the first , The physical distance between units This is the distance attenuation coefficient. For the first , Power interaction between the two units For the first , Power interaction between the two units This is the power weighting coefficient.

[0075] When communication between the two units is interrupted for more than 3 communication cycles or When <0.1, set =0, generate dynamic topological association matrix At the same time, real-time electricity prices are obtained by communicating with the distribution network dispatch center. Compared with the load forecast for the next 15 minutes Complete full-dimensional data collection and preprocessing.

[0076] S2: Generate adaptive multi-objective decisions

[0077] This step is executed by the multi-agent adaptive decision-making module, enabling dynamic adaptation of the decision network to the operating conditions and enhancing the ability to extract key information. First, an attention mechanism processing unit fuses multi-source data to construct a state observation vector. ,in, , For the topological incidence matrix, the first Okay, number Column vector, state observation vector dimension is It covers device status, topology relationships, and external environment information.

[0078] Set attention weight matrix Its diagonal elements The attention weights for the corresponding state dimensions are calculated as follows:

[0079]

[0080] in, The normalized value for the state dimension, with a range of values. , The activation function is Softmax. The weighted state vector... It can adaptively focus on key factors, for example, Mutation magnitude exceeding 10% or When the average electricity price is 1.5 times higher, the corresponding dimension weight is automatically increased to above 0.3, which strengthens the economic benefit orientation and avoids the problem of insufficient adaptability of the fixed weight model to complex operating conditions.

[0081] Will Input adaptive policy network unit, which adopts a dynamic structure design: state vector information entropy For values ​​≤3.2 (simple operating conditions, such as stable load and fixed topology), two hidden layers are used. When the entropy is greater than 3.2 (complex operating conditions, such as drastic load fluctuations and topology reconstruction), the number of layers is automatically increased to 4. The number of neurons in each layer is initially 64. Under complex operating conditions, 8 neurons are added for every 0.1 increase in information entropy, with a maximum of no more than 128 neurons. This ensures representational capability while avoiding decision delays caused by network redundancy.

[0082] The network activation function uses a dynamic ELU function, adjusting the negative interval decay characteristics according to the operating complexity:

[0083]

[0084] in, For dynamic ELU functions, This is the input to the activation function, i.e., the input value of the neural network. , This is the negative interval decay coefficient of the dynamic ELU function, which is relevant when the operating conditions are complex. Approaching 0.5 enhances sensitivity to negative biases (such as insufficient power or low voltage); approaching 0.8 at low complexity improves decision stability, thereby ensuring accurate matching of activation function characteristics under different operating conditions.

[0085] The network output layer has a dual-output structure, outputting respectively... With reactive power auxiliary adjustment , This refers to the initial control action. value range ,and This is used to assist in regulating the reactive power balance of microgrids and adapt to voltage stability control requirements. To achieve multi-objective optimization, a multi-objective loss function is introduced:

[0086]

[0087] in, For economic losses, This represents the average electricity price over the past hour. The economic benefit deviation is quantified based on the average power over the past hour. The stable loss term characterizes the system's operational stability. To mitigate safety risks, ensure that power output remains within safe limits. , These are the maximum active power and the minimum active power, respectively. , , The dynamic weighting coefficients (summing to 1) are dynamically adjusted by the fuzzy logic controller: when system fluctuations are large, the weighting coefficients are increased. The value is increased during peak electricity prices. When the power is close to the limit, increase the value. The value is selected to ensure that the decision-making balances the needs of multiple objectives.

[0088] S3: Distributed Collaboration and Security Correction

[0089] This step is executed by the distributed security collaboration module, constructing a deeply integrated architecture of distributed collaboration and security constraints, enabling global optimization without the intervention of a central node. First, the hard constraint parameters are read through the constraint resolution unit: photovoltaic unit. Energy storage unit ; , =380V; Distributed collaborative constraints , It is obtained through distributed power flow calculation and collaboration, without the need for allocation by a central node.

[0090] The collaborative verification unit communicates with the consensus protocol. Neighboring units with a correlation of ≥0.3 interact and make decisions, selecting highly correlated units to participate in collaboration, reducing unnecessary communication overhead and network latency. Each agent sends its own... And receive from adjacent units Calculate the global cooperative error .when ≤0.05 No correction is needed if the condition is met; otherwise, it is based on topological correlation. Assigning co-correction amounts:

[0091]

[0092] in, For the first The sum of the topological correlations of each unit, after co-correction, yields... This correction method does not require central coordination; global collaboration can be achieved through local interactions between adjacent units, thereby improving control flexibility and anti-interference capabilities.

[0093] The safety correction unit constructs a safety optimization model based on the barrier Lyapunov function and defines the voltage safety barrier function. Frequency security barrier function Power safety barrier function The safety constraints are as follows:

[0094] , ,

[0095] in, , , , , , The attenuation coefficient of the corresponding function has the following value ranges: , , , , , These are the first derivatives of each barrier function with respect to time.

[0096] The coupling relationship between power regulation and voltage and frequency is obtained through system identification. , ,in, , These represent the coupling coefficients of power regulation to voltage and frequency, respectively. If... If the constraints are satisfied, then it is directly used as... Otherwise, construct a safety optimization problem that satisfies safety constraints while preserving the intention of collaborative correction:

[0097]

[0098] St , ,

[0099] This convex quadratic programming problem is solved using the interior-point method (solution time less than 10 ms), ultimately yielding a solution that satisfies both cooperative and security requirements. This effectively avoids local optima and security risks.

[0100] S4: Enhanced Execution Control

[0101] This step is executed by the enhanced execution control module, employing a feedforward-feedback composite control structure and a dynamic reference value generation mechanism to improve power tracking accuracy and anti-interference capability. First, the target power reference value is calculated by the reference value generation unit. Add a ramp rate limit to avoid sudden power surges:

[0102]

[0103] in, It is adaptively set according to the power adjustment capability of the power generation unit to ensure a smooth transition in power adjustment.

[0104] The composite control unit adopts a feedforward-feedback composite control structure, and the feedback channel uses a PR regulator with the following transfer function:

[0105]

[0106] in, For Laplace variables, For harmonic order, This is the proportionality coefficient. The resonant coefficient has a range of values ​​of [values ​​to be filled in]. , This achieves zero steady-state error tracking of the fundamental frequency and suppression of major harmonics. The feedforward channel incorporates grid voltage feedforward. With load disturbance feedforward voltage , These are the observed values ​​of the power grid voltage. These are the observed values ​​of the load current. This is the equivalent impedance of the system, used to preemptively offset the effects of external disturbances.

[0107] The PR regulator output, after being superimposed with the feedforward compensation term, generates the dq-axis voltage reference value. , Six PWM drive signals are generated using the SVM algorithm. The interaction time between adjacent effective vectors and zero vectors is calculated as follows:

[0108] , ,

[0109] in, The duration of action of the first effective vector. The duration of action of the second effective vector. The duration of action of the zero vector. , =100µs is the switching period The reference voltage vector phase angle, =750V DC bus voltage. PWM signal drives IGBT switching transistors to achieve power matching. High-precision tracking.

[0110] S5: Dynamic Strategy Optimization

[0111] This step is executed by the dynamic strategy optimization module, which constructs an effective experience screening and dynamic parameter adjustment strategy update mechanism to achieve online evolution of the decision-making strategy. First, experience data is collected through the experience screening unit. Reward value A comprehensive evaluation of economic benefits, system stability, and operational safety is necessary.

[0112]

[0113] in, , The peak electricity price for the day. , , The larger the value, the better the control effect. The strategy is guided to evolve towards the optimal overall performance through multi-dimensional reward values.

[0114] To address the low optimization efficiency caused by redundancy in the experience base, an information gain filtering mechanism is adopted, defining information entropy. Quantitative indicators characterizing the uncertainty of random variables, calculating the information gain of each piece of experience. ,in, Current state Information entropy Given the state at the next time step Current state The conditional information entropy. Only when When the information gain threshold is adjusted according to the optimization accuracy, it is determined that the experience contains valid new information and stored in the experience replay library (circular buffer, capacity 10). 5 (Item 1) When the experience base reaches its capacity limit, remove the old experience with the least information gain to ensure that the experience base always stores high-value and effective experience, thereby improving the efficiency and effectiveness of strategy optimization.

[0115] The gradient descent unit uses an improved PPO algorithm to update network parameters. , Given a set of parameters consisting of policy network weights and biases, this algorithm uses a clip mechanism to limit the policy update magnitude, avoiding training instability caused by excessively rapid parameter updates. Furthermore, a dynamic parameter adjustment mechanism is introduced to further improve the algorithm's convergence and adaptability. A mini-batch of data is formed by randomly sampling 256 valid experiences from the experience base, and the generalized advantage estimation (GAE) method is used to calculate the advantage function estimate. To improve the accuracy of advantage estimation, its expression is:

[0116]

[0117] in, The advantage function estimate represents the value at time step [missing information]. The relative advantage of taking action, Index for the current time step, This represents the offset for future time steps, used to accumulate the temporal differential error across different offset steps after the current time step. The length of the sampling sequence. This is a discount factor used to weigh the importance of current and future rewards. These are GAE parameters used to balance the bias and variance of the dominance estimate. For time step The timing difference error is calculated using the following formula:

[0118]

[0119] For time step The reward value, compared to the reward value for a single control cycle mentioned earlier. The meaning is consistent. Time-domain representation at a specific time step For value networks (used to estimate state values). For value networks in state The output below, For the parameter set of the value network, For the target value network, For the target value network in state The output below, The parameter set of the target value network is synchronized once every 10 policy updates. = To avoid value estimation bias.

[0120] Introducing dynamic clipping parameters With adaptive learning rate Based on the number of training steps Dynamic adjustment, balancing strategy exploration and convergence stability, its expression is:

[0121] ,

[0122] Early training phase ( When it is smaller), Approaching 0.2, it supports rapid strategy exploration to adapt to diverse operating conditions; in the later stages of training ( When it is large), Approaching 0.1 limits the policy update magnitude and ensures convergence stability; adaptive learning rate. The learning rate gradually decreases with the number of training steps to avoid excessive learning rate in the later stages, which could lead to parameter oscillations and improve convergence accuracy.

[0123] The improved PPO loss function is designed as follows, introducing an entropy regularization term to encourage policy exploration and avoid getting trapped in local optima:

[0124]

[0125] in, This indicates an experience replay library. This indicates expectations regarding the data in the experience base. The probability ratio of the strategies. For the current strategy (parameter is) In state Down Output Action The probability density, The old strategy before the update (parameter is) The corresponding probability density, For the pruning function, the policy probability ratio is restricted to... Within the interval, Here is the entropy regularization coefficient. The entropy of the current strategy is used to increase the randomness of the strategy, encouraging the exploration of better decisions.

[0126] The Adam optimizer is used to minimize the loss function and update the policy network parameters. The updated formula is:

[0127]

[0128] in, For loss function For parameters The gradient is calculated. Each parameter update iteration lasts for 10 rounds to ensure sufficient parameter optimization. A strategy update is performed every 50 control cycles to balance optimization effectiveness and real-time performance. Through this dynamic strategy optimization mechanism, the adaptive strategy network can continuously learn and evolve according to the dynamic changes in microgrid operating conditions, maintaining optimal decision-making performance in complex and ever-changing operating scenarios, and significantly improving the adaptive capability and robustness of the control system.

[0129] Finally, it should be noted that the above embodiments are merely examples for clearly illustrating the present invention and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A microgrid distributed generation control system based on artificial intelligence, characterized in that, include: The data acquisition and topology sensing module acquires the operating status data and communication connectivity status of distributed generation units in the microgrid through deployed sensors and communication links, and dynamically establishes a mathematical model based on an improved Bayesian network to describe topological association information. The multi-agent adaptive decision-making module is connected to the data acquisition and topology perception module. By configuring an independent agent for each power generation unit and integrating state observation vectors and attention mechanisms, it generates dynamically adapted multi-objective decisions. The distributed security collaboration module is connected to the multi-agent adaptive decision-making module. It exchanges decision information with neighboring agents through a consensus protocol and performs corrections based on a security optimization model constructed using the obstacle Lyapunov function. An enhanced execution control module, connected to the distributed security collaboration module, achieves power command tracking control by introducing a target power reference value with an adaptive ramp rate limit and a feedforward-feedback composite control structure. The dynamic policy optimization module is connected to the multi-agent adaptive decision-making module and the distributed security collaboration module, respectively. It updates network parameters online by filtering high-information-gain empirical data and using an improved proximal policy optimization algorithm to achieve continuous evolution and performance optimization of the agent's decision-making strategy.

2. The microgrid distributed generation control system based on artificial intelligence as described in claim 1, characterized in that, The data acquisition and topology sensing module includes: The state sensing unit is used to collect the instantaneous values ​​of the three-phase voltage, the instantaneous values ​​of the three-phase current, the output active power, the reactive power, the effective value of the voltage at the microgrid's point of common coupling, and the system frequency of the distributed generation unit. The dynamic topology sensing unit is used to detect the communication connectivity status and power interaction strength of adjacent power generation units in real time, and to generate a dynamic topology association matrix based on an improved Bayesian network.

3. The microgrid distributed generation control system based on artificial intelligence as described in claim 1, characterized in that, The multi-agent adaptive decision-making module includes: The attention mechanism processing unit is used to fuse state-aware data, topological correlation matrix and external environment parameters from multiple sources to construct a state observation vector, and dynamically weight each dimension of the state observation vector through the attention mechanism. The adaptive policy network unit is used to adaptively adjust the structure of the deep neural network and the parameters of the dynamic activation function based on the information entropy of the state vector, and output the initial control action.

4. The microgrid distributed generation control system based on artificial intelligence as described in claim 1, characterized in that, The distributed security collaboration module includes: The constraint parsing unit is used to store the hard constraints for microgrid operation; The collaborative verification unit is used to exchange decision information with neighboring agents through a consensus protocol and calculate the global collaborative error. The safety correction unit is used to construct a safety optimization model based on the obstacle Lyapunov function and correct the initial decision.

5. The microgrid distributed generation control system based on artificial intelligence as described in claim 1, characterized in that, The enhanced execution control module includes: The reference value generation unit is used to calculate the target power reference value based on the final control action and the current operating power, and to add an adaptive ramp rate limit; The composite control unit adopts a feedforward-feedback composite control structure to achieve zero steady-state error tracking of the fundamental component through a proportional-resonant regulator, and introduces a feedforward compensation term to generate a pulse width modulation drive signal through a space vector modulation algorithm.

6. A microgrid distributed generation control method based on artificial intelligence, implementing the system as described in any one of claims 1 to 5, characterized in that, include: S1: Through the data acquisition and topology sensing module, the operating status data of distributed generation units in the microgrid are collected, and the topology association matrix is ​​dynamically established based on the physical distance and power interaction intensity between generation units through the dynamic topology sensing unit, so as to complete the comprehensive perception and topology modeling of the microgrid's operating status. S2: Based on the operating status data and topological correlation matrix obtained in step S1, a state observation vector is constructed for each power generation unit through the multi-agent adaptive decision-making module and dynamically weighted using the attention mechanism. Then, the initial control action is generated through the adaptive policy network unit. S3: Based on the preliminary control action generated in step S2, the distributed safety collaboration module interacts with neighboring agents to calculate the global collaboration error. The safety correction unit performs collaborative correction and safety constraint verification on the preliminary action based on the safety optimization model constructed by the obstacle Lyapunov function, generating the final control action that satisfies global optimization and safe operation. S4: Based on the final control action obtained in step S3, the target power reference value with adaptive ramp rate limit is calculated by the reference value generation unit in the enhanced execution control module, and the power electronic converter is driven by the feedforward-feedback composite control strategy to achieve high-precision and fast tracking of power commands. S5: Through the dynamic policy optimization module, the system operation experience data is collected and effective experiences with high information gain are screened. The improved proximal policy optimization algorithm is used to dynamically update the policy network parameters, so as to realize the online learning and continuous optimization of the agent's decision-making ability.

7. The microgrid distributed generation control method based on artificial intelligence as described in claim 6, characterized in that, In step S1, the dynamic topology sensing unit calculates the topology correlation degree using the following formula. To generate dynamic topological correlation matrix : ； in, For the first , The physical distance between units This is the distance attenuation coefficient. For the first , Power interaction between the two units For the first , Power interaction between the two units Power weighting coefficients, dynamic topological correlation matrix By topological correlation It is formed by filling in matrix elements.

8. The microgrid distributed generation control method based on artificial intelligence as described in claim 6, characterized in that, In step S2, the adaptive policy network unit uses a dynamic ELU function as the activation function, and its negative interval decay coefficient... Information entropy based on the state observation vector Dynamic adjustment, the adjustment method is as follows: 。 9. The microgrid distributed generation control method based on artificial intelligence as described in claim 6, characterized in that, In step S3, the safety correction unit constructs a safety optimization model based on the barrier Lyapunov function, and the model is required to satisfy the following constraints: ， ， ； in, , , , , , The attenuation coefficient of the corresponding function. , , These are the first-order derivatives of the voltage safety barrier function, frequency safety barrier function, and power safety barrier function with respect to time, respectively.

10. The microgrid distributed generation control method based on artificial intelligence as described in claim 6, characterized in that, In step S4, the reference value generation unit calculates the target power reference value after adding the ramp rate limit using the following method. : ； in, Current operating power, , The maximum permissible rate of change of power per unit time. For maximum permissible power, To control the cycle, This is the final active power adjustment.

Images