Microgrid group two-stage active emergency self-healing learning method based on knowledge-intelligence fusion cooperation
Through the two-stage active emergency self-healing learning method of microgrid group with knowledge-intelligence fusion collaboration, the deep fusion neural network and multi-agent energy management system are used to solve the problem of reducing the toughness of the microgrid system in traditional methods, and realize the microgrid Independent learning and optimization improve safety and stability.
Patent Information
- Application Number
- CN202410853309.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-28
- Publication Date
- 2024-11-08
AI Technical Summary
Traditional centralized regulation strategies are difficult to effectively and safely manage large-scale distributed heterogeneous microgrid groups, especially in the face of local faults and large-scale fluctuations in the source charge, resulting in reduced system resilience, and existing methods rely heavily on microgrids. and accurate understanding of parameters cannot cope with the randomness and volatility of distributed power output, threatening voltage safety and stable operation.
A two-stage active emergency self-healing learning method for microgrid groups with knowledge-intelligence fusion collaboration is proposed. Through deep fusion of neural network autonomous identification algorithm and multi-intelligence group intelligent collaborative network reconstruction, the network is realized accurately and quickly reconstructed and recovered. The specific steps include: data-driven based on coupled uncertainty and spatiotemporal information, combining deep residual shrinking networks and deep adversarial neural networks for autonomous identification, forming an accident dynamic knowledge base; then dividing the agent partitions and building a multi-agent energy management system , and use potential game theory and neural network associative memory for network reconstruction.
It realizes independent learning and optimization of microgrids, improves the safety and stability of the system, reduces the occurrence of faults, improves the level of intelligence, and provides a solid foundation for the future development of microgrids.
Smart Images

Figure CN118920596A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of active safety autonomous learning of multi-dimensional heterogeneous microgrid groups, and relates to a two-stage active emergency self-healing learning method for a microgrid group, in particular to a knowledge-intelligence fusion and collaborative two-stage active emergency self-healing learning method for a microgrid group. Background Art
[0002] Large-scale distributed heterogeneous microgrid groups have the characteristics of highly open interconnection, strong uncertainty, deep integration of information and energy, and enhanced system vulnerability, which makes it difficult for traditional centralized control strategies to control them in a timely, effective, accurate and safe manner. It is urgent to develop artificial intelligence technology to enhance the generalization, collaboration and evolution capabilities of heterogeneous microgrid group models and algorithms in an open environment.
[0003] In response to the problem of reduced system resilience caused by local faults in microgrids and large-scale fluctuations in sources and loads, although model-based methods (such as model-based prediction, multi-time scale, etc.) have also achieved good performance in microgrid fault system regulation, they are heavily dependent on accurate understanding of microgrids and parameters, which is not practical for today's complex microgrid group systems.
[0004] In recent years, the randomness and volatility of distributed power output have seriously threatened the voltage safety and operational stability of microgrids. Therefore, a method that does not rely on models and allows microgrids to autonomously learn to cope with faults is needed.
[0005] After searching, no public documents of the prior art that are identical or similar to the present invention were found. Summary of the invention
[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and propose a two-stage active emergency self-healing learning method for a microgrid group with knowledge-intelligence fusion and collaboration. In the first stage, a data knowledge fusion-driven information-physical coupling uncertain parameter autonomous identification method is proposed. In the second stage, a multi-agent group intelligence collaborative network reconstruction method is proposed, thereby realizing the accurate and rapid reconstruction and recovery of the knowledge-intelligence fusion and collaboration network.
[0007] The present invention solves the practical problem by adopting the following technical solutions:
[0008] A two-stage active emergency self-healing learning method for a microgrid group with knowledge-intelligence fusion collaboration includes the following steps:
[0009] Step S1, based on the data drive of coupling uncertainty and spatiotemporal information, microgrid high-precision information, etc., combined with deep residual shrinkage network and deep adversarial neural network, a deep fusion neural network autonomous identification algorithm is proposed;
[0010] Step S2, based on step S1, using historical data to form microgrid uncertainty prior knowledge;
[0011] Step S3, data knowledge fusion driven, forming an accident dynamic knowledge base of uncertain dynamic parameters such as distributed power output, energy storage charging and discharging power, and randomness of load fluctuations;
[0012] Step S4, based on the optimized parameters, first divide the agent into partitions, then micro-build each agent, and finally form a multi-agent energy management system;
[0013] In step S5, based on potential game theory and neural network associative memory, a network reconstruction algorithm is proposed, and the algorithm is interactively coordinated with the multi-agent energy management system proposed in S4 to jointly form a multi-agent group intelligence collaborative network reconstruction framework, and finally complete the microgrid emergency self-healing learning.
[0014] Moreover, the specific steps of step S1 include:
[0015] Step S11, constructing a residual shrinkage module of a deep residual shrinkage network;
[0016] Input feature vector x l Entering the convolutional layer, x l Dimensions are C x W. x l The residual feature x is obtained through the conversion module l+2 ;
[0017] x l+1 =ReLU(ω l+1 x l +b l+1 )
[0018] x l+2 =ω l+2 ReLU(ω l+1 x l +b l+1 )+b l+2
[0019] Among them, ω l+1 Represent the weight parameters of the l+1th layer and the l+2th layer respectively. l+1 and b l+2 are the bias parameters of the l+1th and l+2th layers. l+2 By taking the absolute value and global average pooling, we get a vector of size C×1 After three layers of fully connected layers, batch normalization, activation functions ReLU and Sigmoid, we get a parameter vector α of size C×1, α∈(0,1), and a threshold of τ:
[0020]
[0021] α uses the shrinkage operator S and the gating parameter to perform a soft threshold operation to achieve shrinkage and obtain the sparse feature y;
[0022]
[0023] The soft thresholded y is added to the residual term to get the final output x l+3 , size is C×W.
[0024] Step S12, constructing and training a deep adversarial neural network to form a deep fusion neural network autonomous identification algorithm;
[0025] The training process of the entire network can be expressed as:
[0026] (1) The first step is to calculate the loss function V(G),
[0027] V(G)=(1-y)log(1-D(G(Z)))
[0028] Among them, y is the parameter of the balanced loss, z is the input, G(z) is the output of the generator, and D(G(z)) is the probability that the data G(z) generated by the generator is a real sample. Figure 3 In, G loss Represents the loss function V(G).
[0029] (2) The second step is to calculate the loss function V(D):
[0030] V(D)=-((1-y)log(1-D((G(Z)))+ylogD(x)))
[0031] Among them, x is the real sample, and D(x) represents the probability that the real sample data is the real sample. Figure 3 Middle, D loss Represents the loss function V(D).
[0032] (3) The third step is to calculate the total loss function, which is expressed as follows:
[0033] m G inm D axV(D,G)=E X~P data [logD(x)]+E X~Pdata [log(1-D(G(z)))]
[0034] When the input x of the discriminator comes from real data, the output is close to 1. When the input z of the discriminator comes from noise variables, the output is close to 0. The stronger the discriminative ability of D, the larger D(x) is, and the smaller D(G(z)) is, which leads to a larger V(D,G). When the generated data is very close to the real data and D cannot make a judgment, the training ends.
[0035] Moreover, the specific steps of step S2 include:
[0036] Step S21, historical data preprocessing and feature extraction: cleaning and marking the historical data to remove outliers and noise;
[0037] Step S22, using ensemble learning technology to integrate the uncertainty information in historical data into the experience pool of deep reinforcement learning.
[0038] Moreover, the specific steps of step S3 include:
[0039] Step S31, establish the microgrid operation cost minimization function as follows:
[0040] Distributed power supply output model:
[0041] P distributed =f(W weather ,X control )
[0042] Among them, P distributed is the output power of the distributed power source, W weather It is the weather conditions, X control is a control variable;
[0043] Load demand model:
[0044] L demand =g(T time ,T load ,X other )
[0045] Among them, L demand is the load demand, T time It's time, T load is the load type, X other are other relevant conditions, such as energy procurement, load response, energy reserves, and outage costs.
[0046] Under the premise of meeting the power supply demand of user loads, the overall operation cost of the microgrid is minimized as the objective function. The objective function includes five parts: battery is the energy storage system cost, C DE is the cost of distributed power equipment, C DG is the power generation cost of distributed generation, C eis the emission cost, C grid is the cost of exchanging electricity with the power grid. As shown in the formula:
[0047] f1=min(C battery +C DE +C DG +C e +C grid )
[0048] Among them, f1 represents the overall operating cost of the microgrid in a period.
[0049]
[0050] In the above formula, C o is the battery charging and discharging maintenance cost, P M It is the battery charging and discharging power.
[0051]
[0052] In the formula, F fuel represents the fuel cost, F main represents equipment cost, Fss represents energy storage cost, K i is the operation and maintenance coefficient of DG unit i, P i is the power of DG unit i.
[0053]
[0054] In the above formula, P DE is the power generated by distributed generation, E i is the penalty cost of type i exhaust gas generated by the equipment, λ i is the emission coefficient of type i exhaust gas.
[0055] Step S32, establish a penalty function for the imbalance of energy supply and demand in the microgrid as follows:
[0056] First, the energy supply and demand balance constraint is:
[0057] ∑P distributed =∑L demand
[0058] This constraint ensures that the total energy supply P in the microgrid distributed and total energy demand L demand balance.
[0059] Distributed power generation capacity constraints:
[0060] P distributed ≤P max
[0061] This constraint limits the output power of the distributed generation to not exceed its maximum capacity Pmax .
[0062] Load demand constraints:
[0063] L demand ≥L min_demand
[0064] This constraint ensures that the load demand meets the minimum demand L min-demand .
[0065] Energy storage system constraints:
[0066] e c ≤e max
[0067] e f ≤e max
[0068] c≤c max
[0069] Among them, e c and e f Represent the charge and discharge rate, e max is the maximum charge and discharge rate, c is the charge and discharge state of the energy storage system, c max is the capacity of the energy storage system. These constraints ensure that the charging and discharging rate and charging and discharging state of the energy storage system are within the specified range.
[0070] The penalty function penalty for the imbalance of energy supply and demand in the microgrid is:
[0071]
[0072] Among them, P i line,req It is a necessary power exchange for microgrids. is the penalty cost for deviating from microgrid dispatch, P i line is the actual power exchange of the microgrid, MG is the total number of partitions, T is the total number of nodes in the partition mg, and the penalty for mg∈(1,MG) in the total △t time is penalty.
[0073] Step S33, integrate the uncertainty parameter information of the operating cost minimization function, energy supply and demand balance constraint function, and penalty function obtained by analysis and modeling into a dynamic knowledge base to form a dynamic knowledge base of microgrid uncertainty accidents; for each uncertainty parameter, record its historical data, statistical characteristics, and dynamic change trend information.
[0074] Step S34, combining the accident dynamic knowledge base with the data knowledge fusion method in steps S1 and S2 to drive intelligent decision-making and optimization of the microgrid system.
[0075] Moreover, the specific steps of step S4 include:
[0076] Step S41, a heuristic method for partitioning intelligent agents is proposed, which uses a trial-and-error method to first divide the nodes according to geography to obtain preliminary partitions, one partition corresponds to one intelligent agent, and after the preliminary partitions, adjustments are made to obtain the final partitions;
[0077] The evaluation method is as follows:
[0078]
[0079] Among them, r n,t represents the power loss of the node, c loss and c switch represents the cost coefficient of power loss and switching operation; the constraint is described by the cost function, including switching loss and voltage loss. n,m,t ,u n,m,t-1 represents the switch state of the controlled switch m at the nth node at time t and time t-1 respectively; Mn represents the total number of switches, m∈[1,Mn], p t loss represents the active power loss of the node at time t;
[0080] Step S42, micro-constructing the agent individuals, and expressing the multi-agent cooperation problem as a Markov decision process in multi-agent deep reinforcement learning;
[0081] A Markov decision process is defined as a tuple (D, S, A, O, T, R, γ), where D = {1, ..., n} is a set of n agents, S is the state space; A = × i∈D A i is the joint action space of all agents, including the individual action space Ai of agent i; O× i∈D O i is the joint observation space, which includes the observation set Oi of agent i; T is the transition probability, R is the shared reward function of each agent, and γ is the discount factor. At each time step t, each agent i first obtains its own observation o according to the environment t i ∈Oi, then based on the random strategy π t i :O i ×A i →[0,1] select a t i ∈A i The strategy delivers actions at T:S×A1×···×A with unknown transition probabilities. n→S enters the next state. Next, receive the joint reward r t ∈R:S×A. The goal of Markov decision making is to find a joint optimal strategy π=× i∈D π i Achieve the maximum expected return J = E π where γ is a discount factor that balances the importance of long-term and short-term rewards, and T is the time horizon.
[0082] Step S43, establishing a multi-agent energy management system;
[0083] The model of the multi-agent energy management system is divided into three levels: reaction level, decision level and collaboration level. In this system, information and instructions flow mutually, and a two-way transmission mechanism is adopted: the intention and cognition of the agent are conveyed from top to bottom, while the instructions of decision and action are conveyed from bottom to top.
[0084] Moreover, the specific steps of step S5 include:
[0085] Step S51, evaluate the vulnerability of different intelligent agents in the microgrid, and perform system optimization and reconstruction within the range of [20,100];
[0086] Step S52, based on potential game theory and neural network associative memory, a network reconstruction algorithm is proposed to perform dynamic reconstruction of the microgrid;
[0087] In step S53, the network reconstruction algorithm of S52 is interactively coordinated with the multi-agent energy management system proposed by S4 to jointly form a multi-agent group intelligence collaborative network reconstruction framework, and finally complete the microgrid emergency self-healing learning.
[0088] Moreover, the specific steps of step S51 include:
[0089] Step S511, using neural network associative memory to identify the current state of the microgrid system; by learning historical data, the neural network can be used to predict the state of the current microgrid system at the next moment, including the power supply and demand of each node and the energy transfer path;
[0090] The neural network associative memory algorithm is as follows:
[0091] For a given area, the number of memory patterns is s, and the memory patterns p(1), p(2), ..., p(m) ∈ Dn, where Dn is the n-dimensional space of the microgrid system state containing n agents, m ∈ {1, 2, ..., s}, and the following associative memory algorithm is given:
[0092] (1)c i is the importance or priority of different agents in the microgrid system, a ijand b ij are the connection strength and power transmission strength between agents, respectively. i is the initial state of each intelligent agent in the microgrid system and the initial input data of solar panels, wind turbines and other equipment; i , a ij , b ij and I i , i, j = 1, 2, ..., n, so that c i > 0, CAB is a non-singular M matrix, where C = diag(c1,c2,...,c n ), A=(a ij ) n×n And B=(b ij ) n×n .
[0093] (2) Identify the voltage fluctuation region 1, 2, ..., s based on the characteristics of the memory pattern p(1), p(2), ..., p(s)∈Dn.
[0094] Step S512, perform node vulnerability analysis, select three index parameters: voltage fluctuation vulnerability, overload vulnerability, and load low power factor vulnerability, take normal operating conditions as the benchmark, and use the deviation degree of the three index parameters as the degree of hazard of the assessment system. The calculation formula of the vulnerability index of voltage fluctuation is: Y = 20|X-1|. That is, when the node voltage is the rated value, the risk value of the node is 0, and when it is within ±5% of the rated value, the risk value of the node is 1. The voltage fluctuation range is normal within (0.95, 1.05), and the calculation formula of the overload vulnerability index is: Y = 10 (X-0.9), otherwise, the calculation formula of the load low power factor vulnerability index is: Y = -10 (X-0.9).
[0095] Step S513, comprehensive vulnerability assessment; constructing the system vulnerability strength ISV calculation model as follows:
[0096]
[0097] Where: γ j represents the probability of the jth type of failure; i represents the structural vulnerability strength of node i; ω i The weight coefficient of the impact of the i-th evaluation index on the system vulnerability, i = 1, 2, 3; U fvb Indicates the deviation of voltage fluctuation; L fvb Indicates the degree of overload deviation; P fvb It indicates the deviation degree of low power factor; M and N indicate the number of faults and the number of nodes respectively.
[0098] Moreover, the specific steps of step S52 include:
[0099] Step S521, constructing a potential function;
[0100] The potential function is as follows:
[0101]
[0102] Among them, G(y) is the overall economic benefit of the microgrid, that is, the sum of the benefits of each intelligent agent, N is the number of nodes, and f i is the sum of the benefits at the node, including the wind power generation system benefit f wt , Photovoltaic power generation system benefits pv 、Battery energy storage system benefits de 、Diesel engine power generation system ba :
[0103] f i =f pv +f wt +f de +f ba
[0104] f wt =μP wt Δt-γ wt P wt Δt
[0105] f pv =μP pv Δt-γ pv P pv Δt
[0106] Where: μ is the unit electricity price; P wt is the output power per unit time of the wind turbine array, P pv is the output power per unit time of the photovoltaic array; γ wt is the unit maintenance cost of wind turbine,array,γ pv is the unit maintenance cost of the photovoltaic array; △t is the length of the time interval.
[0107] f ba =μP ba Δt-γ ba |P ba |Δt
[0108] Where: P ba is the output power of the battery, γ ba is the unit maintenance cost of the battery; P ba Negative means the battery is charged, P ba A positive value indicates that the battery is discharged.
[0109] f de=(μP de -ρ oil D oil -γ de P de )Δt
[0110] Where: The unit maintenance cost of diesel engine is γ de , the revenue function of the diesel engine is f de ; The unit price of fuel is ρ oil , D oil For fuel consumption.
[0111] Step S522: The potential game model of the microgrid needs to satisfy the supply and demand balance. A penalty factor is used to meet this condition. The penalty function is:
[0112]
[0113] Where: α is the penalty factor; ∑ k P k ,∑ m L m are the total output power and total load demand, respectively, and the penalty F is added to the reward r as a negative reward.
[0114] Step S523, convergence criterion of network optimization reconstruction: consistent network optimization reconstruction based on active power load distribution;
[0115] In the actual microgrid system, the active power load distribution consistency algorithm is as follows:
[0116]
[0117] Where: P i is the active power load assigned to agent i; ei is the rated active power load of the intelligent agent; m ij Expressed as weight coefficient, the matrix form is as follows:
[0118] R[k+1]=MR[k]
[0119] Where: M represents the state transfer matrix; R[k] represents the ratio matrix.
[0120] In order to improve the iterative performance of active power allocation, the error ei is introduced:
[0121]
[0122] Substituting the error ei into the active power load distribution consistency algorithm formula, we get the following formula:
[0123]
[0124] After repeated iterations, the active power load distribution ratio converges to a constant, as shown below:
[0125]
[0126] Where: P1*…P n * represents the actual output power of the intelligent agent; P1…Pn represents the maximum output power of the intelligent agent; K represents a constant. At this point, the network optimization reconstruction converges and the algorithm ends.
[0127] Advantages and beneficial effects of the present invention:
[0128] The present invention proposes a two-stage active emergency self-healing learning method for a microgrid group with knowledge-intelligence fusion and collaboration. First, based on the data drive of coupling uncertainty and spatiotemporal information, high-precision information of microgrids, etc., combined with deep residual shrinkage network and deep adversarial neural network, a deep fusion neural network autonomous identification algorithm is proposed. And further use historical data to form microgrid uncertainty prior knowledge; through data knowledge fusion drive, realize the accident dynamic knowledge base of uncertainty dynamic parameters such as distributed power output, energy storage charging and discharging power and randomness of load fluctuation, and the first stage is completed. Through the data-driven optimization of the first stage and the construction of the accident dynamic knowledge base, the uncertainty and dynamic parameter fluctuation in the microgrid can be effectively identified and responded to, the occurrence of accidents can be reduced, and thus the safety and stability of the microgrid can be improved. Taking the parameters optimized in the first stage as the basis for the optimization in the second stage, the intelligent agent partitions are first divided, and then each intelligent agent is micro-built, and finally a multi-agent energy management system is formed; based on potential game theory and neural network associative memory, a network reconstruction algorithm is proposed, and the algorithm is used to interact and collaborate with the multi-agent energy management system proposed by S4 to jointly form a multi-agent group intelligence collaborative network reconstruction framework, thereby completing the second stage. The present invention combines deep learning and neural network technology, enabling the microgrid to learn and optimize autonomously, forming a multi-agent collaborative network, improving the intelligence level of the microgrid, and laying a solid foundation for the future development of microgrids.
[0129] 2. The present invention provides a two-stage active emergency self-healing learning method for a microgrid group with knowledge-intelligence fusion and collaboration of heterogeneous microgrids, comprising: in the first stage, using the high-precision data of the microgrid to analyze the fault information of the microgrid group in real time, and constructing an accident dynamic knowledge base based on the deep fusion neural network autonomous identification algorithm and the microgrid uncertainty prior knowledge; in the second stage, studying the energy management system based on multi-agent deep reinforcement learning to solve the dependence of traditional algorithms on models; exploring the multiple mapping relationships between multi-agents and energy management systems, and studying the network reconstruction algorithm combined with potential game theory to achieve accurate and rapid reconstruction and recovery of the network with knowledge-intelligence fusion and collaboration. BRIEF DESCRIPTION OF THE DRAWINGS
[0130] Figure 1 It is a flow chart of the knowledge-intelligence fusion collaborative two-stage active emergency self-healing learning method for microgrid groups of the present invention;
[0131] Figure 2 It is a schematic diagram of the knowledge-intelligence fusion collaborative two-stage active emergency self-healing learning method for microgrid groups of the present invention;
[0132] Figure 3 It is a network structure diagram of the deep fusion neural network autonomous identification algorithm;
[0133] Figure 4 It is a schematic diagram of a multi-agent energy management system. DETAILED DESCRIPTION
[0134] The embodiments of the present invention are further described in detail below with reference to the accompanying drawings:
[0135] The present invention proposes a two-stage active emergency self-healing learning method for microgrid groups with knowledge-intelligence fusion and collaboration. In the first stage, the microgrid group fault information is analyzed in real time using the microgrid's high-precision data, and a dynamic accident knowledge base is constructed based on the deep fusion neural network autonomous identification algorithm and microgrid uncertainty prior knowledge; in the second stage, an energy management system based on multi-agent deep reinforcement learning is studied to solve the dependence of traditional algorithms on models; the multiple mapping relationships between multi-agents and energy management systems are explored, and a network reconstruction algorithm combined with potential game theory is studied to achieve accurate and rapid reconstruction and recovery of the network with knowledge-intelligence fusion and collaboration.
[0136] A two-stage active emergency self-healing learning method for microgrid groups with knowledge-intelligence fusion and collaboration, such as Figure 1 and Figure 2 As shown, the following steps are included:
[0137] Step S1, based on the data drive of coupling uncertainty and spatiotemporal information, microgrid high-precision information, etc., combined with deep residual shrinkage network and deep adversarial neural network, a deep fusion neural network autonomous identification algorithm is proposed.
[0138] The specific steps of step S1 include:
[0139] Step S11, constructing a residual shrinkage building unit (RSBU) of a deep residual shrinkage network (DRSN);
[0140] like Figure 3 As shown: Input feature vector x l Entering the convolutional layer, x l Dimensions are C x W. x lThe residual feature x is obtained through the conversion module (two convolutional layers, batch normalization layer, activation function ReLU) l+2 ;
[0141] x l+1 =ReLU(ω l+1 x l +b l+1 )
[0142] x l+2 =ω l+2 ReLU(ω l+1 x l +b l+1 )+b l+2
[0143] Among them, ω l+1 Represent the weight parameters of the l+1th layer and the l+2th layer respectively. l+1 and b l+2 are the bias parameters of the l+1th and l+2th layers. l+2 By taking the absolute value and global average pooling, we get a vector of size C×1 Then through x l+2 =ω l+2 ReLU(ω l+1 x l +b l+1 )+b l+2 After the fully connected layer, batch normalization, activation function ReLU and Sigmioid, we get a parameter vector α of size C×1, α∈(0,1), and a threshold of τ:
[0144]
[0145] α uses the shrinkage operator S and the gating parameters to perform a soft threshold operation to achieve shrinkage and obtain the sparse feature y.
[0146]
[0147] The soft thresholded y is added to the residual term to get the final output x l+3 , size is C×W.
[0148] Step S12, constructing and training a deep adversarial neural network to form a deep fusion neural network autonomous identification algorithm;
[0149] The Deep Convolutional Generative Adversarial Network (DCGAN) algorithm is an unsupervised deep learning algorithm consisting of two parts: a generator (G) and a discriminator (D). In the generator, random noise is input to generate data. The discriminator is used to determine whether the input data is real data or artificial data, so it can also be regarded as a binary classifier. The generator and the discriminator improve their imitation ability and discrimination ability respectively through the game learning process, such as Figure 3 As shown:
[0150] The training process of the entire network can be expressed as:
[0151] (1) The first step is to calculate the loss function V(G),
[0152] V(G)=(1-y)log(1-D(G(Z)))
[0153] Among them, y is the parameter of the balanced loss, z is the input, G(z) is the output of the generator, and D(G(z)) is the probability that the data G(z) generated by the generator is a real sample. Figure 3 In, G loss Represents the loss function V(G).
[0154] (2) The second step is to calculate the loss function V(D):
[0155] V(D)=-((1-y)log(1-D(G(Z))+ylogD(x)))
[0156] Among them, x is the real sample, and D(x) represents the probability that the real sample data is the real sample. Figure 3 Middle, D loss Represents the loss function V(D).
[0157] (3) The third step is to calculate the total loss function, which is expressed as follows:
[0158] m G inm D axV(D,G)=E X~P data [logD(x)]+E X~Pdata [log(1-D(G(z)))]
[0159] When the input x of the discriminator comes from real data, the output is close to 1. When the input z of the discriminator comes from noise variables, the output is close to 0. The stronger the discriminative ability of D, the larger D(x) is, and the smaller D(G(z)) is, which leads to a larger V(D,G). When the generated data is very close to the real data and D cannot make a judgment, the training ends.
[0160] Step S2: Based on step S1, the historical data is used to form microgrid uncertainty prior knowledge.
[0161] The specific steps of step S2 include:
[0162] Step S21, historical data preprocessing and feature extraction: clean and mark the historical data, remove outliers and noise, and ensure the accuracy and reliability of the data.
[0163] Extract features from historical data, such as load data, power generation, weather conditions, etc., so that the deep reinforcement learning model can better understand the state of the microgrid system.
[0164] Step S22, using integrated learning technology, integrates the uncertainty information in the historical data into the experience pool of deep reinforcement learning to improve the microgrid's perception and robustness of uncertainty.
[0165] Based on the uncertainty prior knowledge formed by historical data, deep reinforcement learning methods can more accurately perceive the uncertainty in the microgrid system and improve the accuracy of prediction and decision-making.
[0166] Step S3, driven by data knowledge fusion, forms an accident dynamic knowledge base of uncertain dynamic parameters such as distributed power output, energy storage charging and discharging power, and randomness of load fluctuations; at this point, the first stage is completed.
[0167] The specific steps of step S3 include:
[0168] Step S31, establish the microgrid operation cost minimization function as follows:
[0169] Distributed power supply output model:
[0170] P distributed =f(W weather ,X control )
[0171] Among them, P distributed is the output power of the distributed power source, W weather It is the weather conditions, X control is the control variable.
[0172] Load demand model:
[0173] Ldemand =g(T time ,T load ,X other )
[0174] Among them, L demand is the load demand, T time It's time, T load is the load type, X other are other relevant conditions, such as energy procurement, load response, energy reserves, and outage costs.
[0175] Under the premise of meeting the power supply demand of user loads, the overall operation cost of the microgrid is minimized as the objective function. The objective function includes five parts: battery is the energy storage system cost, C DE is the cost of distributed power equipment, C DG is the power generation cost of distributed generation, C e is the emission cost, C grid is the cost of exchanging electricity with the power grid. As shown in the formula:
[0176] f1=min(C battery +C DE +C DG +C e +C grid )
[0177] Among them, f1 represents the overall operating cost of the microgrid in a period.
[0178]
[0179] In the above formula, C o is the battery charging and discharging maintenance cost, P M It is the battery charging and discharging power.
[0180]
[0181] In the formula, F fuel represents the fuel cost, F main represents equipment cost, Fss represents energy storage cost, K i is the operation and maintenance coefficient of DG unit i, P i is the power of DG unit i.
[0182]
[0183] In the above formula, P DE is the power generated by distributed generation, E i is the penalty cost of type i exhaust gas generated by the equipment, λ i is the emission coefficient of type i exhaust gas.
[0184] Step S32, establish a penalty function for the imbalance of energy supply and demand in the microgrid as follows:
[0185] First, the energy supply and demand balance constraint is:
[0186] ∑P distributed =∑L demand
[0187] This constraint ensures that the total energy supply P in the microgrid distributed and total energy demand L demand balance.
[0188] Distributed power generation capacity constraints:
[0189] P distributed ≤P max
[0190] This constraint limits the output power of the distributed generation to not exceed its maximum capacity P max .
[0191] Load demand constraints:
[0192] L demand ≥L min_demand
[0193] This constraint ensures that the load demand meets the minimum demand L min-demand .
[0194] Energy storage system constraints:
[0195] e c ≤e max
[0196] e f ≤e max
[0197] c≤c max
[0198] Among them, e c and e f Represent the charge and discharge rate, e max is the maximum charge and discharge rate, c is the charge and discharge state of the energy storage system, c max is the capacity of the energy storage system. These constraints ensure that the charging and discharging rate and charging and discharging state of the energy storage system are within the specified range.
[0199] The penalty function penalty for the imbalance of energy supply and demand in the microgrid is:
[0200]
[0201] Among them, P i line,req It is a necessary power exchange for microgrids. is the penalty cost for deviating from microgrid dispatch, P i line is the actual power exchange of the microgrid, MG is the total number of partitions, T is the total number of nodes in the partition mg, and the penalty for mg∈(1,MG) in the total △t time is penalty.
[0202] Step S33, the uncertainty parameter information of the operating cost minimization function, energy supply and demand balance constraint function, and penalty function obtained by analysis and modeling is integrated into a dynamic knowledge base to form a dynamic knowledge base of microgrid uncertainty accidents. For each uncertainty parameter, its historical data, statistical characteristics (such as mean, variance), dynamic change trend and other information are recorded.
[0203] Step S34, combining the accident dynamic knowledge base with the data knowledge fusion method in steps S1 and S2 to drive intelligent decision-making and optimization of the microgrid system. At this point, the first stage ends.
[0204] Step S4, using the parameters obtained from the first stage optimization as the basis for the second stage optimization, first divide the agent into partitions, then micro-build each agent, and finally form a multi-agent energy management system.
[0205] The specific steps of step S4 include:
[0206] In step S41, a heuristic method for dividing agents is proposed. The nodes are first divided geographically to obtain preliminary partitions by trial and error. One partition corresponds to one agent. After the preliminary partitions, adjustments are made to obtain the final partitions.
[0207] Specifically, the microgrid is divided into several areas according to the default geographic location information. Then, each node is assigned a certain number of interconnected areas (geographical partitions). Because geographic partitioning cannot guarantee that the reward function and power loss reach the optimal solution, an evaluation will be established after the geographic partitioning. During the evaluation process, the nodes with power loss higher than the average value are recorded, and then this node is reallocated to other intelligent agent partitions with electrical connections to find another partition with the smallest power loss of the node. The above post-partitioning adjustment process will be repeated until the overall power loss reaches the optimal solution and the final partition is obtained.
[0208] The evaluation method is as follows:
[0209]
[0210] Among them, r n,t represents the power loss of the node, c loss and c switchrepresents the cost coefficient of power loss and switching operation. The constraints are described by the cost function, including switching loss and voltage loss. n,m,t ,u n,m,t-1 represents the switch state of the controlled switch m at the nth node at time t and time t-1 respectively. Mn represents the total number of switches, m∈[1,Mn], p t loss Represents the active power loss of the node at time t.
[0211] Step S42, micro-constructing the agent individuals. In multi-agent deep reinforcement learning, the multi-agent cooperation problem is usually expressed as a Markov decision process;
[0212] A Markov decision process is defined as a tuple (D, S, A, O, T, R, γ), where D = {1, ..., n} is a set of n agents, S is the state space; A = × i∈D A i is the joint action space of all agents, including the individual action space Ai of agent i; O× i∈D O i is the joint observation space, which includes the observation set Oi of agent i; T is the transition probability, R is the shared reward function of each agent, and γ is the discount factor. At each time step t, each agent i first obtains its own observation o according to the environment t i ∈Oi, then based on the random strategy π t i :O i ×A i →[0,1] select a t i ∈A i The strategy delivers actions at T:S×A1×···×A with unknown transition probabilities. n →S enters the next state. Next, receive the joint reward r t ∈R:S×A. The goal of Markov decision making is to find a joint optimal strategy π=× i∈D π i Achieve the maximum expected return J = E π where γ is a discount factor that balances the importance of long-term and short-term rewards, and T is the time horizon.
[0213] Step S43, establishing a multi-agent energy management system;
[0214] like Figure 4As shown in the figure, the model of the multi-agent energy management system is divided into three levels: reaction layer, decision layer and collaboration layer; in this system, information and instructions flow mutually, and a two-way transmission mechanism is adopted: the intention and cognition of the agent are conveyed from top to bottom, while the instructions of decision and action are conveyed from bottom to top. Such transmission not only reflects the importance of the agent's autonomous will, but also ensures the optimization of the overall performance of the system and the unity of the decision-making process. The system summarizes three different levels of consciousness, namely action consciousness, self-consciousness and social consciousness. In addition, the model also explains how the agent can refresh its knowledge base in real time by monitoring environmental changes. When a response is needed, the system executes or communicates through the "collaboration-decision-actuator" module to ensure the efficient operation of the system and the smooth progress of coordination.
[0215] In step S5, a network reconstruction algorithm is proposed based on potential game theory and neural network associative memory. This algorithm interacts and collaborates with the multi-agent energy management system proposed in S4 to form a multi-agent group intelligence collaborative network reconstruction framework, and finally completes the emergency self-healing learning of the microgrid.
[0216] The specific steps of step S5 include:
[0217] Step S51, evaluate the vulnerability of different intelligent agents in the microgrid, and perform system optimization and reconstruction within the range of [20,100];
[0218] The specific steps of step S51 include:
[0219] Step S511, using neural network associative memory to identify the current state of the microgrid system; by learning historical data, the neural network can be used to predict the state of the current microgrid system at the next moment, including the power supply and demand of each node, the energy transfer path, etc.;
[0220] The neural network associative memory algorithm is as follows:
[0221] For a given area, the number of memory patterns is s, and the memory patterns p(1), p(2), ..., p(m) ∈ Dn, where Dn is the n-dimensional space of the microgrid system state containing n agents, m ∈ {1, 2, ..., s}, and the following associative memory algorithm is given:
[0222] (1)c i is the importance or priority of different agents in the microgrid system, a ij and b ij are the connection strength and power transmission strength between agents, respectively, i is the initial state of each intelligent agent in the microgrid system and the initial input data of solar panels, wind turbines and other equipment; i, a ij , b ij and I i , i, j = 1, 2, ..., n, so that c i > 0, CAB is a non-singular M matrix, where C = diag(c1,c2,...,c n ), A=(a ij ) n×n And B=(b ij ) n×n .
[0223] (2) Identify the voltage fluctuation region 1, 2, ..., s based on the characteristics of the memory pattern p(1), p(2), ..., p(s)∈Dn.
[0224] Step S512, perform node vulnerability analysis, select three index parameters: voltage fluctuation vulnerability, overload vulnerability, and load low power factor vulnerability, take normal operating conditions as the benchmark, and use the deviation degree of the three index parameters as the degree of hazard of the assessment system. The calculation formula of the vulnerability index of voltage fluctuation is: Y = 20|X-1|. That is, when the node voltage is the rated value, the risk value of the node is 0, and when it is within ±5% of the rated value, the risk value of the node is 1. The voltage fluctuation range is normal within (0.95, 1.05), and the calculation formula of the overload vulnerability index is: Y = 10 (X-0.9), otherwise, the calculation formula of the load low power factor vulnerability index is: Y = -10 (X-0.9).
[0225] Step S513, comprehensive vulnerability assessment; constructing the system vulnerability strength ISV calculation model as follows:
[0226]
[0227] Where: γ j represents the probability of the jth type of failure; i represents the structural vulnerability strength of node i; ω i The weight coefficient of the impact of the i-th evaluation index on the system vulnerability, i = 1, 2, 3; U fvb Indicates the deviation of voltage fluctuation; L fvb Indicates the degree of overload deviation; P fvb It indicates the deviation degree of low power factor; M and N indicate the number of faults and the number of nodes respectively.
[0228] Step S52, based on potential game theory and neural network associative memory, a network reconstruction algorithm is proposed to perform dynamic reconstruction of the microgrid;
[0229] The specific steps of step S52 include:
[0230] Step S521, constructing a potential function; in the framework of potential game theory, microgrid reconstruction refers to the process in which participants in the microgrid system conduct strategic games and cooperation based on their respective interests and goals to reconfigure resources, optimize structures, and formulate operation strategies. This reconstruction process aims to maximize the overall benefits of the system and achieve the optimal system configuration and operation strategy through games and cooperation between participants. The potential function is as follows:
[0231]
[0232] Among them, G(y) is the overall economic benefit of the microgrid, that is, the sum of the benefits of each intelligent agent, N is the number of nodes, and f i is the sum of the benefits at the node, including the wind power generation system benefit f wt , Photovoltaic power generation system benefits pv 、Battery energy storage system benefits de 、Diesel engine power generation system ba :
[0233] f i =f pv +f wt +f de +f ba
[0234] f wt =μP wt Δt-γ wt P wt Δt
[0235] f pv =μP pv Δt-γ pv P pv Δt
[0236] Where: μ is the unit electricity price; P wt is the output power per unit time of the wind turbine array, P pv is the output power per unit time of the photovoltaic array; γ wt is the unit maintenance cost of wind turbine,array,γ pv is the unit maintenance cost of the photovoltaic array; △t is the length of the time interval.
[0237] f ba =μP ba Δt-γ ba |P ba |Δt
[0238] Where: P ba is the output power of the battery, γ ba is the unit maintenance cost of the battery; P ba Negative means the battery is charged, P baA positive value indicates that the battery is discharged.
[0239] f de =(μP de -ρ oil D oil -γ de P de )Δt
[0240] Where: The unit maintenance cost of diesel engine is γ de , the revenue function of the diesel engine is f de ; The unit price of fuel is ρ oil , D oil For fuel consumption.
[0241] Step S522: The potential game model of the microgrid needs to satisfy the supply and demand balance. A penalty factor is used to meet this condition. The penalty function is:
[0242]
[0243] Where: α is the penalty factor; ∑ k P k ,∑ m L m are the total output power and total load demand, respectively, and the penalty F is added to the reward r as a negative reward.
[0244] Step S523, convergence criterion of network optimization reconstruction: consistent network optimization reconstruction based on active power load distribution.
[0245] In the actual microgrid system, the active power load distribution consistency algorithm is as follows:
[0246]
[0247] Where: P i is the active power load assigned to agent i; ei is the rated active power load of the intelligent agent; m ij Expressed as weight coefficient, the matrix form is as follows:
[0248] R[k+1]=MR[k]
[0249] Where: M represents the state transfer matrix; R[k] represents the ratio matrix.
[0250] In order to improve the iterative performance of active power allocation, the error ei is introduced:
[0251]
[0252] Substituting the error ei into the active power load distribution consistency algorithm formula, we get the following formula:
[0253]
[0254] After repeated iterations, the active power load distribution ratio converges to a constant, as shown below:
[0255]
[0256] Where: P1*…P n *Indicates the actual output power of the intelligent system; P1…P n is represented by the maximum output power of the agent; K is represented by a constant. At this point, the network optimization reconstruction converges and the algorithm ends.
[0257] In step S53, the network reconstruction algorithm of S52 is interactively coordinated with the multi-agent energy management system proposed by S4 to jointly form a multi-agent group intelligence collaborative network reconstruction framework, and finally complete the microgrid emergency self-healing learning.
[0258] It should be emphasized that the embodiments of the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific implementation modes. Any other implementation modes derived by those skilled in the art based on the technical solutions of the present invention also fall within the scope of protection of the present invention.
Claims
1. A two-stage active emergency self-healing learning method for a microgrid group with knowledge-intelligence fusion collaboration, characterized by: The following steps are involved: Step S1, based on the data drive of coupling uncertainty and spatiotemporal information, microgrid high-precision information, etc., combined with deep residual shrinkage network and deep adversarial neural network, a deep fusion neural network autonomous identification algorithm is proposed; Step S2, based on step S1, using historical data to form microgrid uncertainty prior knowledge; Step S3, data knowledge fusion driven, forming an accident dynamic knowledge base of uncertain dynamic parameters such as distributed power output, energy storage charging and discharging power, and randomness of load fluctuations; Step S4, based on the optimized parameters, first divide the agent into partitions, then micro-build each agent, and finally form a multi-agent energy management system; In step S5, based on potential game theory and neural network associative memory, a network reconstruction algorithm is proposed, and the algorithm is interactively coordinated with the multi-agent energy management system proposed in S4 to jointly form a multi-agent group intelligence collaborative network reconstruction framework, and finally complete the microgrid emergency self-healing learning.
2. According to claim 1, a two-stage active emergency self-healing learning method for a microgrid group with knowledge-intelligence fusion and coordination is characterized by: The specific steps of step S1 include: Step S11, constructing a residual shrinkage module of a deep residual shrinkage network; Input feature vector x l Entering the convolutional layer, x l Dimensions: C×W; x l The residual feature x is obtained through the conversion module l+2 ; x l+1 =ReLU(ω l+1 x l +b l+1 ) x l+2 =ω l+2 ReLU(ω l+1 x l +b l+1 )+b l+2 Among them, ω l+1 Represent the weight parameters of the l+1th layer and the l+2th layer respectively; b l+1 and b l+2 is the bias parameter of the l+1th layer and the l+2th layer; x l+2 By taking the absolute value and global average pooling, we get a vector of size C×1 After three layers of fully connected layers, batch normalization, activation functions ReLU and Sigmoid, we get a parameter vector α of size C×1, α∈(0,1), and a threshold of τ: α uses the shrinkage operator S and the gating parameter to perform a soft threshold operation to achieve shrinkage and obtain the sparse feature y; The soft thresholded y is added to the residual term to get the final output x l+3 , size is C×W; Step S12, constructing and training a deep adversarial neural network to form a deep fusion neural network autonomous identification algorithm; The training process of the entire network can be expressed as: (1) The first step is to calculate the loss function V(G), V(G)=(1-y)log(1-D(G(Z))) Among them, y is the parameter of the balanced loss, z is the input, G(z) is the output of the generator, and D(G(z)) is the probability that the data G(z) generated by the generator is a real sample; in Figure 3, G loss Represents the loss function V(G); (2) The second step is to calculate the loss function V(D): V(D)=-((1-y)log(1-D(G(Z))+ylogD(x))) Where x is a real sample, and D(x) represents the probability that the real sample data is a real sample. In Figure 3, D loss Represents the loss function V(D); (3) The third step is to calculate the total loss function, which is expressed as follows: m G inm D axV(D,G)=E X~Pdata [logD(x)]+E X~Pdata [log(1-D(G(z)))] When the input x of the discriminator comes from real data, the output is close to 1; when the input z of the discriminator comes from noise variables, the output is close to 0; the stronger the discriminative ability of D, the larger D(x) is, and the smaller D(G(z)) is, which leads to a larger V(D,G); when the generated data is very close to the real data and D cannot make a judgment, the training ends.
3. According to claim 1, a two-stage active emergency self-healing learning method for a microgrid group with knowledge-intelligence fusion collaboration is characterized by: The specific steps of step S2 include: Step S21, historical data preprocessing and feature extraction: cleaning and marking the historical data to remove outliers and noise; Step S22, using ensemble learning technology to integrate the uncertainty information in historical data into the experience pool of deep reinforcement learning.
4. According to claim 1, a two-stage active emergency self-healing learning method for a microgrid group with knowledge-intelligence fusion collaboration is characterized by: The specific steps of step S3 include: Step S31, establish the microgrid operation cost minimization function as follows: Distributed power supply output model: P distributed =f(W weather ,X control ) Among them, P distributed is the output power of the distributed power source, W weather It is the weather conditions, X control is a control variable; Load demand model: L demand =g(T time ,T load ,X other ) Among them, L demand is the load demand, T time It's time, T load is the load type, X other are other relevant conditions, such as energy procurement, load response, energy reserves, and outage costs; Under the premise of meeting the power supply demand of user loads, the overall operation cost of the microgrid is minimized as the objective function. The objective function includes five parts: battery is the energy storage system cost, C DE is the cost of distributed power equipment, C DG is the power generation cost of distributed generation, C e is the emission cost, C grid is the cost of exchanging electricity with the grid, as shown in the following formula: f1=min(C battery +C DE +C DG +C e +C grid ) Among them, f1 represents the overall operation cost of the microgrid in a period; In the above formula, C o is the battery charging and discharging maintenance cost, P M The charging and discharging power of the battery; In the formula, F fuel represents the fuel cost, F main represents equipment cost, Fss represents energy storage cost, K i is the operation and maintenance coefficient of DG unit i, P i is the power of DG unit i; In the above formula, P DE is the power generated by distributed generation, E i is the penalty cost of type i exhaust gas generated by the equipment, λ i is the emission coefficient of type i exhaust gas; Step S32, establish a penalty function for the imbalance of energy supply and demand in the microgrid as follows: First, the energy supply and demand balance constraint is: ∑P distributed =∑L demand This constraint ensures that the total energy supply P in the microgrid distributed and total energy demand L demand balance; Distributed power generation capacity constraints: P distributed ≤P max This constraint limits the output power of the distributed generation to not exceed its maximum capacity P max ; Load demand constraints: L demand ≥L min_demand This constraint ensures that the load demand meets the minimum demand L min-demand ; Energy storage system constraints: and c ≤e max and f ≤e max c≤c max Among them, e c and e f Represent the charge and discharge rate, e max is the maximum charge and discharge rate, c is the charge and discharge state of the energy storage system, c max is the capacity of the energy storage system; these constraints ensure that the charging and discharging rate and charging and discharging state of the energy storage system are within the specified range; The penalty function penalty for the imbalance of energy supply and demand in the microgrid is: in, It is a necessary power exchange for microgrids. is the penalty cost for deviating from microgrid dispatch, is the actual power exchange of the microgrid, MG is the total number of partitions, T is the total number of nodes in the partition mg, and the penalty for mg∈(1,MG) in the total △t time is penalty; Step S33, integrating the uncertainty parameter information of the operating cost minimization function, energy supply and demand balance constraint function, and penalty function obtained by analysis and modeling into a dynamic knowledge base to form a dynamic knowledge base of microgrid uncertainty accidents; for each uncertainty parameter, record its historical data, statistical characteristics, and dynamic change trend information; Step S34, combining the accident dynamic knowledge base with the data knowledge fusion method in steps S1 and S2 to drive intelligent decision-making and optimization of the microgrid system.
5. According to claim 1, a two-stage active emergency self-healing learning method for a microgrid group with knowledge-intelligence fusion collaboration is characterized by: The specific steps of step S4 include: Step S41, a heuristic method for partitioning intelligent agents is proposed, which uses a trial-and-error method to first divide the nodes according to geography to obtain preliminary partitions, one partition corresponds to one intelligent agent, and after the preliminary partitions, adjustments are made to obtain the final partitions; The evaluation method is as follows: Among them, r n,t represents the power loss of the node, c loss and c switch represents the cost coefficient of power loss and switching operation; the constraint is described by the cost function, including switching loss and voltage loss; u n,m,t ,u n,m,t-1 represents the switch state of the controlled switch m at the nth node at time t and time t-1 respectively; Mn represents the total number of switches, m∈[1,Mn], represents the active power loss of the node at time t; Step S42, micro-constructing the agent individuals, and expressing the multi-agent cooperation problem as a Markov decision process in multi-agent deep reinforcement learning; A Markov decision process is defined as a tuple (D, S, A, O, T, R, γ), where D = {1, ..., n} is a set of n agents, S is the state space; A = × i∈D A i is the joint action space of all agents, including the individual action space Ai of agent i; O× i∈D O i is the joint observation space, which includes the observation set Oi of agent i; T is the transition probability, R is the shared reward function of each agent, and γ is the discount factor; at each time step t, each agent i first obtains its own observation o according to the environment t i ∈Oi, then based on the random strategy choose The strategy delivers actions at T:S×A1×···×A with unknown transition probabilities. n →S enters the next state; next, receives the joint reward r t ∈R:S×A; The goal of Markov decision making is to find a joint optimal strategy π=× i∈D π i Achieve maximum expected profit Where γ is a discount factor that balances the importance of long-term and short-term rewards, and T is the time horizon; Step S43, establishing a multi-agent energy management system; The model of the multi-agent energy management system is divided into three levels: reaction level, decision level and collaboration level. In this system, information and instructions flow mutually, and a two-way transmission mechanism is adopted: the intention and cognition of the agent are conveyed from top to bottom, while the instructions of decision and action are conveyed from bottom to top.
6. According to claim 1, a two-stage active emergency self-healing learning method for a microgrid group with knowledge-intelligence fusion collaboration is characterized by: The specific steps of step S5 include: Step S51, evaluate the vulnerability of different intelligent agents in the microgrid, and perform system optimization and reconstruction within the range of [20,100]; Step S52, based on potential game theory and neural network associative memory, a network reconstruction algorithm is proposed to perform dynamic reconstruction of the microgrid; In step S53, the network reconstruction algorithm of S52 is interactively coordinated with the multi-agent energy management system proposed by S4 to jointly form a multi-agent group intelligence collaborative network reconstruction framework, and finally complete the microgrid emergency self-healing learning.
7. A two-stage active emergency self-healing learning method for a microgrid group with knowledge-intelligence fusion and collaboration according to claim 6, characterized in that: The specific steps of step S51 include: Step S511, using neural network associative memory to identify the current state of the microgrid system; by learning historical data, the neural network can be used to predict the state of the current microgrid system at the next moment, including the power supply and demand of each node and the energy transfer path; The neural network associative memory algorithm is as follows: For a given area, the number of memory patterns is s, and the memory patterns p(1), p(2), ..., p(m) ∈ Dn, where Dn is the n-dimensional space of the microgrid system state containing n agents, m ∈ {1, 2, ..., s}, and the following associative memory algorithm is given: (1)c i is the importance or priority of different agents in the microgrid system, a ij and b ij are the connection strength and power transmission strength between agents, respectively. i is the initial state of each intelligent agent in the microgrid system and the initial input data of solar panels, wind turbines and other equipment; i , a ij , b ij and I i , i, j = 1, 2, ..., n, so that c i > 0, CAB is a non-singular M matrix, where C = diag(c1,c2,...,c n ), A=(a ij ) n×n And B=(b ij ) n×n ; (2) Identify the voltage fluctuation region 1, 2, ..., s according to the characteristics of the memory pattern p(1), p(2), ..., p(s)∈Dn; Step S512, perform node vulnerability analysis, select three index parameters of voltage fluctuation vulnerability, overload vulnerability, and load low power factor vulnerability, take normal operating conditions as the benchmark, and use the deviation degree of the three index parameters as the degree of hazard of the assessment system; the calculation formula of the voltage fluctuation vulnerability index is: Y=20|X-1|; that is, when the node voltage is the rated value, the risk value of the node is 0, and when it is within ±5% of the rated value, the risk value of the node is 1; the voltage fluctuation range is normal within (0.95, 1.05), the calculation formula of the overload vulnerability index is: Y=10(X-0.9), otherwise, the calculation formula of the load low power factor vulnerability index is: Y=-10(X-0.9); Step S513, comprehensive vulnerability assessment; constructing the system vulnerability strength ISV calculation model as follows: Where: γ j represents the probability of the jth type of failure; i represents the structural vulnerability strength of node i; ω i The weight coefficient of the impact of the i-th evaluation index on the system vulnerability, i = 1, 2, 3; U fvb Indicates the deviation of voltage fluctuation; L fvb Indicates the degree of overload deviation; P fvb It indicates the deviation degree of low power factor; M and N indicate the number of faults and the number of nodes respectively.
8. The two-stage active emergency self-healing learning method for a microgrid group with knowledge-intelligence fusion and collaboration according to claim 6 is characterized by: The specific steps of step S52 include: Step S521, constructing a potential function; The potential function is as follows: Among them, G(y) is the overall economic benefit of the microgrid, that is, the sum of the benefits of each intelligent agent, N is the number of nodes, and f i is the sum of the benefits at the node, including the wind power generation system benefit f wt , Photovoltaic power generation system benefits pv 、Battery energy storage system benefits de 、Diesel engine power generation system ba : f i =f pv +f wt +f de +f ba f wt =μP wt Δt-γ wt P wt Δt f pv =μP pv Δt-γ pv P pv Δt Where: μ is the unit electricity price; P wt is the output power per unit time of the wind turbine array, P pv is the output power per unit time of the photovoltaic array; γ wt is the unit maintenance cost of wind turbine,array,γ pv is the unit maintenance cost of the photovoltaic array; △t is the length of the time interval; f ba =μP ba Δt-γ ba |P ba |Δt Where: P ba is the output power of the battery, γ ba is the unit maintenance cost of the battery; P ba Negative means the battery is charged, P ba A positive value indicates that the battery is discharged; f de =(μP de -r oil D oil -c de P de )Δt Where: The unit maintenance cost of diesel engine is γ de , the profit function of the diesel engine is f de ; The unit price of fuel is ρ oil , D oil For fuel consumption; Step S522: The potential game model of the microgrid needs to satisfy the supply and demand balance. A penalty factor is used to meet this condition. The penalty function is: Where: α is the penalty factor; ∑ k P k ,∑ m L m are the total output power and total load demand, respectively. The penalty F is added to the reward r as a negative reward; Step S523, convergence criterion of network optimization reconstruction: consistent network optimization reconstruction based on active power load distribution; In the actual microgrid system, the active power load distribution consistency algorithm is as follows: Where: P i is the active power load assigned to agent i; ei is the rated active power load of the intelligent agent; m ij Expressed as weight coefficient, the matrix form is as follows: R[k+1]=MR[k] Where: M represents the state transfer matrix; R[k] represents the ratio matrix; In order to improve the iterative performance of active power allocation, the error ei is introduced: Substituting the error ei into the active power load distribution consistency algorithm formula, we get the following formula: After repeated iterations, the active power load distribution ratio converges to a constant, as shown below: Where: P1*…P n *Indicates the actual output power of the intelligent system; P1…P n It is represented as the maximum output power of the intelligent agent; K is represented as a constant; at this point, the network optimization reconstruction converges and the algorithm ends.
Citation Information
Cited By
Power grid energy management optimization platform based on big data
CN119647871A
