Electric power system intelligent perception decision generation and optimization method and system based on behavior reasoning

Abstract

The invention relates to the technical field of electric power system automation, and discloses an electric power system intelligent perception decision generation and optimization method and system based on behavior reasoning, and the method comprises the steps: constructing an equipment behavior feature space through monitoring the operation parameters of equipment in an electric power system in real time; based on the equipment behavior feature space, predicting equipment behaviors by adopting a time sequence reasoning model, and generating a behavior reasoning result; constructing a Markov decision process model in combination with the behavior reasoning result and the environmental parameters, and generating an optimal decision meeting security constraints; and establishing a multi-objective optimization model according to the optimal decision, and solving a Pareto optimal solution set through an improved optimization algorithm to realize operation optimization of the power system. According to the method, intelligent perception, decision generation and optimization of the power system are realized through cooperative work of all layers.

Description

Technical Field

[0001] The present invention relates to the technical field of power system automation, and particularly to a method and system for generating and optimizing intelligent perception decisions of a power system based on behavior reasoning. Background Art

[0002] With the development of the new power system towards high - proportion renewable energy access and multi - energy coupling, traditional simulation analysis methods based on physical models face problems such as lagging dynamic response and insufficient multi - source heterogeneous data processing capabilities. The existing technologies have the following limitations:

[0003] Existing data - driven methods lack the modeling of the causal relationship of equipment operation behavior, resulting in poor decision interpretability; traditional optimization algorithms are difficult to adapt to the dynamic topology changes of the power system and cannot achieve real - time decision generation; the weight allocation mechanism in the multi - objective optimization process is fixed and lacks scenario - oriented adaptability. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method for generating and optimizing intelligent perception decisions of a power system based on behavior reasoning, which can achieve intelligent perception, decision generation and optimization of the power system through the collaborative work among layers.

[0006] To solve the above - mentioned technical problems, the present invention provides the following technical solution. A method for generating and optimizing intelligent perception decisions of a power system based on behavior reasoning includes: constructing an equipment behavior feature space by real - time monitoring the operation parameters of equipment in the power system; predicting the equipment behavior by using a time - series reasoning model based on the equipment behavior feature space to generate a behavior reasoning result; constructing a Markov decision process model by combining the behavior reasoning result and environmental parameters to generate an optimal decision that meets safety constraints; establishing a multi - objective optimization model according to the optimal decision and solving the Pareto optimal solution set through an improved optimization algorithm to realize the operation optimization of the power system.

[0007] As a preferred solution of the method for generating and optimizing intelligent perception decisions of a power system based on behavior reasoning according to the present invention, wherein: the construction of the equipment behavior feature space includes an equipment state matrix, an equipment association adjacency matrix and an equipment reliability index vector.

[0008] The equipment state matrix is generated by collecting the voltage, current and power parameters of all equipment in the power system, and the matrix dimension is determined by the number of equipment and the feature dimension.

[0009] The equipment association adjacency matrix is generated according to the physical connection relationship between equipment, and the matrix elements represent the direct relevance between equipment.

[0010] The device reliability index vector is generated by comprehensively calculating the historical failure data, maintenance records, and inherent characteristics of the device.

[0011] As a preferred solution of the power system intelligent perception decision generation and optimization method based on behavior reasoning according to the present invention, wherein: the time series reasoning model includes using a double-layer long short-term memory network model combined with an attention mechanism to perform time series reasoning on the device behavior characteristics;

[0012] The first layer of the model receives the device behavior feature sequence, and the second layer performs high-level feature extraction on the output of the first layer; the attention mechanism realizes the fusion of local features and global features through the relevance score between the global state encoding and the current hidden state.

[0013] As a preferred solution of the power system intelligent perception decision generation and optimization method based on behavior reasoning according to the present invention, wherein: the prediction of the device behavior includes training a double-layer LSTM time series reasoning model, and using stochastic gradient descent (SGD) or an improved algorithm to update the parameters of the model. The parameters include the weight matrices W f 、W i 、W c 、W o and the bias vectors b f 、b i 、b c 、b o in the LSTM unit, as well as the parameters in the Attention mechanism;

[0014] During the training process, the historical data of the device behavior is divided into a training set and a validation set according to the time series. The model is optimized by minimizing the loss function. At the same time, a regularization method and Dropout technology are used to prevent overfitting. The trained model is deployed to the corresponding computing platform to receive device behavior data in real time, and behavior reasoning is performed through the double-layer LSTM time series reasoning model and the Attention mechanism to provide a decision basis.

[0015] As a preferred solution of the power system intelligent perception decision generation and optimization method based on behavior reasoning according to the present invention, wherein: the construction of the Markov decision process model includes the device behavior feature space and environmental parameters, and the action space covers all executable operations of the devices;

[0016] The hierarchical reinforcement learning architecture includes a policy network, a value network, and a constraint network, and the decision is made to satisfy the safety constraints through an adaptive penalty coefficient.

[0017] As a preferred solution of the power system intelligent perception decision generation and optimization method based on behavior reasoning according to the present invention, wherein: the establishment of the multi-objective optimization model includes the operating cost, risk coefficient, and carbon emissions, which are expressed as:

[0018] min[f1(x), f2(x), f3(x)] = [Operating cost, Risk coefficient, Carbon emission]

[0019] For a system with two or more devices, the operating cost is expressed as the sum of the operating costs of each device. Suppose there are n devices in the system, and the operating cost function of each device i is c i (x), then the operating cost where x is the decision variable vector;

[0020] The risk coefficient is represented by the product of the failure probability and the risk loss, that is where p i (x) is the failure probability of device i under the decision variable x, and l i (x) is the loss when device i fails;

[0021] The carbon emission comes from the energy combustion process of the power generation equipment. Suppose the carbon emission coefficient of power generation equipment i is e i , and its power generation power is p i (x), then the carbon emission

[0022] As a preferred solution of the intelligent perception decision generation and optimization method for power systems based on behavior reasoning described in the present invention, wherein: the solution of the Pareto optimal solution set includes enhancing the diversity of the optimization solution set by introducing a behavior similarity constraint, and the formula of the behavior similarity constraint is

[0023]

[0024] where y i and y j are two individuals in the population, H(y) is the behavior feature vector of individual y, which is obtained by mapping through the device behavior feature space obtained in the behavior perception modeling stage, ||·|| represents the Euclidean norm, and ∈ is a preset parameter used to control the attenuation speed of similarity.

[0025] As a preferred solution of the intelligent perception decision generation and optimization system for power systems based on behavior reasoning described in the present invention, wherein: it includes a data perception and feature modeling module, a behavior reasoning and prediction module, a dynamic decision generation module, and a multi-objective optimization module;

[0026] The data perception and feature modeling module provides basic data support for the system and forms a quantifiable behavior feature space through multi-dimensional feature mapping and device relationship modeling;

[0027] The behavior reasoning and prediction module mines the behavior patterns of devices from the time series dimension, combines the attention mechanism to enhance the ability to extract key information, and provides accurate reasoning basis for decision-making generation;

[0028] The dynamic decision-making generation module balances system stability, economy, and risk in a complex dynamic environment, and ensures the real-time and safety of decisions through a hierarchical reinforcement learning architecture;

[0029] The multi-objective optimization module synthesizes the conflicting requirements of multiple objectives and outputs an optimization solution set that can be flexibly selected.

[0030] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the intelligent perception decision-making generation and optimization method for power systems based on behavior reasoning.

[0031] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the steps of the intelligent perception decision-making generation and optimization method for power systems based on behavior reasoning.

[0032] Advantages of the present invention: When a fault occurs, this method can quickly locate the problem and take effective measures, greatly improving the reliability of the power system, reducing the power outage time and economic losses caused by faults. It effectively promotes the application of high-proportion renewable energy in the power system, promotes the green and low-carbon transformation of the power system, reduces the dependence on traditional fossil energy, reduces carbon emissions, and has significant environmental benefits.

[0033] When the operating state of the power system changes rapidly, it makes timely responses and generates decisions. Combining the Markov decision process and the hierarchical reinforcement learning architecture, comprehensively considering system states, actions, and environmental factors, it generates decisions that not only meet safety constraints but also maximize long-term cumulative rewards, improving the scientificity and effectiveness of decisions and ensuring the stable and efficient operation of the power system.

[0034] By adopting an improved NSGA III algorithm and introducing behavior similarity constraints, more Pareto optimal solutions can be found, providing richer choices for decision-makers. The "behavior fingerprint" modeling technology is pioneered to realize the digital mapping of device operation characteristics, enrich the expression of behavior characteristics, and provide comprehensive data support for subsequent analysis. The dynamic causal reasoning mechanism designs a causal effect evaluation function to enhance the reliability of behavior reasoning and the scientificity of decision-making basis. The hierarchical reinforcement learning architecture constructs a three-level network to generate optimal strategies under safety constraints, adapt to the dynamic topological changes of the power system, break through the limitations of traditional methods in terms of dynamic adaptability and interpretability, and lay a core technical foundation for building a new generation of power intelligent decision-making systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0036] Figure 1 Schematic diagram of the architecture of the intelligent perception decision-making generation and optimization method for power systems based on behavior reasoning provided by an embodiment of the present invention. Detailed implementation manners

[0037] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the detailed implementation manners of the present invention with reference to the accompanying drawings of the specification. Obviously, the described embodiments are some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0038] Embodiment 1, referring to Figure 1 , which is the first embodiment of the present invention. This embodiment provides an intelligent perception decision-making generation and optimization method for power systems based on behavior reasoning, including:

[0039] S1: By real-time monitoring the operating parameters of devices in the power system, construct a device behavior feature space.

[0040] Furthermore, to accurately describe the behavior characteristics of devices in the power system, establish a device behavior feature space Ω=(X, A, R). In this feature space, it includes a device state matrix X, a device association adjacency matrix A, and a device reliability index vector R:

[0041] Device state matrix X: X∈R n×d , where n represents the number of devices in the power system, and d represents the selected feature dimensions of each device. For example, for a power system containing various devices such as generators, transformers, and transmission lines, the number of devices n is the sum of these devices; the feature dimension d can cover operating parameters such as the voltage, current, and power of the device, and through these parameters, the operating state of the device at a certain moment can be comprehensively reflected.

[0042] Device association adjacency matrix A: A∈{0, 1} n×n , this matrix is used to represent the connection relationship between devices. When the element A ij = 1 in the matrix, it means that there is a direct association between device i and device j; when A ijWhen it is 0, it means there is no direct connection between device i and device j. Taking the power transmission network as an example, if the transformer at node i is connected to the power transmission line at node j, then A ij = 1, otherwise it is 0.

[0043] Device reliability index vector R: R ∈ R n , and each element R i in the vector corresponds to the reliability index of a device. This index is a quantitative parameter for measuring the normal operation ability of a device within a certain period of time, and can be comprehensively calculated through various factors such as the historical fault data, maintenance records, and inherent characteristics of the device. For example, indicators such as failure probability and mean time between failures are used to represent the reliability of the device.

[0044] On this basis, through the "behavior fingerprint" modeling technology, the device behavior entropy feature quantity BE is introduced i Formula:

[0045]

[0046] where P k is the transition probability of the device in the state, and K is the number of operation state types. This formula realizes the digital mapping of the device operation characteristics, quantifies the device behavior characteristics from a new dimension, enriches the expression form of the behavior characteristics, and complements the device behavior characteristic space Ω, providing more comprehensive data support for the subsequent in-depth analysis of the device behavior. Among them, the calculation method of P k is as follows:

[0047] It should be noted that when calculating the state transition probability P k , the historical operation data of the device can be used for statistical analysis.

[0048] Taking a certain transformer in the IEEE 39 - bus system as an example, collect its operation - state data over a long period of time. These states can be divided into several discrete states according to key parameters such as voltage, current, and power. Suppose the operation states of the transformer are divided into \(K = 3\) states: normal, slightly abnormal, and severely abnormal. Count the number of times that when the transformer is in state \(i\) at a certain moment, it transfers to state \(k\) (\(k = 1,2,3\) corresponding to the above three states) at the next moment. For example, among the 1000 time steps counted, it is found that when the transformer is in the normal state, there are 800 times that it remains in the normal state at the next moment, 150 times that it changes to the slightly abnormal state, and 50 times that it changes to the severely abnormal state. Then the probability \(P1\) of transferring from the normal state to the normal state is \(800\div1000 = 0.8\); the probability \(P2\) of transferring to the slightly abnormal state is \(150\div1000 = 0.15\); the probability \(P3\) of transferring to the severely abnormal state is \(50\div1000 = 0.05\). By such statistical calculations of a large amount of historical data, the transfer probabilities \(P\) between different states of different devices can be determined. k 。

[0049] Furthermore, in order to better extract and characterize device behavior information from the device behavior feature space, the present invention uses an improved graph convolutional network for behavior characterization. Its specific formula is:

[0050]

[0051] where: \(H\) (l) represents the feature matrix of the graph convolutional network at the \(l\) - th layer, which contains the extraction results of the current layer for device behavior features. Through continuous updating, it gradually learns more advanced and more abstract device behavior features. \(\sigma\) is the activation function.

[0052] is an improvement on the original adjacency matrix \(A\). This improvement is to avoid information loss caused by overly sparse connection relationships of some nodes during the calculation process. By processing the adjacency matrix in this way, the connection information between devices can be better retained, and the learning ability of the model for device relationships can be improved.

[0053] is the degree matrix, which is a diagonal matrix, and the elements on its diagonal are equal to the sum of the elements in the \(i\) - th row of

[0054] \(W\) (l)is the trainable parameter matrix of the l-th layer. During the model training process, the values of these parameters are continuously adjusted through the backpropagation algorithm, enabling the model to better fit the data and learn the internal relationships between device behavior characteristics.

[0055] During the actual training process, by continuously inputting data such as the device state matrix X, the device association adjacency matrix A, and the device reliability index vector R, iterative calculations are performed using the above improved graph convolution network formula, and the feature matrix H is gradually updated, thereby effectively representing the device behavior and providing accurate feature information for subsequent behavior reasoning, decision generation, and optimization.

[0056] S2: Based on the device behavior feature space, a temporal reasoning model is used to predict the device behavior and generate a behavior reasoning result.

[0057] Furthermore, after perceiving and modeling the device behavior, the present invention designs a two-layer LSTM (Long Short-Term Memory) temporal reasoning model. At the same time, the Attention mechanism is introduced to achieve the fusion of local features and global features, further enhancing the model's ability to capture complex behavior patterns. In addition, a causal effect evaluation function CE is designed:

[0058] CE = τ ATE + λ·Gov(τ(x), x)

[0059] where τ ATE is the average treatment effect, λ is the balance coefficient, realizing the causal inference of decision-making impacts, working in coordination with the two-layer LSTM and the Attention mechanism, and enhancing the reliability of behavior reasoning and the scientific nature of decision-making basis.

[0060] The output of a single-layer LSTM unit can be expressed as h t = LSTM(x t , h t-1 , c t-1 ), and the current hidden state h t and cell state c t-1 are calculated based on the current input x t-1 , the hidden state h t of the previous moment, and the cell state c t

[0061] The present invention adopts a two-layer LSTM structure. The first layer of LSTM receives the device behavior feature sequence output in the behavior perception modeling stage as input, and its output is used as the input of the second layer of LSTM. The second layer of LSTM further processes the output of the first layer to obtain a higher-level feature representation With this double - layer structure, the model can learn more complex time - series patterns and improve its reasoning ability about device behavior.

[0062] It should be noted that based on the double - layer LSTM, the Attention mechanism is introduced to achieve the fusion of local features and global features:

[0063]

[0064] Among them, H global is the global state encoding, which can be a comprehensive representation of the entire device behavior sequence, for example, obtained by average pooling or max pooling of the hidden states at all times; represents the i - th element in H global ; the score function is used to calculate the correlation score between the hidden state h t at the current time and each element in the global state encoding. Common score functions include dot - product attention, scaled dot - product attention, etc. Here, dot - product attention is taken as an example:

[0065]

[0066] The correlation scores are transformed into a probability distribution through the Softmax function:

[0067]

[0068] Among them, N is the length of the global state encoding H global . α t,i represents the attention weight of the current time t to the i - th element in the global state encoding.

[0069] Finally, the fused feature vector s t is obtained through weighted summation:

[0070]

[0071] Furthermore, when training the double - layer LSTM time - series reasoning model, Stochastic Gradient Descent (SGD) or its improved algorithms (such as Adam, Adagrad, etc.) are used to update the model parameters, including the weight matrices W f , W i , W c , W o and the bias vectors b f , b i , b c , b o in the LSTM cells, as well as the relevant parameters in the Attention mechanism.

[0072] During the training process, the historical data of device behavior is divided into a training set and a validation set according to the time series. The model is optimized by minimizing the loss function (such as the mean square error loss function where y t is the true value, is the model prediction value, and T is the length of the time series). At the same time, to prevent overfitting, regularization methods (such as L1 and L2 regularization) and Dropout technology can be used.

[0073] In practical applications, the trained model is deployed to the corresponding computing platform to receive device behavior data in real time. Behavioral reasoning is performed through a two-layer LSTM time series inference model and an Attention mechanism, providing a basis for subsequent decision-making.

[0074] S3: Combine the behavioral reasoning results and environmental parameters to construct a Markov decision process model and generate an optimal decision that satisfies safety constraints.

[0075] Furthermore, to generate optimal decisions for device management in a complex and dynamically changing environment, a Markov Decision Process (MDP) model is constructed. At the same time, the idea of a hierarchical reinforcement learning architecture is incorporated to construct a three-level network of "policy network - value network - constraint network":

[0076]

[0077] where π θ (a|s) is the action probability distribution output by the policy network, η is the adaptive penalty coefficient, is the expected cumulative reward evaluated by the value network, is the degree of constraint violation output by the constraint network, s is the state, and a is the action. This network architecture is used to generate an optimal policy under safety constraints. Combined with the MDP model, it further optimizes the decision-making process to ensure that the decision not only meets the system safety requirements but also maximizes the long-term cumulative reward.

[0078] It should be noted that to measure the pros and cons of taking a certain action in a certain state, a composite reward function R is designed, and its expression is:

[0079] R = αR stability + βR economy + γR risk

[0080] Stability reward R stability, which is used to evaluate the impact of an action on system stability. It can be defined by calculating factors such as the degree of fluctuation of the device state and the voltage stability index of the system. For example, if an action can keep the voltage fluctuation of the device within a small range and maintain the voltage stability of the system, a higher stability reward is given.

[0081] A possible definition method is where x j,t represents a key state parameter (such as voltage, power, etc.) of device j at time t, and ω j is the weight coefficient of device j, which reflects the importance of this device in the system.

[0082] Economic reward R economy : Mainly consider the economic cost and benefit of the action. For example, for a power generation device, if the action taken can reduce the power generation cost and improve the energy utilization efficiency, a higher economic reward is given. It can be defined as R economy = benefit - cost, where the benefit can be the sales revenue of the electric energy generated by the device, and the cost can include fuel cost, equipment maintenance cost, etc.

[0083] Risk reward R risk : Used to evaluate the risks that an action may bring. Risks can include equipment failure risk, system collapse risk, etc. For example, if an action increases the failure probability of the device, a lower risk reward is given. It can be defined by calculating the failure probability of the device, risk assessment indicators, etc.;

[0084] Such as where p i is the failure probability of device i.

[0085] The discount factor γ ∈ [0, 1] is used to balance the importance of current rewards and future rewards. A larger γ value indicates more emphasis on future long-term rewards, while a smaller γ value pays more attention to current immediate rewards. In practical applications, the appropriate discount factor needs to be selected according to specific problems and goals. For example, for some long-term planning decision-making problems, a larger γ value can be selected; while for some short-term emergency decision-making problems, a smaller γ value can be selected.

[0086] Furthermore, the coefficients (α, β, γ) are used to adjust the weights of each part in the composite reward function, and they are dynamically adjusted through fuzzy inference.

[0087] For each input variable, multiple fuzzy sets are defined. For example, for the load level, three fuzzy sets of "low load", "medium load", and "high load" can be defined, and the corresponding membership functions are defined for each fuzzy set. The membership function is used to describe the degree to which the input variable belongs to a certain fuzzy set. Common membership functions include triangular membership functions, trapezoidal membership functions, etc.

[0088] According to expert knowledge and experience, a series of fuzzy rules are formulated. For example, Rule 1: If the load level is "low load" and the average reliability of the equipment is "high", then α takes a smaller value, β takes a larger value, and γ takes a smaller value; Rule 2: If the load level is "high load" and the average reliability of the equipment is "low", then α takes a larger value, β takes a smaller value, and γ takes a larger value.

[0089] When dynamically adjusting the coefficients (α, β, γ) of the composite reward function using fuzzy inference, the system load level and the average reliability of the equipment are used as input variables. For the system load level, triangular membership functions are used to define the three fuzzy sets of "low load", "medium load", and "high load". Assume that the load range of the system is [0, P max , when the load is less than 0.3P max the membership function of "low load" is:

[0090]

[0091] When the load L is between 0.3P max and 0.7P max the membership function of "medium load" is:

[0092]

[0093] When the load L is greater than 0.7P max the membership function of "high load" is:

[0094]

[0095] For the average reliability of the equipment, triangular membership functions are also used to define the three fuzzy sets of "low reliability", "medium reliability", and "high reliability". Assume that the value range of the average reliability of the equipment is [0, 1], when the average reliability R avg is less than 0.3, the membership function of "low reliability" is:

[0096]

[0097] The membership function of "medium reliability" is: When the average reliability R avg is between 0.3 and 0.7,

[0098]

[0099] When the average reliability R avg is greater than 0.7, the membership function of "high reliability" is:

[0100]

[0101] According to these membership functions, the membership degrees of the system belonging to each fuzzy set can be calculated based on the current load level of the system and the average reliability of the equipment. Then, combined with the pre-established fuzzy rules, such as "if the load level is 'low load' and the average reliability of the equipment is 'high', then α takes a smaller value, β takes a larger value, and γ takes a smaller value", fuzzy reasoning and defuzzification operations are carried out to dynamically adjust the coefficients (α, β, γ) in the composite reward function.

[0102] Furthermore, after constructing the Markov decision process and determining the reward function and the coefficient adjustment method, the optimal policy π is solved * to generate dynamic decisions.

[0103] In an alternative embodiment, the optimal policy π * adopts the value iteration algorithm. Specifically, the value iteration algorithm solves the optimal policy by continuously iteratively updating the state value function V(s). The state value function V(s) represents the expected value of the long-term cumulative reward that can be obtained starting from state s and following the optimal policy. Its iteration formula is:

[0104]

[0105] where k represents the number of iterations, V k+1 (s) is the value function of state s at the (k + 1)-th time, V k (s′) is the value function of state s′ at the k-th iteration, and P is the state transition probability. When the difference between V k+1 (s) and V k (s) is less than a preset threshold, the algorithm is considered to converge. At this time, the optimal policy can be calculated according to the state value function V(s):

[0106]

[0107] In another alternative embodiment, the optimal policy π * adopts the policy iteration algorithm. Specifically, the policy iteration algorithm is divided into two steps: policy evaluation and policy improvement. In the policy evaluation step, the state value function V π (s) under the current policy π is calculated and obtained by solving the system of equations:

[0108]

[0109] In the policy improvement step, the policy is improved according to the current state value function V π (s) to obtain a new policy π', that is:

[0110]

[0111] Repeat these two steps continuously until the policy converges to obtain the optimal policy π * .

[0112] Through the above process, at each time step, according to the current system state s ∈ S, the corresponding action a ∈ A is selected through the optimal policy π * (s), thus realizing the generation of dynamic decision-making. At the same time, the policy output by the policy network in the hierarchical reinforcement learning architecture cooperates with the optimal policy solving process. On the premise of meeting safety constraints, it further optimizes the decision-making selection, improving the scientificity and effectiveness of decision-making.

[0113] S4: Establish a multi-objective optimization model based on the optimal decision-making, and solve the Pareto optimal solution set through an improved optimization algorithm to realize the operation optimization of the power system.

[0114] Furthermore, in the actual system operation, it is often necessary to consider multiple conflicting objectives simultaneously, such as operating cost, risk coefficient, and carbon emission. It is usually impossible to achieve the optimal simultaneously among these objectives. Therefore, multi-objective optimization is required to find a set of compromise optimal solutions, that is, Pareto optimal solutions. This part will elaborate in detail on how to establish a Pareto front search model and use an improved NSGAIII algorithm for solving.

[0115] The established multi-objective optimization problem contains three objective functions, corresponding to operating cost, risk coefficient, and carbon emission respectively, and can be expressed as:

[0116] min[f1(x), f2(x), f3(x)] = [operating cost, risk coefficient, carbon emission]

[0117] Operating cost f1(x): The operating cost usually includes the energy consumption cost, maintenance cost, labor cost, etc. of the equipment. For a system containing multiple devices, the operating cost can be expressed as the sum of the operating costs of each device. Assume there are n devices in the system, and the operating cost function of each device i is c i (x), then the operating cost where x is the decision variable vector, which contains various parameters affecting the system operation, such as the operating power and operating time of the equipment.

[0118] Risk coefficient f2(x): The risk coefficient is used to measure various risks that the system may face during operation, such as equipment failure risk, safety accident risk, etc. It can be obtained through a comprehensive evaluation of the reliability index of the equipment, historical failure data, and the operating state of the system. For example, the risk coefficient can be expressed as the product of the failure probability and the risk loss, that is where p i (x) is the failure probability of device i under the decision variable x, and l iThe loss when device i fails is \(L(x)\).

[0119] Carbon emissions \(f_3(x)\): Carbon emissions are closely related to the type and intensity of energy consumption in the system. Taking the power system as an example, carbon emissions mainly come from the energy combustion process of power generation equipment. Assume that the carbon emission coefficient of power generation equipment i is \(e\) i , and its power generation is \(p_i(x)\), then the carbon emissions i are

[0120] It should be noted that the optimization problem needs to satisfy a series of constraints, which can be expressed as:

[0121] \(g_j(x)\leq0, j = 1,\cdots,m\) j

[0122] These constraints can be divided into equality constraints and inequality constraints. Common constraints include:

[0123] Device operation limits, such as the upper and lower power limits of the device, can be expressed as

[0124]

[0125] where and are the minimum and maximum powers of device i, respectively.

[0126] For the power system, the power balance constraint needs to be satisfied, that is where \(P\) load is the total load of the system.

[0127] Environmental constraints, such as the total carbon emissions not exceeding a certain upper limit, can be expressed as \(f_3(x)\leq E\) max , where \(E\) max is the maximum allowable carbon emissions.

[0128] Furthermore, in order to ensure the behavioral diversity of the optimization solution set, the present invention improves the NSGA-III algorithm and introduces a behavioral similarity constraint. The behavioral similarity is calculated by the following formula:

[0129]

[0130] where \(y_i\) i and \(y_j\) j

[0131] are two individuals in the population, \(H(y)\) is the behavioral feature vector of individual y obtained through the behavioral perception modeling stage, \(\|\cdot\|\) represents the Euclidean norm, and \(\epsilon\) is a preset parameter used to control the attenuation speed of similarity.

[0132] It should be noted that a group of initial individuals is randomly generated as a population, and each individual y is a decision variable vector that needs to satisfy the constraint g j (y) ≤ 0, j = 1, …, m.

[0133] Perform non-dominated sorting on the individuals in the population and divide the individuals into different non-dominated ranks. The lower the non-dominated rank of an individual, the better it is, and there is no domination relationship among individuals within the same non-dominated rank.

[0134] For each individual y i , calculate its behavioral similarity sim(y j , y i , y j ) with other individuals y in the population.

[0135] When calculating the crowding distance, in addition to considering the crowding degree of the objective space, a behavioral similarity constraint is also introduced. For individual y i , its crowding distance d i can be expressed as:

[0136]

[0137] Among them, is the traditional crowding distance in the objective space, λ is a weight coefficient used to adjust the influence degree of behavioral similarity on the crowding distance, and M is the size of the population.

[0138] Perform selection operations according to the non-dominated rank and crowding distance of individuals, and preferentially select individuals with a low non-dominated rank and a large crowding distance to enter the next generation population. Perform crossover and mutation operations on the selected individuals to generate new individuals. The crossover operation can adopt methods such as single-point crossover and multi-point crossover, and the mutation operation can adopt methods such as random mutation and Gaussian mutation. The newly generated individuals need to satisfy the constraint conditions:

[0139] g j (y) ≤ 0, j = 1, …, m

[0140] Repeat the above steps until the termination condition is met, such as reaching the maximum number of iterations or the convergence of the objective function value.

[0141] A grid search method can be used to determine appropriate values for the parameters σ and λ in the improved NSGAIII algorithm. For a multi-objective optimization task on an IEEE 39-bus system, the value ranges for σ and λ are first set. Assume that the value range of σ is [0.1, 10] and the value range of λ is [0, 1]. This range is divided into several discrete points. For example, σ is taken in the range [0.1, 10] with a step size of 0.1, and λ is taken in the range [0, 1] with a step size of 0.01. This constitutes a series of parameter combinations. For each parameter combination, the improved NSGAIII algorithm is run on the IEEE 39-bus system, and the multi-objective optimization results (such as the coverage and diversity of the Pareto optimal solution set) are used as the evaluation basis. By comparing the performance of the algorithm under different parameter combinations, the σ and λ values that achieve the optimal results for this system are selected as the final parameter settings. For example, after many experiments, it was found that when σ = 1.5 and λ = 0.5, the Pareto optimal solution set obtained by the algorithm has the highest coverage and better diversity. Therefore, this set of parameter values is used when it is actually applied to the IEEE 39-bus system.

[0142] Example 2, an embodiment of the present invention, provides a method for generating and optimizing intelligent power system decisions based on behavioral reasoning. To verify the beneficial effects of the present invention, scientific demonstration was conducted through experiments. As a small, localized power system, a microgrid has unique operating characteristics and requirements. This example selects a microgrid consisting of distributed power sources (such as photovoltaic and wind turbines), energy storage devices (battery energy storage systems), and several local loads as the test object. The number of devices is determined based on the device types within the microgrid. Assume a total of 50 devices, including 10 distributed power sources, 20 energy storage units, and 20 loads. The feature dimension d is selected as 6, covering voltage, current, power, light intensity (for photovoltaic devices), temperature (affecting energy storage performance), and load demand. The device state matrix X is populated by real-time monitoring and collection of device operating parameters; the device association adjacency matrix A is constructed based on the electrical connection relationships between devices; and the device reliability index vector R is calculated by integrating historical device failure data, maintenance records, and inherent characteristics. "Behavioral fingerprint" modeling technology and an improved graph convolutional network are used to deeply explore device behavioral characteristics.

[0143] A behavioral inference engine was built using a two-layer LSTM temporal inference model and the Attention mechanism. The training data consisted of device behavior data collected every 10 minutes over the past six months in the microgrid, with 70% of the data used as the training set and 30% as the validation set. During training, the Adam optimization algorithm was used with a learning rate of 0.0005, an L2 regularization coefficient of 0.0005, and a dropout probability of 0.3. After training, the model can predict and infer the future behavior of microgrid devices.

[0144] Construct a Markov decision process (MDP) model applicable to microgrids. For the operation requirements of microgrids, establish a Pareto front search model. Solve it using the improved NSGA III algorithm, set the parameters σ = 1.2 and λ = 0.6 (determined by grid search), initialize the population size to 80, and the maximum number of iterations to 150.

[0145] When a fault occurs in the microgrid, the average fault recovery time of the traditional genetic algorithm is 30 minutes; for the method of combining behavior inference based on the support vector machine regression model and decision-making using the particle swarm optimization algorithm, the average fault recovery time is 25 minutes; while the average fault recovery time of the method of the present invention is only 15 minutes, which is 50% shorter than that of the traditional genetic algorithm and 40% shorter than that of the compared AI model combined with the traditional optimization algorithm. The operating cost of the microgrid optimized by the traditional genetic algorithm is 5000 yuan per day; the operating cost of the method based on the support vector machine regression model and the particle swarm optimization algorithm is 4500 yuan per day; the method of the present invention reduces the operating cost to 3500 yuan per day by optimizing the power generation and energy storage strategies, which is 30% lower than that of the traditional genetic algorithm and 22.2% lower than that of the compared method. Under the traditional genetic algorithm, the fluctuation range of the power balance degree of the microgrid is relatively large, with an average deviation of ±5%; the average deviation of the method based on the support vector machine regression model and the particle swarm optimization algorithm is ±3%; the method of the present invention controls the average deviation of the power balance degree within ±1.5% through accurate behavior inference and optimized decision-making, significantly improving the stability of the microgrid.

[0146] Embodiment 3, an embodiment of the present invention, provides an intelligent perception decision generation and optimization method for power systems based on behavior inference. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments. Select a regional power grid as an example of a high-proportion new energy system, in which the installed capacity ratio of new energy (wind power, photovoltaic) reaches 70%.

[0147] Determine the number of devices n, assuming a total of 200 devices including various new energy power generation devices, traditional thermal power devices, transmission lines, and substations. Set the feature dimension d to 8. In addition to the basic voltage, current, and power parameters, it also includes the wind speed (wind power) of new energy devices, light intensity (photovoltaic), coal consumption rate of thermal power devices, and temperature of transmission lines (affecting transmission losses), etc. Collect the device operation parameters to fill the device state matrix X; construct the device association adjacency matrix A according to the power grid topology; calculate the device reliability index vector R by integrating the device historical fault data, maintenance records, and inherent characteristics. Use the "behavior fingerprint" modeling technology and the improved graph convolutional network for device behavior feature mining.

[0148] Build a behavior inference engine with a double-layer LSTM time-series inference model and an Attention mechanism. The training data uses the device behavior data collected every 15 minutes in the regional power grid in the past year. 75% is used as the training set and 25% is used as the validation set. During training, the Adagrad optimization algorithm is adopted, with a learning rate of 0.001, an L1 regularization coefficient of 0.0003, and a Dropout probability set to 0.25. After training, it is used to predict the future behavior of the device. Construct an MDP model that conforms to the operating characteristics of a high-proportion new energy system. Establish a Pareto front search model and solve it using the improved NSGAIII algorithm, setting the parameters σ = 1.8 and λ = 0.4 (determined through grid search), initializing the population size to 120, and the maximum number of iterations to 200.

[0149] When using the traditional particle swarm optimization algorithm for scheduling, the new energy consumption rate is 75%; when using the method of combining behavior prediction with a traditional neural network and genetic algorithm for decision-making, the new energy consumption rate is 80%; the method of the present invention improves the new energy consumption rate to 90% by optimizing the scheduling strategies of power generation equipment and transmission networks, which is 20% higher than the traditional particle swarm optimization algorithm and 12.5% higher than the compared AI model combined with the traditional optimization algorithm.

[0150] The system risk coefficient under the traditional particle swarm optimization algorithm is 0.3; the risk coefficient of the method based on the traditional neural network and genetic algorithm is 0.25; the method of the present invention reduces the risk coefficient to 0.18 by comprehensively considering various risk factors and optimizing decisions, which is 40% lower than the traditional particle swarm optimization algorithm and 28% lower than the compared method. The system carbon emission under the traditional genetic algorithm is 1000 tons per day; the carbon emission of the method based on the traditional neural network and particle swarm optimization algorithm is 900 tons per day; the method of the present invention reduces the carbon emission to 750 tons per day by optimizing the operation of thermal power equipment and new energy equipment, which is 25% lower than the traditional genetic algorithm and 16.7% lower than the compared method.

[0151] Example 4 is an embodiment of the present invention, which provides an intelligent perception decision generation and optimization method for power systems based on behavior inference. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0152] Taking a wind turbine in a high - proportion new - energy system as an example, collect its operation data every 10 minutes in the past year. According to parameters such as wind speed and power, divide the operation status into 4 states: normal power generation, derated power generation, standby, and fault. Count the number of times of transitioning to other states at the next moment in different states. Suppose that in the 10,000 time steps counted, when in the normal power - generation state, there are 8,000 times of remaining in the normal power - generation state at the next moment, 1,000 times of transitioning to derated power generation, 800 times of transitioning to standby, and 200 times of transitioning to fault. Then the probability P1 of transitioning from the normal power - generation state to the normal power - generation state is 8000÷10000 = 0.8; the probability P2 of transitioning to the derated power - generation state is 1000÷10000 = 0.1; the probability P3 of transitioning to the standby state is 800÷10000 = 0.08; the probability P4 of transitioning to the fault state is 200÷10000 = 0.02. Through such statistical analysis of a large amount of historical data, determine the transition probability P between different states of each device k .

[0153] In the scenarios of micro - grid and high - proportion new - energy system, the grid - search method is used to determine the parameters σ and λ. Taking the high - proportion new - energy system as an example, set the value range of σ to be [0.5, 3] with a step size of 0.1; the value range of λ to be [0.1, 0.9] with a step size of 0.1. For each group of parameter combinations, run the improved NSGAIII algorithm on the high - proportion new - energy system, and use the multi - objective optimization results (such as indicators like the coverage rate and diversity of the Pareto - optimal solution set) as the evaluation basis. After a large number of experiments, it is found that when σ = 1.8 and λ = 0.4, the coverage rate of the Pareto - optimal solution set obtained by the algorithm in this system is the highest and the diversity is better. Therefore, this group of parameter values is adopted in this scenario. In the micro - grid scenario, the optimal parameter combination of σ = 1.2 and λ = 0.6 is also determined through grid - search

[0154] In the micro - grid and high - proportion new - energy system, the system load level and the average reliability of equipment are used as the input variables of fuzzy inference. For the system load level, a triangular membership - function is used to define the fuzzy sets of "low load", "medium load", and "high load". Suppose the load range of the micro - grid is [0, 1000kW]. When the load L is less than 200kW, the membership - function of "low load" is:

[0155]

[0156] When the load L is between 200kW and 600kW, the membership - function of "medium load" is:

[0157]

[0158] When the load L is greater than 600kW, the membership - function of "high load" is:

[0159]

[0160] For the average reliability of the device, assuming the value range is [0, 1], the triangular membership function is used to define the fuzzy sets of "low reliability", "medium reliability", and "high reliability". When the average reliability R avg is less than 0.3, the membership function of "low reliability" is:

[0161]

[0162] When the average reliability R avg is between 0.3 and 0.7, the membership function of "medium reliability" is:

[0163]

[0164] When the average reliability R avg is greater than 0.7, the membership function of "high reliability" is:

[0165]

[0166] According to these membership functions, the membership degrees of the system belonging to each fuzzy set are calculated based on the current load level and the average reliability of the device, and then fuzzy reasoning and defuzzification operations are carried out in combination with the pre-established fuzzy rules (such as "if the load level is 'low load' and the average reliability of the device is 'high', then α takes a smaller value, β takes a larger value, and γ takes a smaller value") to dynamically adjust the coefficients (α, β, γ) in the composite reward function.

[0167] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not restrictive. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

[0168] Embodiment 5, the fifth embodiment of the present invention, which is different from the previous four embodiments in that:

[0169] When the above-mentioned functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs.

[0170] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a predefined sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or in combination with these instruction execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in combination with an instruction execution system, apparatus, or device.

[0171] More specific examples (non-exhaustive list) of computer-readable media include the following: electrical connection parts with one or more wirings (electronic devices), portable computer disk cartridges (magnetic devices), random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memories), fiber optic devices, and portable compact disc read-only memories (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, then editing, interpreting, or otherwise processing it as necessary, and then storing it in a computer memory.

[0172] It should be understood that each part of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or combination of the following techniques well known in the art can be used: discrete logic circuits with logic gate circuits for implementing logical functions on data signals, application specific integrated circuits with suitable combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0173] Embodiment 6 is an embodiment of the present invention, which provides an intelligent perception decision-making generation and optimization system for a power system based on behavior reasoning, including a data perception and feature modeling module, a behavior reasoning and prediction module, a dynamic decision-making generation module, and a multi-objective optimization module;

[0174] The data perception and feature modeling module provides basic data support for the system, and forms a quantifiable analysis behavior feature space through multi-dimensional feature mapping and device relationship modeling;

[0175] The behavior reasoning and prediction module mines the behavior laws of devices from the time series dimension, and combines the attention mechanism to enhance the key information extraction ability, providing an accurate reasoning basis for decision-making generation;

[0176] The dynamic decision-making generation module balances the system stability, economy, and risk in a complex dynamic environment, and ensures the real-time and security of decisions through a hierarchical reinforcement learning architecture;

[0177] The multi-objective optimization module synthesizes the conflicting requirements of multiple objectives and outputs a flexible selection of optimization solution sets.

[0178] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. An intelligent perception decision-making generation and optimization method for power systems based on behavior reasoning, characterized in that: including Constructing a device behavior feature space by real-time monitoring of the operating parameters of devices in the power system; Based on the device behavior feature space, using a time series reasoning model to predict device behavior and generating a behavior reasoning result; Combining the behavior reasoning result and environmental parameters to construct a Markov decision process model and generating an optimal decision that satisfies safety constraints; Establishing a multi-objective optimization model according to the optimal decision, and solving the Pareto optimal solution set through an improved optimization algorithm to realize the operation optimization of the power system.

2. The intelligent perception decision generation and optimization method for power systems based on behavior reasoning according to claim 1, characterized in that: The construction of the device behavior feature space includes a device state matrix, a device association adjacency matrix, and a device reliability index vector. The device state matrix is generated by collecting the voltage, current, and power parameters of all devices in the power system, and the matrix dimension is determined by the number of devices and the feature dimension. The device association adjacency matrix is generated according to the physical connection relationship between devices, and the matrix elements represent the direct relevance between devices. The device reliability index vector is generated by comprehensively calculating the historical fault data, maintenance records, and inherent characteristics of the device.

3. The intelligent perception decision generation and optimization method for power systems based on behavior reasoning according to claim 2, wherein: The time series reasoning model includes using a double-layer long short-term memory network model combined with an attention mechanism to perform time series reasoning on device behavior characteristics. The first layer of the model receives the device behavior feature sequence, and the second layer performs high-level feature extraction on the output of the first layer; the attention mechanism realizes the fusion of local features and global features through the correlation score between the global state encoding and the current hidden state.

4. The intelligent perception decision generation and optimization method for power systems based on behavior reasoning according to claim 3, characterized in that: The prediction of device behavior includes training a two-layer LSTM time series inference model, and using Stochastic Gradient Descent (SGD) or an improved algorithm to update the parameters of the model. The parameters include the weight matrices W f 、W i 、W c 、W o and the bias vectors b f 、b i 、b c 、b o and the parameters in the Attention mechanism; During the training process, the historical data of device behavior is divided into a training set and a validation set according to the time series. The model is optimized by minimizing the loss function. At the same time, regularization methods and Dropout techniques are used to prevent overfitting. The trained model is deployed to the corresponding computing platform to receive device behavior data in real time, and behavior reasoning is performed through a double-layer LSTM time series reasoning model and an Attention mechanism to provide a decision-making basis.

5. The intelligent perception decision generation and optimization method for a power system based on behavior reasoning according to claim 4, characterized in that: The construction of the Markov decision process model includes a device behavior feature space and environmental parameters, and the action space covers all executable operations of the devices. The hierarchical reinforcement learning architecture includes a policy network, a value network, and a constraint network, and the decision satisfies the safety constraints through an adaptive penalty coefficient.

6. The method for generating and optimizing intelligent perception decision of power system based on behavior reasoning according to claim 5, characterized in that: The establishment of the multi-objective optimization model includes the operating cost, risk coefficient, and carbon emission, which are expressed as: min[f1(x),f2(x),f3(x)] = [operating cost, risk coefficient, carbon emission For a system consisting of two or more devices, the operating cost is expressed as the sum of the operating costs of individual devices. Suppose there are n devices in the system, and the operating cost function of each device i is c i (x), then the operating cost where x is the decision variable vector; The risk coefficient is represented by the product of the failure probability and the risk loss, that is where p i (x) is the failure probability of device i under the decision variable x, and l i (x) is the loss when device i fails; Carbon emissions come from the energy combustion process of power generation equipment. The carbon emission coefficient of power generation equipment i is set as e i , and its power generation is p i (x), then the carbon emissions 7. The method for generating and optimizing the intelligent perception decision of the power system based on behavior reasoning according to claim 6, characterized in that: The solution of the Pareto optimal solution set includes enhancing the diversity of the optimization solution set by introducing a behavior similarity constraint. The formula of the behavior similarity constraint is where y i and y j are two individuals in the population, H(y) is the behavior feature vector of individual y, which is obtained by mapping through the device behavior feature space obtained in the behavior perception modeling stage, ||·|| represents the Euclidean norm, and ∈ is a preset parameter used to control the attenuation rate of similarity.

8. A system adopting the method for generating and optimizing intelligent perception decision of a power system based on behavior reasoning as described in any one of claims 1 to 7, characterized in that: including a data perception and feature modeling module, a behavior reasoning and prediction module, a dynamic decision generation module, and a multi-objective optimization module; The data perception and feature modeling module provides basic data support for the system, and forms a quantifiable and analyzable behavior feature space through multi-dimensional feature mapping and device relationship modeling. The behavior reasoning and prediction module mines the device behavior law from the time series dimension, combines the attention mechanism to enhance the key information extraction ability, and provides an accurate reasoning basis for decision generation. The dynamic decision-making generation module balances system stability, economy, and risk in a complex dynamic environment, and ensures the real-time performance and safety of decisions through a hierarchical reinforcement learning architecture; The multi-objective optimization module synthesizes multi-objective conflicting requirements and outputs an optimizable solution set that can be flexibly selected.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 7.

Images