Fusion of mamba prediction with multi-agent active distribution network voltage control method

Abstract

The present application relates to the technical field of power system intelligent control, and more particularly to an active distribution network voltage control method fusing Mamba prediction and multi-agent, which is mainly aimed at the problems of lack of foresight and low inference efficiency of prediction model in existing multi-agent reinforcement learning, and proposes the following technical scheme: step one, dividing the region based on the topology structure of the active distribution network; step two, predicting the node voltage and power of the distribution network; step three, constructing an optimization target with the minimum node voltage disturbance deviation and the minimum control cost as the target, and reconstructing the active distribution network voltage control problem; step four, realizing the active distribution network voltage collaborative control.The present application gives the voltage control foresight by fusing Mamba efficient prediction and MATD3 multi-agent collaborative decision-making, reduces the computational complexity, relieves the dimension disaster, and improves the accuracy, real-time performance and economy of the distribution network voltage stability control.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for power systems, and in particular to an active distribution network voltage control method that integrates Mamba prediction and multi-agent systems. Background Technology

[0002] The intermittent and random nature of photovoltaic (PV) power output, along with the complexity of load fluctuations, has transformed the distribution network from a traditional unidirectional radial network into a complex system with bidirectional power flow. This has led to increased voltage fluctuations and a sharp rise in the risk of exceeding limits, seriously threatening the safe and stable operation of the power grid and power quality. Traditional voltage control methods, such as reactive power optimization based on sensitivity analysis and secondary voltage control, typically rely on accurate grid models and linearization assumptions. When faced with the nonlinear and high-dimensional characteristics of the distribution network and the rapid fluctuations of renewable energy sources, these methods suffer from severe model mismatch, limited control effectiveness, and computational time constraints that cannot meet real-time requirements. Methods based on classical reinforcement learning, such as deep Q-networks and deep deterministic policy gradients, have been attempted for voltage control. However, single-agent architectures struggle to effectively coordinate the numerous dispersed control resources in the grid and suffer from the "curse of dimensionality." More importantly, these methods are essentially based on current or historical states, lacking the ability to predict future dynamic evolution trends of the system. This results in short-sighted control strategies, prone to control oscillations or action delays, and unable to effectively mitigate the rapid fluctuations of renewable energy sources. Multi-agent reinforcement learning (MAL) methods, with their centralized training and distributed execution framework, alleviate the challenge of non-stationarity in multi-agent environments to some extent. However, the state space of existing MML methods typically only contains measurement data at the current moment, resulting in a lack of foresight in decision-making. Furthermore, time-series prediction models, such as Long Short-Term Memory (LSTM) networks, which form the basis of MML perception, suffer from weak long-range dependency capture capabilities and low training parallelism. While Transformer models offer high accuracy, the computational complexity of their self-attention mechanism increases quadratically with sequence length, requiring enormous computational resources and making real-time inference difficult on resource-constrained distribution network edge controllers. Therefore, this invention proposes an active distribution network voltage control method that integrates Mamba prediction and multi-agent approaches. Summary of the Invention

[0003] The purpose of this invention is to address the problems in the prior art where existing multi-agent reinforcement learning lacks foresight and the prediction model has low inference efficiency, and to propose an active distribution network voltage control method that integrates Mamba prediction and multi-agent systems.

[0004] The technical solution of this invention: an active distribution network voltage control method integrating Mamba prediction and multi-agent systems, comprising the following steps:

[0005] Step 1: Divide the area based on the active distribution network topology, collect the node voltage, active power and reactive power of each area through smart meters, and preprocess the data.

[0006] Step 2: Predict the voltage and power of distribution network nodes using the Mamba algorithm with fused wavelet transform. The Mamba algorithm with fused wavelet transform includes feature mapping, state space construction, wavelet transform, construction of wavelet state transformation module with local attention mechanism, gating fusion, and state space model update steps.

[0007] Step 3: Construct an optimization objective that minimizes node voltage disturbance deviation and control cost, set node voltage constraints and photovoltaic inverter power constraints, and reconstruct the active distribution network voltage control problem through a partially observable Markov model;

[0008] Step 4: Using a centralized training-distributed execution architecture based on the MATD3 algorithm, each agent generates control actions based on local observations and global prediction information to achieve active voltage coordination control of the distribution network.

[0009] Optionally, the preprocessed data in step one includes the node voltage, photovoltaic active power, photovoltaic reactive power, load active power, and load reactive power of the sub-region. The sub-region is determined by dividing the active distribution network into node sets, load node sets, branch sets, and photovoltaic node sets.

[0010] Optionally, in step two, the wavelet transform extracts the low-frequency and high-frequency components of the state space using a low-pass filter and a high-pass filter, respectively.

[0011] The wavelet state transformation module of the local attention mechanism generates enhanced representations of low-frequency and high-frequency components through querying, key-value vector calculation, and filtering of local window neighbor sets.

[0012] Optionally, in step two, the gated fusion integrates the output of the wavelet state transformation module of the local attention mechanism with the state space of the Mamba algorithm through the gated vector, the final gated weight matrix, and the final gated bias vector, thereby updating and obtaining the predicted values ​​of node voltage and power.

[0013] Optionally, the node voltage constraint in step three defines the upper and lower bounds of the node voltage;

[0014] Photovoltaic inverter power constraints include upper and lower limits of photovoltaic active power, upper and lower limits of photovoltaic reactive power, and photovoltaic rated power and capacity factor constraints.

[0015] Optionally, the partially observable Markov model in step three includes the state space, action space, local observation set, and reward function;

[0016] The state space contains photovoltaic reactive power control instructions on the nodes, and the action space corresponds to the set of photovoltaic reactive power control instructions for each agent.

[0017] Optionally, in step four, the MATD3 algorithm outputs control actions through the Actor network, calculates the evaluation function through the Critic network, samples samples using the experience replay pool, updates network parameters through policy gradients and Q-values, and uses soft update coefficients to update the parameters of the TargetActor network and the TargetCritic network.

[0018] Optionally, the Actor network retained after the MATD3 algorithm training converges is used for online decision-making, and the Actor network of each agent realizes real-time optimization control of regional autonomy based on the information of the controlled sub-region.

[0019] In summary, this application includes at least one of the following beneficial technical effects:

[0020] This invention introduces the Mamba state-space model with linear computational complexity and integrates wavelet transform and local attention mechanism to construct a prediction module. This not only solves the shortcomings of long short-term memory networks in terms of weak long-range dependency capture ability and low training parallelism, but also avoids the problem that the computational complexity of the Transformer model's self-attention mechanism increases quadratically with the sequence length. This prediction module can efficiently process massive time-series operation data of distribution networks and accurately output the prediction results of node voltage and power, providing reliable forward-looking information support for voltage control and overcoming the limitations of traditional control methods that rely only on the current or historical state.

[0021] Furthermore, by constructing a partially observable Markov decision process with Mamba prediction information as the core, the future dynamic evolution trend of the system is incorporated into the state space of multi-agent decision-making. Combined with the optimized training of the MATD3 algorithm, each distributed agent can not only perceive the current operating state of the sub-region when generating control actions, but also predict the future impact of photovoltaic output and load fluctuations. This design effectively solves the short-sightedness problem of traditional reinforcement learning control strategies, avoids voltage control oscillations and action delays, significantly improves the ability to smooth out rapid fluctuations in new energy sources, and ensures that the voltage of the distribution network remains stable within a reasonable range.

[0022] Furthermore, by adopting a centralized training and distributed execution architecture of the MATD3 algorithm, combined with a regional division strategy for active distribution network topology, each agent only needs to make decisions based on local observation information and global prediction information of the controlled sub-region. This architecture not only ensures global coordination among multiple agents through centralized training, but also realizes real-time optimization control of regional autonomy through distributed execution. It effectively alleviates the dimensionality curse faced by single agent architecture when coordinating a large number of dispersed control resources, and improves the efficiency and flexibility of multi-agent collaborative control.

[0023] In summary, this invention, by integrating Mamba's efficient prediction with MATD3's multi-agent collaborative decision-making, endows voltage control with foresight, reduces computational complexity, mitigates the curse of dimensionality, and improves the accuracy, real-time performance, and economy of distribution network voltage stability control. Attached Figure Description

[0024] Figure 1 A schematic diagram of the active distribution network voltage control optimization solution framework based on the MATD3 algorithm proposed in this invention;

[0025] Figure 2 This is a partitioned topology diagram of the IEEE 33-node active distribution network in the embodiment;

[0026] Figure 3 This is a graph showing the cumulative reward changes in the example;

[0027] Figure 4 The figure shows a comparison of different control algorithms in the embodiments. Detailed Implementation

[0028] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0029] The components of the embodiments of the invention described and shown in the accompanying drawings can typically be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention.

[0030] Example

[0031] Reference Figure 1 As shown, the active distribution network voltage control method integrating Mamba prediction and multi-agent systems proposed in this invention includes the following steps:

[0032] Step 1: Divide the area based on the active distribution network topology, collect the node voltage, active power and reactive power of each area through smart meters, and preprocess the data.

[0033] The preprocessed data in step one includes the node voltage, photovoltaic active power, photovoltaic reactive power, load active power, and load reactive power of the sub-region. The sub-region is determined by dividing the active distribution network into node sets, load node sets, branch sets, and photovoltaic node sets.

[0034] Assuming the active distribution network is divided into Each region has a load node set of [number]. The set of photovoltaic nodes is B. ;

[0035] Among them, sub-regions In the middle, the set of nodes is The load node set is The set of branches is The photovoltaic node set is , ;

[0036] The boundary of an active distribution network is defined by a set of edges. To indicate, among which, Indicates the first The node and the first Edges between nodes;

[0037] The node voltage and photovoltaic active power are collected by smart meters. Photovoltaic reactive power Load active power and load reactive power .

[0038] Step 2: Predict the voltage and power of distribution network nodes using the Mamba algorithm with fused wavelet transform. The Mamba algorithm with fused wavelet transform includes feature mapping, state space construction, wavelet transform, construction of wavelet state transformation module with local attention mechanism, gating fusion, and state space model update steps.

[0039] Among them, the formula is:

[0040]

[0041] Pair of sub-regions Input Collect node data in real time and perform feature mapping;

[0042] in:

[0043]

[0044] Subregion The Data collected by smart meters during each control cycle Indicates learnable weights, Representing sub-regions Middle node Voltage, photovoltaic active power, photovoltaic reactive power, load active power, load reactive power;

[0045] Using the formula:

[0046]

[0047] Calculate the Mamba algorithm's first step. The state space of each regulatory cycle;

[0048] in, Subregion The The state space of each regulatory cycle Represents the state transition matrix. Represents the input matrix, Indicates that subregion m is in the first... The hidden state of a regulatory cycle;

[0049] Using the formula:

[0050]

[0051]

[0052] right Perform wavelet transform, where, and They represent the first The low-frequency and high-frequency components of the state space of each control cycle. and These represent the low-pass filter and the high-pass filter, respectively. Indicates the filter length. , Indicates the first The state space of each regulatory cycle;

[0053] In step two, wavelet transform extracts the low-frequency and high-frequency components of the state space through low-pass and high-pass filters, respectively; the wavelet state transformation module of the local attention mechanism generates enhanced representations of the low-frequency and high-frequency components through query, key-value vector calculation, and local window neighbor set filtering.

[0054] Using the formula:

[0055]

[0056]

[0057]

[0058]

[0059]

[0060]

[0061]

[0062] Construct a wavelet state transformation module based on a local attention mechanism;

[0063] in, , and These represent the low-frequency components. The query, key, and value vector. , and They represent high-frequency components respectively. The query, key, and value vector. , , , , and Represents low-frequency components With high frequency components Learnable parameters for the query, key, and value vectors. and These represent the low-frequency components. With high frequency components Attention weights Indicates the scaling factor. Indicates position The set of all neighbors within a local window. Represented as the first The final enhanced representation generated from the low-frequency components at each location, Represented as the first The final enhanced representation generated from the high-frequency components at each location, Indicates splicing characteristics, This indicates weighted output. express function, Represents the gate weight matrix. This represents the gating bias vector. Represents the fusion weight matrix. This represents element-wise product.

[0064] Step 2, gated fusion, fuses the output of the wavelet state transformation module of the local attention mechanism with the state space of the Mamba algorithm through the gated vector, the final gated weight matrix, and the final gated bias vector, and updates the predicted values ​​of node voltage and power.

[0065] Among them, the formula is:

[0066]

[0067]

[0068] The output of the wavelet state transformation module based on the local attention mechanism is gated and fused with the state space of the Mamba algorithm.

[0069] in, Represents the gate vector, This represents the final gating weight matrix. This represents the final gating bias vector. This represents the state space after gating fusion. ;

[0070] Using the formula:

[0071] ,

[0072] ;

[0073] Update the state-space model of the Mamba algorithm to obtain the active distribution network sub-region. Predicted values ​​of node voltage and power;

[0074] in, Represents the state transition matrix, Represents the input matrix, Represents the output matrix, Indicates active distribution network sub-area The first in Predicted output for each regulatory cycle, .

[0075] Step 3: Construct an optimization objective that minimizes node voltage disturbance deviation and control cost, set node voltage constraints and photovoltaic inverter power constraints, and reconstruct the active distribution network voltage control problem through a partially observable Markov model;

[0076] The node voltage constraint in step three defines the upper and lower bounds of the node voltage;

[0077] Photovoltaic inverter power constraints include upper and lower limits of photovoltaic active power, upper and lower limits of photovoltaic reactive power, and photovoltaic rated power and capacity factor constraints.

[0078] Specifically, using the formula:

[0079]

[0080] Build region The objective function is to minimize the mid-node voltage disturbance deviation and control cost. ;

[0081] in, For reference voltage, For the set of regulatory cycles, This represents the node voltage during the t-th control cycle;

[0082] Using the formula:

[0083]

[0084] Construct the objective function Node voltage constraints;

[0085] in, , These represent the upper and lower bounds of the voltage, respectively.

[0086] Using the formula:

[0087]

[0088]

[0089]

[0090]

[0091] Establish constraints on the power of photovoltaic inverters in active distribution network voltage control:

[0092] in, , Representing nodes respectively The upper and lower bounds of photovoltaic active power. , Representing nodes respectively The upper and lower bounds of photovoltaic reactive power. Represents a node Up to the rated power of photovoltaic, Indicates at node Photovoltaic reactive power control commands, Indicates the photovoltaic capacity factor;

[0093] Reconstruct the voltage control problem using a partially observable Markov model;

[0094] The partially observable Markov model in step three includes the state space, action space, local observation set, and reward function;

[0095] The state space contains photovoltaic reactive power control instructions on the nodes, and the action space corresponds to the set of photovoltaic reactive power control instructions for each intelligent agent.

[0096] Multi-agent set is represented as Each agent manages a section of the active distribution network;

[0097] Using the formula:

[0098]

[0099] Indicates the first In each control cycle, the voltage-controlled portion is observable in the state space of the Markov model. , Indicates active distribution network sub-area The first in Predicted output for each control cycle, Represents any active distribution network zone;

[0100] Using the formula:

[0101]

[0102]

[0103] Indicates the first One control cycle, the action space A in the observable Markov model of the voltage-controlled part;

[0104] in, Indicates the first The action space of an intelligent agent Indicates the first Each regulatory cycle at the node Photovoltaic reactive power control commands, Represents any node Subregion The set of photovoltaic nodes in the system;

[0105] Using the formula:

[0106]

[0107]

[0108]

[0109] Indicates the first Each control cycle, the local observation set in the partially observable Markov model of voltage control. ;

[0110] in, Indicates in the region The local observation set in Indicates in the region The active distribution network operating status measured in the middle;

[0111] Using the formula:

[0112]

[0113] Indicates the first In each control cycle, the reward function in the observable Markov model for voltage control is... .

[0114] Step 4: Using a centralized training-distributed execution architecture based on the MATD3 algorithm, each agent generates control actions based on local observations and global prediction information to achieve active voltage coordination control of the distribution network.

[0115] In step four, the MATD3 algorithm outputs control actions through the Actor network, calculates the evaluation function through the Critic network, samples samples using the experience replay pool, updates network parameters through policy gradient and Q value, and uses soft update coefficients to update the parameters of the TargetActor network and the TargetCritic network.

[0116] The Actor network retained after the MATD3 algorithm training converges is used for online decision-making. The Actor network of each agent realizes real-time optimization control of regional autonomy based on the information of the controlled sub-region.

[0117] Specifically, using the formula:

[0118]

[0119]

[0120] Computational intelligent agents The Action function of Actor network and evaluation function of Critic network for each regulation cycle ;

[0121] in, , These represent the parameters of the Actor network and the Critic network, respectively. Representation strategy, Indicates the first The agent in the th... Control actions output in each control cycle, Indicates the first The agent in the th... The state-action value function value corresponding to each regulation cycle Indicates the first The agent in the th... Local observation vectors for each regulation cycle Indicates the first The state space of each regulatory cycle Indicates the first The actions of all agents in a single control cycle constitute a composite action vector. The parameter is strategy function Indicates the first An Actor policy network of agents, with parameters as follows: , Indicates the number of agents;

[0122] Using the formula:

[0123]

[0124]

[0125] Update the policy gradient of the Actor network parameters and the Q-value of the Critic network;

[0126] in, Indicates the target Q value. Represents the gradient operator, The mathematical expectation operator is used to select the network with the smallest Q-value from the two Critic networks as the update target, and is expressed as:

[0127]

[0128] in, The parameters representing the Target Critic network, Indicates the first The agent in the th... The immediate rewards obtained during each regulatory cycle;

[0129] Using the formula:

[0130]

[0131]

[0132]

[0133]

[0134] Update the parameters of the Actor network and Target Actor network, as well as the Critic network and Target Critic network.

[0135] in, The parameters representing the Target Critic network, and The learning rates are for the Actor network and the Critic network, respectively. This is the soft update coefficient. Indicates the first The loss function of the Critic network for each agent;

[0136] During training, the number of samples randomly taken from the experience replay pool is... experience samples ;

[0137] Using the formula:

[0138]

[0139]

[0140] Computational intelligent agents The Actor network strategy and the gradient of the loss function of the Critic network;

[0141] After the voltage control algorithm converges, the network parameters are no longer updated; only the Actor network is retained for online decision-making. Each agent's Actor network only needs information about the controlled sub-region. This enables real-time optimization control of regional autonomy, outputting a set of all actions. .

[0142] Based on the above, an experiment on voltage control of an IEEE 33-node active distribution network based on Mamba-MATD3 was generated:

[0143] 1. Experimental environment and configuration:

[0144] This embodiment verifies the proposed method in an IEEE 33-node active distribution network simulation environment. The hardware configuration is: AMD Ryzen 7 5800H CPU (3.20GHz), NVIDIA GeForce RTX 3050 Laptop GPU, and 16.0GB of memory; the software environment is a Python 3.6 platform under Windows.

[0145] Topology and zoning of active distribution networks, such as Figure 2 As shown: The IEEE 33 nodes are divided into 5 regions, and distributed photovoltaics are connected at nodes 2, 8, 14, 23, 27 and 31; the root node voltage is 1.00 pu, and the node voltage constraint range is 0.95-1.05 pu.

[0146] 2. Data and algorithm parameter settings:

[0147] Datasets: The training data consisted of photovoltaic power output data from the Belgian Elia Group and residential electricity load data from Portugal; the test data consisted of a 30-minute segment randomly selected from the dataset to verify the distributed execution performance of the algorithm.

[0148] The parameters of the MATD3 algorithm are shown in Table 1:

[0149] Table 1 MATD3 Algorithm Parameters

[0150] Parameter name set up Training batch size 256 Experience replay pool capacity 50000 Strategy update frequency 2 Target strategy smoothing coefficient 0.2 Soft update coefficient 0.005 Discount factor 0.98 Actor Network Learning Rate 0.0001 Critic Network Learning Rate 0.0002

[0151] 3. Algorithm training process:

[0152] The training phase is set to 5000 rounds (each round containing 500 steps). The cumulative reward change during algorithm training is as follows: Figure 3 As shown:

[0153] Rounds 0-1000: Cumulative rewards fluctuate steadily. The agent is in the exploration phase, interacting with the environment through random actions and storing experience samples in the replay pool.

[0154] 1000-4000 rounds: The cumulative reward increases significantly, the agent learns effective control strategies, and the decision convergence speed is improved by combining the voltage-power prediction information of the Mamba model;

[0155] After 4000 rounds: the cumulative reward tends to stabilize, the Actor / Critic network parameters are optimized to the best, forming a generalizable voltage control strategy, and the system voltage deviation is effectively suppressed.

[0156] 4. Verification of control effect:

[0157] Reference Figure 4 The node voltage distribution of the five schemes, namely "no control", "TD3", "MADDPG", "MATD3" and "Mamba-MATD3", is compared and the results are as follows:

[0158] Uncontrolled: Node voltage exceeds limits;

[0159] TD3 (single agent): alleviates over-limit but lacks regional coordination, resulting in large voltage deviation;

[0160] MADDPG / MATD3 (Multi-Agent): Enables regional collaboration, but lacks a prediction mechanism, resulting in a delayed response to new energy fluctuations.

[0161] Mamba-MATD3: By integrating Mamba's time-series prediction with MATD3's multi-agent collaboration, voltage trends can be sensed in advance and controlled proactively, resulting in minimal node voltage fluctuations and deviations, and the most stable system operation, thus verifying the superiority of the method.

[0162] This embodiment demonstrates that the Mamba-MATD3 method can effectively address the intermittency of photovoltaic power and load fluctuations in active distribution networks, achieving stable and precise control of the global voltage.

[0163] The above specific embodiments are merely optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A method for active distribution network voltage control integrating Mamba prediction and multi-agent systems, characterized in that, Includes the following steps: Step 1: Divide the area based on the active distribution network topology, collect the node voltage, active power and reactive power of each area through smart meters, and preprocess the data. Step 2: Predict the voltage and power of distribution network nodes using the Mamba algorithm with fused wavelet transform. The Mamba algorithm with fused wavelet transform includes feature mapping, state space construction, wavelet transform, construction of wavelet state transformation module with local attention mechanism, gating fusion, and state space model update steps. Step 3: Construct an optimization objective that minimizes node voltage disturbance deviation and control cost, set node voltage constraints and photovoltaic inverter power constraints, and reconstruct the active distribution network voltage control problem through a partially observable Markov model; Step 4: Using a centralized training-distributed execution architecture based on the MATD3 algorithm, each agent generates control actions based on local observations and global prediction information to achieve active voltage coordination control of the distribution network. In step two, wavelet transform extracts the low-frequency and high-frequency components of the state space using low-pass and high-pass filters, respectively. The wavelet state transformation module of the local attention mechanism generates enhanced representations of low-frequency and high-frequency components through querying, key-value vector calculation, and filtering of local window neighbor sets. The second step, gated fusion, fuses the output of the wavelet state transformation module of the local attention mechanism with the state space of the Mamba algorithm through the gated vector, the final gated weight matrix, and the final gated bias vector, and updates the predicted values ​​of node voltage and power.

2. The active distribution network voltage control method integrating Mamba prediction and multi-agent technology as described in claim 1, characterized in that, The preprocessed data in step one includes the node voltage, photovoltaic active power, photovoltaic reactive power, load active power, and load reactive power of the sub-region. The sub-region is determined by dividing the active distribution network into node sets, load node sets, branch sets, and photovoltaic node sets.

3. The active distribution network voltage control method integrating Mamba prediction and multi-agent technology as described in claim 1, characterized in that, The node voltage constraint in step three defines the upper and lower bounds of the node voltage; Photovoltaic inverter power constraints include upper and lower limits of photovoltaic active power, upper and lower limits of photovoltaic reactive power, and photovoltaic rated power and capacity factor constraints.

4. The active distribution network voltage control method integrating Mamba prediction and multi-agent technology as described in claim 2, characterized in that, The partially observable Markov model in step three includes the state space, action space, local observation set, and reward function; The state space contains photovoltaic reactive power control instructions on the nodes, and the action space corresponds to the set of photovoltaic reactive power control instructions for each agent.

5. The active distribution network voltage control method integrating Mamba prediction and multi-agent technology according to claim 1, characterized in that, In step four, the MATD3 algorithm outputs control actions through the Actor network, calculates evaluation functions through the Critic network, samples samples using the experience replay pool, updates network parameters through policy gradients and Q-values, and uses soft update coefficients to update the parameters of the TargetActor network and the TargetCritic network.

6. The active distribution network voltage control method integrating Mamba prediction and multi-agent technology according to claim 5, characterized in that, The Actor network retained after the MATD3 algorithm training converges is used for online decision-making. The Actor network of each agent realizes real-time optimization control of regional autonomy based on the information of the controlled sub-region.

Images