Model-data hybrid driving-based regional energy internet toughness online evaluation method

Abstract

The invention belongs to the technical field of energy internet, and particularly relates to a regional energy internet toughness online evaluation method based on model-data hybrid driving. According to the method, firstly, a comprehensive energy network physical model is constructed, a toughness evaluation index system is established, the physical model covers energy nodes and line models, and energy activities are simulated through the law of conservation of energy and the like; and then carrying out real-time monitoring and risk early warning, collecting data by using an intelligent sensor, and predicting the risk in combination with machine learning. Then, based on physical model output and real-time data, a deep learning algorithm is adopted to evaluate network toughness, and a CNN-RNN model, adversarial training and Bayesian reasoning are fused to determine a weight; and finally, according to an evaluation result, applying a hybrid optimization algorithm and a reinforcement learning optimization strategy, and considering market factors to adjust energy operation. The advantages of the model and the data are integrated, the toughness evaluation accuracy, the risk early warning timeliness and the strategy optimization effectiveness are effectively improved, and the method is of great significance to energy internet development.

Description

Technical Field

[0001] The present invention belongs to the technical field of energy internet, and in particular relates to resilience assessment and optimization of regional energy systems, and specifically to an online resilience assessment method for regional energy internet based on model-data hybrid drive. Background Art

[0002] In the rapid development of the energy industry, the scale of regional energy Internet continues to expand, the structure becomes more complex, and the challenges it faces are becoming increasingly severe. Traditional energy network assessment methods have many defects and cannot meet the needs of modern energy management.

[0003] On the one hand, some existing assessment methods rely too much on physical models. Although these models can simulate energy flow, conversion and equipment operation based on theoretical and empirical parameters, in actual applications, due to the complexity and uncertainty of the energy system, such as dynamic changes in energy demand, sudden equipment failures and external environmental interference, the prediction results that rely solely on physical models deviate greatly from the actual situation. For example, in the face of sudden natural disasters or unexpected equipment downtime, physical models cannot accurately reflect the actual operating status and potential risks of the system in a timely manner, resulting in an inaccurate assessment of energy network resilience and difficulty in providing a reliable basis for subsequent decision-making.

[0004] On the other hand, although some data-driven methods can use a large amount of operating data to mine potential laws, they lack an in-depth understanding and effective integration of the physical nature of energy systems. Such methods are easily affected by issues such as data quality, data integrity, and data noise during data processing, and may draw one-sided or inaccurate conclusions. Moreover, when faced with new operating conditions or special situations not covered in the training data, their generalization ability is weak, and they cannot effectively cope with the complex and changing energy network environment, and it is difficult to comprehensively evaluate the network's resilience performance under different disturbances.

[0005] In summary, the existing energy network assessment technology has obvious deficiencies in accuracy, reliability and adaptability. A new method is urgently needed to overcome these shortcomings and achieve effective assessment and optimization of regional energy Internet resilience. Summary of the invention

[0006] In view of the above deficiencies in the prior art, the purpose of the present invention is to provide an online assessment method for regional energy Internet resilience based on model-data hybrid drive to solve the problem that the energy network assessment method in the prior art relies on a single physical model resulting in large deviations, or the data-driven method has limitations.

[0007] In order to achieve the above objectives, the present invention adopts the following technical solutions:

[0008] The online assessment method of regional energy internet resilience based on model-data hybrid drive includes the following steps:

[0009] S1. Build a basic model and develop a physical model of a comprehensive energy network. The physical model integrates key parameters such as energy flow, capacity and failure probability, and simulates energy flow and conversion through system dynamics principles. Based on the physical model, establish a set of network resilience assessment indicators.

[0010] S2. Real-time monitoring and risk warning, by deploying smart sensors to collect energy network operation data, and using machine learning technology to predict potential risks and failure points, and quickly respond to abnormal situations;

[0011] S3. Network resilience assessment, based on real-time monitoring and risk warning, uses deep learning algorithms combined with the physical model output of the integrated energy network and the real-time data of the current energy network collected by the real-time monitoring system in S2 to assess network resilience;

[0012] S4. Strategy optimization and adjustment: Based on the resilience assessment results of S3, use optimization algorithms and reinforcement learning methods to formulate and adjust optimization strategies to improve the resilience of the energy network and ensure that the network can recover quickly in the face of various disturbances.

[0013] Furthermore, the physical model of the integrated energy network includes energy node and energy line models;

[0014] The method for constructing the energy node model is as follows: for each node in the energy network, its input energy flow vector, output energy flow vector and stored energy amount are defined. According to the law of conservation of energy, the energy dynamic balance equation of the node is:

[0015]

[0016] S n (t+Δt) The amount of energy stored in node n at time t+Δt; I n,i (t) is the i-th input energy flow of node n at time t; n,j (t) is the jth output energy flow of node n at time t; S n (t) The amount of energy stored in node n at time t; Δt is the time step; η out is the conversion efficiency of the output energy flow; η in is the conversion efficiency of the input energy flow;

[0017] The formula represents the dynamic balance of node energy. The amount of energy stored in the node at time t+Δt is equal to the amount of energy stored in the node at time t plus the total amount of input energy flow adjusted by conversion efficiency, minus the total amount of output energy flow adjusted by conversion efficiency.

[0018] Furthermore, the energy line model construction method is as follows: for the energy line connecting the nodes, its transmission capacity, transmission loss coefficient and failure probability are defined, and the energy transmission amount on the line satisfies the following relationship:

[0019]

[0020] T i (t) is the energy transmission amount of line l at time t; Q m,j (t) is the i-th output energy flow of the line starting node m at time t; M s is the node set at the starting end of the line; C l is the transmission capacity; α l is the transmission loss coefficient;

[0021] The formula represents the energy transmission amount on the line. The energy transmission amount of the line at time t is equal to (1-α l ) multiplied by the line transmission capacity C l The sum of the energy flows output by the node at the start of the line The smaller value in .

[0022] The network resilience evaluation index system includes resilience quantitative index and fault impact evaluation index.

[0023] The resilience quantitative index is defined as a comprehensive measure of the network's ability to maintain normal energy supply and its ability to recover after disturbances within a given time interval. In addition to considering the network's ability to maintain normal energy supply and its ability to recover after disturbances, the resilience quantitative index also takes into account the quality factor of energy supply. The specific calculation formula is:

[0024]

[0025] Among them, t re The time required for the network to recover from a fault state to a normal operating state; E n,def (t) is the energy shortage of node n at time t; E n,nom (t) is the normal energy demand of node n; Q elec (t) is the power quality indicator.

[0026] Furthermore, the fault impact assessment index includes a fault impact range index and a fault severity index:

[0027] The fault impact range index evaluation index is used to measure the ratio of the number of nodes affected after a fault occurs to the total number of nodes, namely:

[0028]

[0029] Among them, FIR represents the fault impact range index; N affected is the number of affected nodes; N total is the total number of nodes in the network;

[0030] The fault severity index is calculated based on the ratio of the total energy shortage caused by the fault to the total normal energy supply of the network:

[0031]

[0032] Among them, t fail The time when the fault occurred.

[0033] Furthermore, in S2, the method of real-time monitoring and risk warning includes:

[0034] S21. Intelligent sensor deployment and data collection, using graph theory-based methods to determine the optimal deployment location of sensors, constructing the topology of the energy network G = (V, E), where V is the node set and E is the edge set; defining the node importance index I v , comprehensively considering factors such as node connectivity, energy flow and impact on system stability; the node importance index I v It is expressed as:

[0035]

[0036] k v The connectivity of the node; F in,v is the energy inflow; F out,v Energy outflow; ω 1 and ω 2 is the weight coefficient;

[0037] Then, the nodes are sorted according to the importance index, and sensors are deployed preferentially on important nodes and their adjacent edges to ensure that network operation data can be collected comprehensively and accurately;

[0038] S22. Adjust the frequency of data collection. According to the dynamic characteristics of the energy network and the historical data of fault occurrence, the frequency of data collection is increased for areas with large fluctuations in energy flow or frequent faults; for relatively stable areas, the frequency is appropriately reduced;

[0039] S23. The machine learning risk prediction model is constructed, combining support vector machine (SVM) and naive Bayes (NB) for risk prediction; for normal operation data, SVM is used for classification and anomaly detection; for fault data with certain probability distribution characteristics, NB is used for prediction; through a decision fusion module, the prediction results of the two are combined according to the characteristics of the data and the confidence of the model, and the formula is expressed as:

[0040]

[0041] Among them, y S is the prediction result of SVM; y N is the prediction result of NB; C S and C N are the confidence of the prediction results of SVM and NB models respectively; y pred Final prediction results.

[0042] The method for evaluating network resilience in S3 includes:

[0043] S31. The convolutional neural network (CNN) and recurrent neural network (RNN) fusion model uses CNN to extract the spatial features of the energy network physical model output and real-time data, and RNN to capture the time series features. For the input data, firstly, the convolution operation is performed on the CNN layer to obtain the feature map; then, the feature map is flattened and input into the RNN layer for time series processing, then:

[0044] h RNN (t) = RNN(x CNN (t),h RNN (t-1));

[0045] Among them, x CNN (t) is the output of the CNN layer; h RNN (t) is the hidden state of the RNN layer;

[0046] Finally, the RNN output is converted into the network resilience evaluation result through the fully connected layer;

[0047] S32. Adversarial training enhancement, introduces the idea of ​​generative adversarial network (GAN), constructs a discriminator and a generator, where the generator is used to generate false data similar to the fault situation, and the discriminator is used to distinguish the real data from the data generated by the generator; during the training process, by continuously optimizing the generator and the discriminator, the evaluation model can better cope with complex fault situations. The loss function of the discriminator is:

[0048]

[0049] The loss function of the generator is:

[0050]

[0051] Train the model by minimizing the loss function of the discriminator and the generator;

[0052] Among them, L D is the loss function of the discriminator; x is the input data; z is a random noise vector; D(x) is the output of the discriminator, indicating the probability that the input data x is the real data; G(z) is the output of the generator; pd (x) is the probability distribution of the real data x; p z (z) is the probability distribution of noise z; E represents expectation;

[0053] S33. Bayesian reasoning is used to determine the weights of the physical model and real-time data in the evaluation. According to Bayesian theorem, the posterior distribution is:

[0054] p(RI model ,RI data |RI final )∝p(RI final |RI model ,RI data )p(RI model )p(RI data );

[0055] Among them, RI data is the evaluation result based on real-time data; RI model is the evaluation result of the physical model; RI final is the final resilience evaluation result; RI model and RI data Obey Gaussian distribution; prior distribution is p(RI model )~N(μ model ,σ 2 model ) and p(RIdata)~N(μ data ,σ 2 data );

[0056] By approximately calculating the posterior distribution, we can obtain the weights ω of the physical model and real-time data. model and ω data , satisfying ω model +ω data =1, and RI final =ω model RI model +ω data RI data ;

[0057] Among them, RI model Represents the evaluation result based on the physical model; RI data Indicates the evaluation results based on real-time data; RI final Represents the final toughness assessment result.

[0058] Furthermore, the strategy optimization and adjustment step in S4 includes:

[0059] S41. Optimization algorithm, introduces a hybrid optimization algorithm that combines simulated annealing algorithm with genetic algorithm; in the iterative process of genetic algorithm, the Metropolis criterion of simulated annealing algorithm is used to accept inferior solutions with a certain probability, increase the diversity of the population, and avoid falling into the local optimum; for each generation of new individuals x generated by genetic algorithm new , calculate its fitness f(x new ), and the current optimal individual x best The fitness comparison f(x best ),if Where T is the temperature parameter of simulated annealing, then accept x new As the new current optimal individual; at the same time, considering the impact of energy market price fluctuations on the optimization strategy, the energy cost term is added to the optimization objective function, and the energy production, storage and transmission strategies are dynamically adjusted according to the real-time energy market price. The objective function becomes:

[0060]

[0061] Among them, RI is the toughness index; C total is the total cost; P marke t is the energy market price at time t; E pu (t) is the amount of energy purchased; α, β, γ are weight coefficients;

[0062] S42. Reinforcement learning environment and agent design;

[0063] S421. Environmental modeling: The operating status, fault conditions and resilience assessment results of the energy network are used as the environmental state of reinforcement learning. The state space of the environmental state is defined as:

[0064] S={s 1 ,s 2 ,…,s n};

[0065] si includes node energy storage, line transmission capacity utilization, current fault information and network resilience index;

[0066] Add market environment information to the state of the reinforcement learning environment, such as energy price trends, energy policy changes, etc.

[0067] Define the action space as a set of possible optimization strategies, such as adjusting energy production plans, changing energy storage strategies, optimizing line transmission allocation, etc.

[0068] S422. Agent design, using the deep Q network (DQN) or deep deterministic policy gradient (DDPG) algorithm to build an agent. The agent selects actions based on the state of the environment and learns the optimal strategy through interaction with the environment. In DQN, the Q value function of the agent is:

[0069]

[0070] Q(s,a;θ) is the state-action value function; it represents the expected cumulative reward of taking action a in a given state S, where θ is the parameter of the function; E r,τ is the expectation of reward r and discount factor τ; r represents the reward obtained immediately after the agent takes action a in the current state S; the discount factor τ is used to measure the importance of future rewards relative to current rewards; s′ is the next state; a′ is the next action; θ′ is the neural network parameter of the next Q function; It is to take the maximum value of all possible next actions a′;

[0071] Learn effective strategies to improve the resilience of energy networks. In DDPG, the agent consists of an actor network (Actor) and a critic network (Critic). The actor network outputs actions based on the state, and the critic network evaluates the value of the action. The strategy is optimized through the collaborative training of the two.

[0072] In summary, due to the adoption of the above technical solution, the beneficial technical effects of the invention are:

[0073] In terms of the accuracy of resilience assessment, the physical model and real-time monitoring data of the integrated energy network are integrated to give full play to the advantages of both. The physical model accurately simulates the flow and conversion of energy based on the principles of system dynamics, providing a solid theoretical basis for the assessment; the real-time data reflects the current actual state of the energy network and captures the dynamic changes and emergencies in the system operation. The combination of the two enables the network resilience assessment to more accurately reflect the true response capability and recovery speed of the energy network in the face of various disturbances. For example, when assessing the resilience of a regional energy network under extreme weather conditions, the physical model can provide the theoretical operating parameters of energy equipment and the expected change trend of energy flow under such conditions, while the real-time data can provide real-time feedback on actual equipment failure information, energy supply shortages, and energy demand fluctuations at the user end, so that the resilience assessment results are more in line with the actual situation and provide a reliable basis for subsequent strategy formulation.

[0074] In terms of the timeliness of risk warning, the deployment of intelligent sensors and machine learning technology have been used to achieve effective prediction and rapid response to potential risks and fault points. The sensor deployment method ensures the effective monitoring of key nodes and lines, and can collect comprehensive and accurate operation data in a timely manner. Combining the risk prediction model of support vector machine and naive Bayes, it can effectively analyze and judge different types of data, and issue early warning signals in time before a fault occurs. This enables operation and maintenance personnel to take measures in advance to reduce the possibility of failure or mitigate the impact of failure, effectively ensuring the safe and stable operation of the energy network. For example, when the model predicts that a transmission line may fail due to line aging and overload in the future, the operation and maintenance personnel can adjust the energy transmission path or maintain the line in advance to avoid failures and ensure the continuity of energy supply.

[0075] In terms of the effectiveness of strategy optimization, the optimization algorithm and reinforcement learning method are used to formulate and adjust the optimization strategy based on the resilience assessment results, which significantly improves the resilience of the energy network. The hybrid optimization algorithm introduced can avoid falling into the local optimal solution during the optimization process and search for a better strategy space. At the same time, considering the price fluctuations in the energy market and other practical factors, the energy production, storage and transmission strategies are dynamically adjusted to achieve the goal of reducing costs and improving energy efficiency while ensuring the reliability of energy supply.

[0076] In addition, the method of the present invention can adapt to the dynamic changes and development of the energy network by continuously updating and improving the evaluation index system, optimizing the model structure and algorithm parameters. It has good adaptability and scalability, and provides strong technical support for the sustainable development of regional energy Internet.

[0077] In summary, the present invention effectively solves the deficiencies in the prior art and has important application value and practical significance in the field of energy Internet resilience assessment and optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 A flowchart of the online assessment method for regional energy internet resilience based on model-data hybrid drive;

[0079] Figure 2 A flowchart of the real-time monitoring and risk early warning method;

[0080] Figure 3 A flowchart of the network resilience assessment method;

[0081] Figure 4 Flowchart of strategy optimization and adjustment method. DETAILED DESCRIPTION

[0082] In order to make the purpose, technical solution and advantages of the invention more clear, the invention is further described in detail below in combination with the embodiments. It should be understood that the specific embodiments described here are only used to explain the invention and are not used to limit the invention.

[0083] like Figure 1 As shown in the figure, the online assessment method of regional energy internet resilience based on model-data hybrid drive includes the following steps:

[0084] S1. Build a basic model and develop a physical model of an integrated energy network. The physical model integrates key parameters such as energy flow, capacity and failure probability, and simulates energy flow and conversion through system dynamics principles. Based on the physical model, establish a set of network resilience assessment indicators, which will be defined according to the characteristics and behaviors of the physical model to quantitatively evaluate the resilience of the network in the face of various disturbances.

[0085] S2. Real-time monitoring and risk warning, by deploying smart sensors to collect energy network operation data, and using machine learning technology to predict potential risks and failure points, and quickly respond to abnormal situations;

[0086] S3. Network resilience assessment, based on real-time monitoring and risk warning, uses deep learning algorithms to combine the physical model output of the integrated energy network and the real-time data of the current energy network collected by the real-time monitoring system in S2 to evaluate network resilience; the physical model provides a theoretical description and behavior prediction of the energy network, while the real-time data provides a direct observation of the current state of the network. The hybrid method can more accurately simulate and evaluate the network's response capability and recovery speed in the face of different fault conditions;

[0087] S4. Strategy optimization and adjustment: Based on the resilience assessment results of S3, use optimization algorithms and reinforcement learning methods to formulate and adjust optimization strategies to improve the resilience of the energy network and ensure that the network can recover quickly in the face of various disturbances.

[0088] The physical model of the integrated energy network includes energy node and energy line models;

[0089] The method for constructing the energy node model is as follows: for each node in the energy network (such as a power station, a substation, an energy storage facility, an energy consumption node, etc.), its input energy flow vector, output energy flow vector and stored energy amount are defined. According to the law of conservation of energy, the energy dynamic balance equation of the node is:

[0090]

[0091] S n (t+Δt) The amount of energy stored in node n at time t+Δt; In,i (t) is the i-th input energy flow of node n at time t; n,j (t) is the jth output energy flow of node n at time t; S n (t) The amount of energy stored in node n at time t; Δt is the time step; η out is the conversion efficiency of the output energy flow; η in is the conversion efficiency of the input energy flow;

[0092] The formula represents the dynamic balance of node energy. The amount of energy stored in a node at time t+Δt is equal to the amount of energy stored in the node at time t plus the total amount of input energy flow adjusted by conversion efficiency, minus the total amount of output energy flow adjusted by conversion efficiency.

[0093] The method for constructing the energy line model is as follows: for energy lines (such as power transmission lines, gas pipelines, etc.) connecting nodes, their transmission capacity, transmission loss coefficient and failure probability are defined, and the energy transmission amount on the line satisfies the following relationship:

[0094]

[0095] T i (t) is the energy transmission amount of line l at time t; Q m,j (t) is the i-th output energy flow of the line starting node m at time t; M s is the node set at the starting end of the line; C l is the transmission capacity; α l is the transmission loss coefficient;

[0096] The formula represents the energy transmission amount on the line. The energy transmission amount of the line at time t is equal to (1-α l ) multiplied by the line transmission capacity C l The sum of the energy flows output by the node at the start of the line The smaller value in ; this means that the energy transmission amount is limited by the transmission capacity of the line and the total output energy flow of the starting node, and the transmission loss is taken into account.

[0097] The network resilience evaluation index system includes resilience quantitative index and fault impact evaluation index.

[0098] The resilience quantitative index is defined as a comprehensive measure of the network's ability to maintain normal energy supply and its ability to recover after disturbances within a given time interval. In addition to considering the network's ability to maintain normal energy supply and its ability to recover after disturbances, the resilience quantitative index also takes into account the quality factor of energy supply. The specific calculation formula is:

[0099]

[0100] Among them, t re The time required for the network to recover from a fault state to a normal operating state; E n,def (t) is the energy shortage of node n at time t; E n,nom (t) is the normal energy demand of node n; Q elec (t) is the power quality indicator;

[0101] Fault impact assessment indicators include fault impact range indicators and fault severity indicators:

[0102] The fault impact range index evaluation index is used to measure the ratio of the number of nodes affected after a fault occurs to the total number of nodes, namely:

[0103]

[0104] Among them, FIR represents the fault impact range index; N affected is the number of affected nodes; N total is the total number of nodes in the network;

[0105] The fault severity index is calculated based on the ratio of the total energy shortage caused by the fault to the total normal energy supply of the network:

[0106]

[0107] Among them, t fail The time when the fault occurred.

[0108] In S2, the method of real-time monitoring and risk warning includes:

[0109] S21. Intelligent sensor deployment and data collection, using graph theory-based methods to determine the optimal deployment location of sensors, constructing the topology of the energy network G = (V, E), where V is the node set and E is the edge set; defining the node importance index I v , comprehensively considering factors such as node connectivity, energy flow and impact on system stability; the node importance index I v It is expressed as:

[0110]

[0111] k v The connectivity of the node; F in,v is the energy inflow; F out,v Energy outflow; ω 1 and ω 2 is the weight coefficient;

[0112] Then, the nodes are sorted according to the importance index, and sensors are deployed preferentially on important nodes and their adjacent edges to ensure that network operation data can be collected comprehensively and accurately;

[0113] S22. Adjust the frequency of data collection. According to the dynamic characteristics of the energy network and the historical data of fault occurrence, the frequency of data collection is increased for areas with large fluctuations in energy flow or frequent faults; for relatively stable areas, the frequency is appropriately reduced;

[0114] S23. The machine learning risk prediction model is constructed, combining support vector machine (SVM) and naive Bayes (NB) for risk prediction; for normal operation data, SVM is used for classification and anomaly detection; for fault data with certain probability distribution characteristics, NB is used for prediction; through a decision fusion module, the prediction results of the two are combined according to the characteristics of the data and the confidence of the model, and the formula is expressed as:

[0115]

[0116] Among them, y S is the prediction result of SVM; y N is the prediction result of NB; C S and C N They are the confidence of the prediction results of SVM and NB models respectively; ypred is the final prediction result.

[0117] The method for evaluating network resilience in S3 includes:

[0118] S31. Convolutional neural network (CNN)-recurrent neural network (RNN) fusion model, using CNN to extract the spatial features of the energy network physical model output and real-time data, and RNN to capture time series features. For input data (such as energy flow matrix, node state sequence, etc.), first pass through the CNN layer for convolution operation to obtain a feature map; then, the feature map is flattened and input into the RNN layer (such as LSTM or gated recurrent unit GRU) for time series processing, then:

[0119] h RNN (t) = RNN(x CNN (t),h RNN (t-1));

[0120] Among them, x CNN (t) is the output of the CNN layer; h RNN (t) is the hidden state of the RNN layer;

[0121] Finally, the RNN output is converted into the network resilience evaluation result through the fully connected layer;

[0122] S32. Adversarial training enhancement, introduces the idea of ​​generative adversarial network (GAN), constructs a discriminator and a generator, where the generator is used to generate false data similar to the fault situation, and the discriminator is used to distinguish the real data from the data generated by the generator; during the training process, by continuously optimizing the generator and the discriminator, the evaluation model can better cope with complex fault situations. The loss function of the discriminator is:

[0123]

[0124] The loss function of the generator is:

[0125]

[0126] Train the model by minimizing the loss function of the discriminator and the generator;

[0127] Among them, L D is the loss function of the discriminator; x is the input data; z is a random noise vector; D(x) is the output of the discriminator, indicating the probability that the input data x is the real data; G(z) is the output of the generator; p d (x) is the probability distribution of the real data x; p z (z) is the probability distribution of noise z; E represents expectation;

[0128] S33. Bayesian reasoning is used to determine the weights of the physical model and real-time data in the evaluation. According to Bayesian theorem, the posterior distribution is:

[0129] p(RI model ,RI data |RI final )∝p(RI final |RI model ,RI data )p(RI model )p(RI data );

[0130] Among them, RI data is the evaluation result based on real-time data; RI model is the evaluation result of the physical model; RI final is the final resilience evaluation result; RI model and RI data Obey Gaussian distribution; prior distribution is p(RI model )~N(μ model ,σ 2 model ) and p(RIdata)~N(μ data ,σ 2 data );

[0131] By approximately calculating the posterior distribution, we can obtain the weights ω of the physical model and real-time data. model and ω data , satisfying ω model +ω data =1, and RI final =ω model RI model +ω data RI data ;

[0132] Among them, RI model Represents the evaluation result based on the physical model; RI data express

[0133] Evaluation results based on real-time data; RI final Represents the final toughness assessment result.

[0134] The steps of strategy optimization and adjustment in S4 include:

[0135] S41. Optimization algorithm, introduces a hybrid optimization algorithm that combines simulated annealing algorithm with genetic algorithm; in the iterative process of genetic algorithm, the Metropolis criterion of simulated annealing algorithm is used to accept inferior solutions with a certain probability, increase the diversity of the population, and avoid falling into the local optimum; for each generation of new individuals x generated by genetic algorithm new , calculate its fitness f(x new ), and the current optimal individual x best The fitness comparison f(x best ),if (where T is the simulated annealing temperature parameter), then accept x new As the new current optimal individual; at the same time, considering the impact of energy market price fluctuations on the optimization strategy, the energy cost term is added to the optimization objective function, and the energy production, storage and transmission strategies are dynamically adjusted according to the real-time energy market price. The objective function becomes:

[0136]

[0137] Among them, RI is the toughness index; C total is the total cost; P market is the energy market price at time t; E pu (t) is the amount of energy purchased; α, β, γ are weight coefficients;

[0138] S42. Reinforcement learning environment and agent design;

[0139] S421. Environmental modeling: The operating status, fault conditions and resilience assessment results of the energy network are used as the environmental state of reinforcement learning. The state space of the environmental state is defined as:

[0140] S={s 1 ,s 2 ,…,s n};

[0141] s i Including node energy storage, line transmission capacity utilization, current fault information and network resilience index;

[0142] Add market environment information to the state of the reinforcement learning environment, such as energy price trends, energy policy changes, etc.

[0143] Define the action space as a set of possible optimization strategies, such as adjusting energy production plans, changing energy storage strategies, optimizing line transmission allocation, etc.

[0144] S422. Agent design, using the deep Q network (DQN) or deep deterministic policy gradient (DDPG) algorithm to build an agent. The agent selects actions based on the state of the environment and learns the optimal strategy through interaction with the environment. In DQN, the Q value function of the agent is:

[0145]

[0146] Q(s,a;θ) is the state-action value function; it represents the expected cumulative reward of taking action a in a given state S, where θ is the parameter of the function; E r,τ is the expectation of reward r and discount factor τ; r represents the reward obtained immediately after the agent takes action a in the current state S; the discount factor τ is used to measure the importance of future rewards relative to current rewards; s′ is the next state; a′ is the next action; θ′ is the neural network parameter of the next Q function; It is to take the maximum value of all possible next actions a′;

[0147] Learn effective strategies to improve the resilience of energy networks. In DDPG, the agent consists of an actor network (Actor) and a critic network (Critic). The actor network outputs actions based on the state, and the critic network evaluates the value of the action. The strategy is optimized through the collaborative training of the two.

[0148] The above description is a preferred embodiment of the invention and is not intended to limit the invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the invention should be included in the protection scope of the invention.

Claims

1. An online assessment method for regional energy internet resilience based on model-data hybrid drive, characterized in that: The following steps are involved: S1. Build a basic model and develop a physical model of an integrated energy network. The physical model integrates key parameters of energy flow, capacity and failure probability, and simulates energy flow and conversion through system dynamics principles. Based on the physical model, establish a set of network resilience assessment indicators, which will be defined according to the characteristics and behaviors of the physical model to quantitatively evaluate the resilience of the network in the face of various disturbances. S2. Real-time monitoring and risk warning, by deploying smart sensors to collect energy network operation data, and using machine learning technology to predict potential risks and failure points, and quickly respond to abnormal situations; S3. Network resilience assessment, based on real-time monitoring and risk warning, uses deep learning algorithms to combine the physical model output of the integrated energy network and the real-time data of the current energy network collected by the real-time monitoring system in S2 to evaluate network resilience; the physical model provides a theoretical description and behavior prediction of the energy network, while the real-time data provides a direct observation of the current state of the network. The hybrid method can more accurately simulate and evaluate the network's response capability and recovery speed in the face of different fault conditions; S4. Strategy optimization and adjustment: Based on the resilience assessment results of S3, use optimization algorithms and reinforcement learning methods to formulate and adjust optimization strategies.

2. The online assessment method for regional energy internet resilience based on model-data hybrid drive according to claim 1 is characterized in that: The physical model of the integrated energy network includes energy node and energy line models; The method for constructing the energy node model is as follows: for each node in the energy network, its input energy flow vector, output energy flow vector and stored energy amount are defined. According to the law of conservation of energy, the energy dynamic balance equation of the node is: S n (t+Δt) The amount of energy stored in node n at time t+Δt; I n,i (t) is the i-th input energy flow of node n at time t; n,j (t) is the jth output energy flow of node n at time t; S n (t) The amount of energy stored in node n at time t; Δt is the time step; η out is the conversion efficiency of the output energy flow; η in is the conversion efficiency of the input energy flow; The formula represents the dynamic balance of node energy. The amount of energy stored in the node at time t+Δt is equal to the amount of energy stored in the node at time t plus the total amount of input energy flow adjusted by conversion efficiency, minus the total amount of output energy flow adjusted by conversion efficiency.

3. The online assessment method for regional energy internet resilience based on model-data hybrid drive according to claim 2 is characterized in that: The energy line model construction method is as follows: for the energy line connecting the nodes, its transmission capacity, transmission loss coefficient and failure probability are defined, and the energy transmission amount on the line satisfies the following relationship: T i (t) is the energy transmission amount of line l at time t; Q m,j (t) is the i-th output energy flow of the line starting node m at time t; M s is the node set at the starting end of the line; C l is the transmission capacity; α l is the transmission loss coefficient; The formula represents the energy transmission amount on the line. The energy transmission amount of the line at time t is equal to (1-α l ) multiplied by the line transmission capacity C l The sum of the energy flows output by the node at the start of the line The smaller value in .

4. The online assessment method for regional energy internet resilience based on model-data hybrid drive according to claim 1 is characterized in that: The network resilience evaluation index system includes resilience quantitative index and fault impact evaluation index. The resilience quantitative index is defined as a comprehensive measure of the network's ability to maintain normal energy supply and its ability to recover after disturbances within a given time interval. In addition to considering the network's ability to maintain normal energy supply and its ability to recover after disturbances, the resilience quantitative index also takes into account the quality factor of energy supply. The specific calculation formula is: Among them, t re The time required for the network to recover from a fault state to a normal operating state; E n,def (t) is the energy shortage of node n at time t; E n,nom (t) is the normal energy demand of node n; Q elec (t) is the power quality indicator.

5. The online assessment method for regional energy internet resilience based on model-data hybrid drive according to claim 4 is characterized in that: Fault impact assessment indicators include fault impact range indicators and fault severity indicators: The fault impact range index evaluation index is used to measure the ratio of the number of nodes affected after a fault occurs to the total number of nodes, namely: Among them, FIR represents the fault impact range index; N affected is the number of affected nodes; N total is the total number of nodes in the network; The fault severity index is calculated based on the ratio of the total energy shortage caused by the fault to the total normal energy supply of the network: Among them, t fail The time when the fault occurred.

6. The online assessment method for regional energy internet resilience based on model-data hybrid drive according to claim 1 is characterized in that: In S2, the method of real-time monitoring and risk warning includes: S21. Intelligent sensor deployment and data collection, using graph theory-based methods to determine the optimal deployment location of sensors, constructing the topology of the energy network G = (V, E), where V is the node set and E is the edge set; defining the node importance index I v , comprehensively considering the node connectivity, energy flow and factors affecting system stability; the node importance index I v It is expressed as: k v The connectivity of the node; F in,v is the energy inflow; F out,v Energy outflow; ω1 and ω2 are weight coefficients; Then, the nodes are sorted according to the importance index, and sensors are deployed preferentially on important nodes and their adjacent edges to ensure that network operation data can be collected comprehensively and accurately; S22. Adjust the frequency of data collection. According to the dynamic characteristics of the energy network and the historical data of fault occurrence, the frequency of data collection is increased for areas with large fluctuations in energy flow or frequent faults; for relatively stable areas, the frequency is appropriately reduced; S33. The machine learning risk prediction model is constructed, combining support vector machine (SVM) and naive Bayes (NB) for risk prediction; for normal operation data, SVM is used for classification and anomaly detection; for fault data with certain probability distribution characteristics, NB is used for prediction; through a decision fusion module, the prediction results of the two are combined according to the characteristics of the data and the confidence of the model, and the formula is expressed as: Among them, y S is the prediction result of SVM; y N is the prediction result of NB; C S and C N are the confidence of the prediction results of SVM and NB models respectively; y pred Final prediction results.

7. The online assessment method for regional energy internet resilience based on model-data hybrid drive according to claim 1 is characterized in that: The method for evaluating network resilience in S3 includes: S31. Convolutional neural network (CNN)-recurrent neural network (RNN) fusion model, using CNN to extract the spatial features of the energy network physical model output and real-time data, RNN to capture time series features. For the input data, firstly, convolution operation is performed through the CNN layer to obtain the feature map; then, the feature map is flattened and input into the RNN layer (such as LSTM or gated recurrent unit GRU) for time series processing, then: h RNN (t)=RNN(x CNN (t),h RNN (t-1)); Among them, x CNN (t) is the output of the CNN layer; h RNN (t) is the hidden state of the RNN layer; Finally, the RNN output is converted into the network resilience evaluation result through the fully connected layer; S32. Adversarial training enhancement, introduces the idea of ​​generative adversarial network (GAN), constructs a discriminator and a generator, where the generator is used to generate false data similar to the fault situation, and the discriminator is used to distinguish the real data from the data generated by the generator; during the training process, by continuously optimizing the generator and the discriminator, the evaluation model can better cope with complex fault situations. The loss function of the discriminator is: The loss function of the generator is: Train the model by minimizing the loss function of the discriminator and the generator; Among them, L D is the loss function of the discriminator; x is the input data; z is a random noise vector; D(x) is the output of the discriminator, indicating the probability that the input data x is the real data; G(z) is the output of the generator; p d (x) is the probability distribution of the real data x; p z (z) is the probability distribution of noise z; E represents expectation; S33. Bayesian reasoning is used to determine the weights of the physical model and real-time data in the evaluation. According to Bayesian theorem, the posterior distribution is: p(RI model ,RI data |RI final )∝p(RI final |RI model ,RI data )p(RI model )p(RI data ); Among them, RI data is the evaluation result based on real-time data; RI model is the evaluation result of the physical model; RI final is the final resilience evaluation result; RI model and RI data Obey Gaussian distribution; prior distribution is p(RI model) ~N(μ model ,σ 2 model ) and p(RI data )~N(μ data ,σ 2 data ); By approximately calculating the posterior distribution, we can obtain the weights ω of the physical model and real-time data. model and ω data , satisfying ω model +ω data =1, and RI final =ω model RI model +ω data RI data ; Among them, RI model Represents the evaluation result based on the physical model; RI data Indicates the evaluation results based on real-time data; RI final Represents the final toughness assessment result.

8. The online assessment method for regional energy internet resilience based on model-data hybrid drive according to claim 1 is characterized in that: The steps of strategy optimization and adjustment in S4 include: S41. Optimization algorithm, introduces a hybrid optimization algorithm that combines simulated annealing algorithm with genetic algorithm; in the iterative process of genetic algorithm, the Metropolis criterion of simulated annealing algorithm is used to accept inferior solutions with a certain probability, increase the diversity of the population, and avoid falling into the local optimum; for each generation of new individuals x generated by genetic algorithm new , calculate its fitness f(x new ), and the current optimal individual x best The fitness comparison f(x best ),if Where T is the temperature parameter of simulated annealing, then accept x new As the new current optimal individual; at the same time, considering the impact of energy market price fluctuations on the optimization strategy, the energy cost term is added to the optimization objective function, and the energy production, storage and transmission strategies are dynamically adjusted according to the real-time energy market price. The objective function becomes: Among them, RI is the toughness index; C total is the total cost; P marke t is the energy market price at time t; E pu (t) is the amount of energy purchased; α, β, γ are weight coefficients; S42. Reinforcement learning environment and agent design; S421. Environmental modeling: The operating status, fault conditions and resilience assessment results of the energy network are used as the environmental state of reinforcement learning. The state space of the environmental state is defined as: S={s1,s2,…,s n }; s i Including node energy storage, line transmission capacity utilization, current fault information and network resilience index; Add market environment information to the state of the reinforcement learning environment, such as energy price trends and energy policy changes; Define the action space as a set of possible optimization strategies, such as adjusting energy production plans, changing energy storage strategies, and optimizing line transmission allocation; S422. Agent design, using the deep Q network (DQN) or deep deterministic policy gradient (DDPG) algorithm to build an agent. The agent selects actions based on the state of the environment and learns the optimal strategy through interaction with the environment. In DQN, the Q value function of the agent is: Q(s,a;θ) is the state-action value function; it represents the expected cumulative reward of taking action a in a given state S, where θ is the parameter of the function; E r,τ is the expectation of reward r and discount factor τ; r represents the reward obtained immediately after the agent takes action a in the current state S; the discount factor τ is used to measure the importance of future rewards relative to current rewards; s′ is the next state; a′ is the next action; θ′ is the neural network parameter of the next Q function; It is to take the maximum value of all possible next actions a′; Learn effective strategies to improve the resilience of energy networks. In DDPG, the agent consists of an actor network (Actor) and a critic network (Critic). The actor network outputs actions based on the state, and the critic network evaluates the value of the action. The strategy is optimized through the collaborative training of the two.

Images