A power system intelligent control and decision method based on a multi-layer cognitive architecture

Abstract

This invention discloses an intelligent control and decision-making method for power systems based on a multi-layer cognitive architecture. The method includes: receiving real-time measurement data from the power system through a perception layer and performing state estimation based on an extended Kalman filter algorithm to generate a system state vector; generating a candidate set of control strategies based on the system state vector using a liquid neural network through an intuition layer; verifying the feasibility of the candidate control strategies based on neural symbolic constraints through a logic layer and prioritizing the verified candidate control strategies to obtain the selected control strategies; rapidly solving the AC optimal power flow problem using physical information neural operators through a physical layer and outputting physical field prediction results; and selecting and generating control commands through a decision coordination layer using multi-objective Pareto optimization and reinforcement learning agents based on the physical field prediction results. This invention enables fast, accurate, and reliable intelligent control of power systems.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for power systems, and more specifically to an intelligent control and decision-making method for power systems based on a multi-layer cognitive architecture. Background Technology

[0002] Intelligent control of power systems is a core technology of modern energy systems. Through advanced algorithms and models, it enables the safe, stable, and efficient operation of power systems. With the large-scale integration of renewable energy sources such as wind power and photovoltaics, the operating conditions of power systems are becoming increasingly complex, load fluctuations are intensifying, and network topology is dynamically changing. The system faces multi-dimensional technical challenges, including real-time performance, physical feasibility, and multi-objective collaborative optimization.

[0003] From an engineering perspective, intelligent control of power systems currently faces three main challenges: First, there is a conflict between computational efficiency and real-time performance; traditional optimization methods take more than 10 minutes to solve, failing to meet the requirements for second-level control. Second, satisfying physical constraints is difficult; existing methods often employ simplified DC power flow models, resulting in errors of 10-30% in scenarios with high renewable energy penetration. Third, multi-objective coordination capabilities are insufficient; traditional fixed-weight methods cannot dynamically balance multiple dimensions such as safety, economy, and environment. Furthermore, purely data-driven methods lack interpretability, making it difficult to meet the traceability requirements for safe power system operation.

[0004] Therefore, how to overcome the shortcomings of existing technologies in terms of real-time performance, physical constraint satisfaction, multi-objective collaborative optimization, environmental adaptability and interpretability, so as to achieve fast, accurate and reliable intelligent control of power systems, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, the present invention proposes a power system intelligent control and decision-making method based on a multi-layer cognitive architecture, so as to overcome the above problems or at least partially solve the above problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution;

[0007] In a first aspect, embodiments of the present invention provide a method for intelligent control and decision-making in power systems based on a multi-layer cognitive architecture, comprising the following steps: The system receives real-time measurement data from the power system through the sensing layer and performs state estimation based on the extended Kalman filter algorithm to generate a system state vector. Based on the system state vector, the intuitive layer uses a liquid neural network to generate a set of control strategy candidates. The feasibility of the control strategy candidate set is verified by the logic layer based on neural symbol constraints, and the candidate control strategies that pass the verification are prioritized to obtain the candidate control strategies. The physical layer utilizes physical information neural operators to quickly solve the AC optimal power flow problem for the candidate control strategies, and outputs physical field prediction results including node marginal electricity price, generation scheduling, line power flow and congestion probability. Furthermore, based on the prediction results of the physical field, the decision coordination layer utilizes multi-objective Pareto optimization and reinforcement learning agents to select policies and generate control commands.

[0008] Preferably, the real-time measurement data of the power system includes PMU measurement data, SCADA telemetry data, and / or weather forecast data.

[0009] Preferably, when performing state estimation based on the extended Kalman filter algorithm, a covariance matrix is ​​output synchronously, which is used to calculate the Mahalanobis distance of the measurement residuals based on the covariance matrix, and to issue an alarm for abnormal measurement data based on the Mahalanobis distance and a preset threshold.

[0010] Preferably, the liquid neural network includes multiple liquid time constant units, and the time constant of each liquid time constant unit is dynamically adjusted according to the power grid pressure index; the power grid pressure index is obtained by a weighted combination of node voltage deviation and system frequency deviation.

[0011] Preferably, the neural symbol constraint is a first-order logic rule, including one or more of voltage constraints, thermal limitation constraints, reserve capacity constraints, ramp rate constraints, and power balance constraints.

[0012] Preferably, the feasibility of the candidate control strategies is verified, and the verified control strategies are prioritized, including: The neural symbolic constraints are converted into SMT formulas, and the SMT solver is used to verify whether the candidate control strategy satisfies all the constraint rules. For the candidate control strategies that pass the verification, a comprehensive score is calculated based on the priority of the constraint rules, and the strategies are then ranked according to the priority of the comprehensive scores.

[0013] Preferably, the physical information neural operator includes, from input to output, an encoder, a Fourier block layer, and three parallel prediction heads connected thereto; The three parallel prediction heads are used to output the node marginal electricity price, generation scheduling, and line power flow, respectively.

[0014] Preferably, the physical layer is further used to identify congested lines and calculate the congestion probability based on the line power flow predicted by the neural operator of the physical information.

[0015] Preferably, the physical information neural operator is trained using a joint loss function that includes data fidelity loss, physical residual loss, and boundary condition loss; The physical residual loss includes the power balance equation residual sub-loss, voltage constraint residual sub-loss, and power flow consistency constraint residual loss.

[0016] Preferably, the multi-objective Pareto optimization includes five optimization objectives: safety, reliability, economy, environment, and response speed. A non-dominated sorting genetic algorithm is used to generate a Pareto front solution set. The reinforcement learning agent is used to learn to select the optimal control strategy from the Pareto front solution set under different operating conditions based on historical operating data and control effect feedback.

[0017] Preferably, the decision coordination layer is also used to generate interpretable reports, including SHAP value feature importance analysis, rule trigger logs, inference path visualization, and confidence scores, for operators to view and review.

[0018] Secondly, this embodiment also provides a power system intelligent control and decision-making system based on a multi-layer cognitive architecture. This system is used to implement the power system intelligent control and decision-making method based on a multi-layer cognitive architecture as described in any of the preceding embodiments, including: The perception layer module is used to receive real-time measurement data of the power system and perform state estimation based on the extended Kalman filter algorithm to generate a system state vector. The intuition layer module is used to generate a set of control strategy candidates based on the system state vector using a liquid neural network; The logic layer module is used to perform feasibility verification on the candidate set of control strategies based on neural symbol constraints, and to prioritize the candidate control strategies that pass the verification to obtain the candidate control strategies. The physical layer module is used to quickly solve the AC optimal power flow problem for the candidate control strategy using physical information neural operators, and outputs physical field prediction results including node marginal electricity price, generation scheduling, line power flow and congestion probability. The decision coordination layer module is used to select policies and generate control commands based on the physical field prediction results, using multi-objective Pareto optimization and reinforcement learning agents.

[0019] This invention provides a method for intelligent control and decision-making in power systems based on a multi-layer cognitive architecture. Compared with existing technologies, the advantages include: 1. Real-time response within seconds is achievable. This invention achieves state estimation delay within 50ms through extended Kalman filtering; it maintains stable response under normal operating conditions and responds quickly under abnormal operating conditions through the adaptive time constant mechanism of liquid neural network, avoiding the slow response or oscillation problems of traditional fixed time constant methods; furthermore, it replaces traditional numerical optimization methods with physical information neural operators, shortening the AC optimal power flow solution time from more than 10 minutes to less than 1 second, meeting the timeliness requirements of real-time control of power systems.

[0020] 2. Strictly meets physical constraints. This invention uses a physical information neural operator to solve the complete AC power flow equations. Compared with the traditional simplified DC power flow model, it effectively reduces voltage prediction error and line power flow prediction error in scenarios with high renewable energy penetration. By introducing physical residual loss and boundary condition loss during the training phase, it ensures that the prediction results meet basic physical constraints such as power balance, voltage limitation, and thermal limitation.

[0021] 3. Strong multi-objective collaborative optimization capability. This invention generates a non-dominated solution set encompassing five objectives: safety, reliability, economy, environment, and response speed, through Pareto optimization, avoiding the limitations of traditional fixed-weight methods. It also learns optimal trade-off strategies under different operating conditions through reinforcement learning agents, achieving dynamic adjustment of objective weights. Experiments show that compared to traditional single-objective optimization methods, this invention improves economy by 12-18% and reduces carbon emissions by 8-15% while ensuring safety.

[0022] 4. Strong environmental adaptability and interpretability. This invention utilizes an adaptive mechanism based on a liquid neural network that dynamically adjusts the time constant according to grid pressure indicators. Under normal operating conditions, it maintains stable output to suppress measurement noise, and under abnormal operating conditions, it automatically reduces the time constant to achieve rapid response, thus adapting to dynamic environments such as load fluctuations, topology changes, and uncertainties in new energy sources. Through neural symbolic constraints, the logical reasoning process of control decisions is made transparent, generating interpretable reports including SHAP value analysis, rule trigger logs, and inference path visualization, meeting the traceability requirements for safe power system production. Under abnormal operating conditions, the system can provide decision confidence assessment and risk warnings, enhancing operators' trust in the intelligent control system.

[0023] 5. Modularity and scalability. This invention adopts a multi-layered cognitive architecture with decoupled design at each layer, supporting independent upgrades and replacements; it supports expansion from a single scenario to joint optimization across multiple scenarios, exhibiting good engineering applicability. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the extended Kalman filter state estimation process for the perception layer provided in an embodiment of the present invention.

[0026] Figure 2 This is a schematic diagram of the adaptive time constant adjustment mechanism of the intuitive layer liquid neural network provided in an embodiment of the present invention.

[0027] Figure 3 This is a schematic diagram of the logic layer neural symbolic constraint reasoning process provided in an embodiment of the present invention.

[0028] Figure 4 This is a schematic diagram of the physical layer physical information neural operator architecture provided in an embodiment of the present invention.

[0029] Figure 5 This is a schematic diagram of the Pareto optimization and reinforcement learning strategy selection process of the decision coordination module provided in this embodiment of the invention.

[0030] Figure 6 This is a schematic diagram illustrating the interpretability output provided in an embodiment of the present invention.

[0031] Figure 7 This is a schematic diagram of the overall architecture of the multi-layer cognitive architecture power system intelligent control and decision-making system provided in this embodiment of the invention.

[0032] Figure 8 This is a timing diagram of the system control cycle provided in an embodiment of the present invention.

[0033] Figure 9 This is a comparison chart of the physical constraint satisfaction rate of the physical information neural operator provided in the embodiments of the present invention.

[0034] Figure 10 This is a graph showing the adaptive time constant response of the liquid neural network provided in an embodiment of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] This invention discloses an intelligent control and decision-making method for power systems based on a multi-layer cognitive architecture, comprising the following steps: The system receives real-time measurement data from the power system through the sensing layer and performs state estimation based on the extended Kalman filter algorithm to generate a system state vector. Based on the system state vector, the intuitive layer uses a liquid neural network to generate a set of control strategy candidates. The feasibility of the control strategy candidate set is verified by the logic layer based on neural symbol constraints, and the verified control strategies are prioritized to obtain the candidate control strategies. The physical layer employs physical information neural operators to quickly solve the AC optimal power flow problem for the candidate control strategies, and outputs physical field prediction results including node marginal electricity price, generation scheduling, line power flow and congestion probability. Furthermore, based on the prediction results of the physical field, the decision coordination layer utilizes multi-objective Pareto optimization and reinforcement learning agents to select policies and generate control commands.

[0037] This application constructs an end-to-end intelligent control closed-loop system for power systems through a multi-layered progressive inference architecture, realizing real-time intelligent control of power systems and supporting flexible switching and collaborative optimization under various operating conditions (including normal economic dispatch, rapid response to abnormal conditions, and adaptation to topology changes).

[0038] In one embodiment, the perception layer is used to receive real-time measurement data of the power system, perform state estimation through the extended Kalman filter algorithm, and generate a system state vector; wherein, the real-time measurement data of the power system includes one or more multi-source heterogeneous data from phasor measurement unit (PMU) measurement data, SCADA telemetry data, and meteorological forecast data.

[0039] In some implementations, a data acquisition and interface module is configured to collect real-time operational data of the power system: PMU measurement data is typically provided at a rate of 50 frames per second or 60 frames per second for node voltage amplitude. Voltage phase angle ,frequency and the rate of change of frequency ROCOF ( The voltage amplitude range is 0.95-1.05 pu, and the voltage phase angle range is... arrive Radius, frequency range 49.8-50.2 Hz; SCADA telemetry data provides generator active power output in 2-4 second cycles. Unproductive efforts Load demand Information such as power flow and transmission lines; meteorological forecast data including wind speed, solar intensity, temperature, and other renewable energy output forecasts. These raw data, after time synchronization, bad data detection, and format standardization, are formed into a standardized system status observation sequence. This information is then transmitted to the perception layer.

[0040] In this embodiment, the perception layer is used to fuse multi-source measurement data, suppress measurement noise, and estimate the complete state vector of the power system online; the estimation process is as follows: Figure 1 As shown, specifically, the extended Kalman filter algorithm is used based on the nonlinear dynamic model of the power system to analyze the system state observation sequence. Recursive state estimation is performed, where the nonlinear dynamic model of the power system consists of state equations and measurement equations. The state equations describe the dynamic evolution of the system state over time, while the measurement equations describe the mapping relationship between observable measurement data and the indirectly observable system state. Together, they constitute a complete nonlinear state-space model, specifically expressed as follows:

[0041]

[0042] In the formula, For a moment The system state vector includes the voltage magnitude of each node. Voltage phase angle System frequency Rate of change of frequency Load demand A total of 64 features; For the state dimension; To control the input vector; It is a nonlinear state transition function that describes the system state from time 1 to 2. At the time The evolutionary pattern; For nonlinear measurement functions, the system state vector is... It is mapped to an observable measurement space, that is, it describes the nonlinear mapping relationship between the system state and the actual measurement data; For process noise, The process noise covariance matrix; To measure noise, To measure the noise covariance matrix; It represents the system state observation sequence, including voltage, phase angle, frequency, ROCOF measured by PMU, and active power and reactive power measured by SCADA.

[0043] The Extended Kalman Filter (EKF) algorithm extends the Kalman filter to nonlinear systems by linearizing the nonlinear function. The specific recursive process is divided into a prediction step and a correction step: Prediction step, based on the system state equation Predict the state at the next moment and its covariance matrix ;

[0044]

[0045] in, For a moment State estimate; For time-based Information on time The predicted state value; For a moment The estimation error covariance matrix; For a moment The prediction error covariance matrix; State transition function exist The Jacobian matrix at point A is calculated as follows:

[0046] Calibration step, based on the measurement equation and Kalman gain Update state estimation and its covariance matrix :

[0047]

[0048] in, Here is the Kalman gain matrix. ; The dimension of the observation vector; For measurement function exist The Jacobian matrix at point A is calculated as follows:

[0049] It is the identity matrix; To measure the residual vector.

[0050] The final output is the fused system state vector. ,in Representing the state dimension (in this embodiment) (Including characteristics such as voltage, phase angle, frequency, and load at each node). Furthermore, to identify measurement anomalies, based on the covariance matrix... Calculate the Mahalanobis distance of the measurement residuals ,

[0051] in, Let be the residual covariance matrix.

[0052] Then, an anomaly alarm is triggered based on the Mahalanobis distance and a preset threshold, i.e., when the Mahalanobis distance... Exceeding the preset threshold (usually taken) When the value is within the 99.7% confidence interval, an anomaly alarm is triggered, and suspicious measurement points are marked as bad data and removed or downweighted.

[0053] In this embodiment, the sampling period of the extended Kalman filter algorithm is set to 20ms (corresponding to 50Hz), and the process noise covariance is... Based on the dynamic characteristics of the power system, it is set as a diagonal matrix, with diagonal elements as follows: pu²; Measurement noise covariance Based on the measurement accuracy settings of the PMU and SCADA, the standard deviation of the PMU voltage measurement noise is: pu, the standard deviation of phase angle measurement noise is rad, SCADA power measurement noise standard deviation is With this configuration, the EKF state estimation delay is approximately 30-50ms, and the root mean square error (RMSE) of the state estimation error is approximately 0.02 pu, which meets the requirements for real-time control.

[0054] In one embodiment, the intuition layer is used based on the system state vector. A candidate set of control strategies is generated using a liquid neural network. , This represents the number of candidate control strategies; in this embodiment, M = 10-20. Each candidate control strategy includes the predicted bidding price for each node. (Unit: MWh and electricity forecast) (Unit: MW).

[0055] The liquid neural network serves as the core control strategy generator and comprises multiple layers of liquid time constant units. In one specific implementation, the complete liquid neural network architecture includes an input projection layer, two layers of liquid time constant units, a multi-head attention layer, and an output head.

[0056] The input projection layer is:

[0057] In the formula, This is the system state vector output by the perception layer; The projection weight matrix maps the input from 64 dimensions to 128 dimensions; It is the bias vector; The activation function for the Gaussian error linear unit; This is a layer normalization operation.

[0058] Liquid time constant unit layer:

[0059]

[0060] in, These are the outputs of the liquid time constant units for the first and second layers, respectively. and These are the first and second layers of liquid time constant units, with each layer having a hidden dimension of 128.

[0061] Multi-head attention layer:

[0062] in, It is a 4-head attention mechanism with a head dimension of 32, used to capture the spatial dependencies between different nodes.

[0063] Output head: Price prediction header:

[0064] Battery prediction head:

[0065] in, Predicted bidding prices for each node (unit: / MWh). The number of power grid nodes (in this embodiment, ); Power forecast for each node (unit: MW); These are the output weight matrices for the price prediction head and the power prediction head, respectively; These are the output bias vectors for the price prediction head and the power prediction head, respectively; It is a soft positive definite activation function, ensuring that the power prediction output is non-negative.

[0066] In this implementation scheme, the total number of parameters in the liquid neural network architecture is approximately 158,000, and the latency for a single forward inference is approximately 10-20ms (approximately 19ms when the batch size is 1 and approximately 10.8ms when the batch size is 100).

[0067] In one specific embodiment, the kinetic equations of the liquid time constant element are expressed based on ordinary differential equations (ODEs):

[0068] In the formula, For the first A hidden neuron at time... The hidden state; The input vector; For the input weight matrix, For the hidden layer dimension; This is a cyclic weight matrix; It is the bias vector; The activation function (the tanh function is used in this embodiment); For the first The liquid time constant of each neuron is dynamically adjusted to adapt to grid pressure. In some preferred embodiments, the time constant of each liquid time constant unit is based on grid pressure indicators. Dynamic adjustments are made to achieve adaptive response to power grid conditions, maintaining stable output under normal conditions and responding rapidly under abnormal conditions. Specifically, the liquid time constant... The adjustment rules are as follows:

[0069] in, It is the minimum time constant (corresponding to the fast response mode). This represents the maximum time constant (corresponding to the stationary filtering mode). This is the threshold value for the pressure index; when the power grid is operating normally ( The liquid neural network maintains a large time constant to suppress measurement noise; when the power grid experiences an anomaly ( The liquid neural network automatically switches to a fast response mode, reducing the time constant to about 0.1 and increasing the response speed by about 9 times.

[0070] Power grid pressure indicators It is obtained by a weighted combination of node voltage deviation and system frequency deviation, and the process is as follows: Figure 2 ,Right now

[0071] Represents the node voltage deviation vector. , pu is the rated voltage. Represents the system frequency deviation vector. , Hz is the rated frequency. , These are the corresponding weighting coefficients, preferably... , , This represents the L2 norm.

[0072] In one embodiment, the logic layer is used to verify the feasibility of the control strategy candidate set based on neural symbolic constraints, wherein the neural symbolic constraints are first-order logic rules, including one or more of voltage constraints, thermal limitation constraints, reserve capacity constraints, and ramp rate constraints. In some specific implementations, voltage constraint rules... , For nodes voltage amplitude, pu、 pu represents the upper and lower voltage limits; the thermal limiting rule requires that the apparent power flow of all lines not exceed the thermal limit, i.e. ,in For the collection of routes, For the line The power flow in the view, For the line Thermal limitations (apparent power capacity); reserve capacity constraints require the system to maintain sufficient spinning reserve capacity, i.e. etc., among which For generator sets, For generator Those who have made contributions For generator Maximum output The required spinning reserve capacity (typically 10-15% of total load); the ramp rate constraint rule limits the rate of change of generator power, i.e. ,in For generator The change in power, For generator Climbing speed limit (unit: MW / min). To control the cycle (in this embodiment) The power balance constraint rule requires that the total power generation equals the total load plus network losses, i.e. ,in For the set of load nodes, For nodes The load demand, This is for network losses (typically 2-5% of the total load).

[0073] To further optimize the above technical solution, such as Figure 3 As shown, the rule buffer stores the above constraint rules in YAML format, and assigns a priority weight to each rule. Rule priority reflects the importance of the constraint in actual operation; for example, voltage constraints and thermal limits usually have the highest priority. Economic constraints have a lower priority. Operators can dynamically modify rule priorities through configuration files without retraining the model.

[0074] Furthermore, the aforementioned first-order logic rules are converted into SMT formulas, and a solver is used to verify their satisfiability. For each candidate strategy output by the intuition layer... The SMT solver determines whether the strategy satisfies all constraint rules:

[0075] in, For a set of constraint rules; For rules Applied to strategy The logical formula; For logical AND operation. If then strategy If it is a feasible strategy, then it is filtered out.

[0076] For feasible strategies, a comprehensive score is calculated based on rule priority. This score is used to rank feasible strategies and output the top-ranked strategies. The strategy with the highest score is then used for further evaluation by the physical layer module. The overall score is calculated as follows:

[0077] in, For indicator functions, if the rule If the condition is met, the value is 1; otherwise, it is 0.

[0078] In this embodiment, the single-constraint solution latency of the logic layer is approximately 50-100ms (depending on the number of constraint rules and candidate strategies). Through incremental solution and constraint caching optimization, the solution latency can be further reduced to 20-30ms.

[0079] In one embodiment, the physical layer is used for the control strategy to be selected. Physical field prediction is performed. Specifically, the physical information neural operator is used to quickly solve the AC-OPF optimal power flow problem. The physical information neural operator achieves millisecond-level AC-OPF solution by transmitting global information in the frequency domain, and the single inference delay is much smaller than that of traditional numerical optimization methods.

[0080] In this embodiment, the physical information neural operator is based on a Fourier neural operator architecture and achieves millisecond-level AC-OPF solving by performing global information transfer in the frequency domain. Its structure is referenced... Figure 4Specifically, it includes one encoder, six Fourier blocks (each block contains a spectral convolutional layer and an activation function), and three decoder heads (which output the node marginal price (LMP) and generation dispatch respectively). Line flow ).in, The encoder is represented as:

[0081] In the formula, The input grid state characteristics include the voltage amplitude (pu), voltage phase angle (rad), frequency (Hz), rate of change of frequency (ROCOF) (Hz / s), load demand (pu), and current (pu) of each node. The number of power grid nodes (in this embodiment) ); The encoder weight matrix maps the input from 6 dimensions to a 128-dimensional feature space; This is the encoder bias vector.

[0082] Fourier blocks (6 layers in total):

[0083] Each Fourier block contains spectral convolutional layers, local convolutional layers, and residual connections:

[0084] in, It is a local 1×1 convolution weight matrix; For the first The Fourier convolution operator of the layer retains the first 16 frequency modes; This is the GELU activation function.

[0085] Decoder Heads (3 output heads in total): LMP (Local Marginal Price) forecast header;

[0086] Power generation dispatch forecast head:

[0087] Line power flow prediction head:

[0088] in, Forecast of marginal electricity price at nodes (unit: / MWh) This represents the number of nodes in the power grid. For power generation dispatch forecast (unit: pu). This refers to the number of generators. Line apparent power flow prediction (unit: pu). Number of lines; These are the upper and lower limits of the generator's active power output, respectively. These are the weight matrices for the nodal marginal electricity price forecast head, the generation dispatch forecast head, and the line power flow forecast head, respectively. These are the bias vectors for each output head.

[0089] As a priority, based on the line power flow predicted by the physical information neural operator, congested lines are identified and congestion probabilities are calculated, thereby outputting physical field prediction results that include node marginal electricity prices, generation scheduling, line power flow, and congestion probabilities.

[0090] Furthermore, the physical information neural operator is trained using a joint loss function that includes data fidelity loss, physical residual loss, and boundary condition loss; this joint loss function is expressed as:

[0091] In the formula, For the joint loss function; For data fidelity loss; This refers to the physical residual loss (including power balance equation residuals, voltage constraint residuals, and power flow consistency constraint residuals). For boundary condition loss, weighting coefficients , In some implementation schemes, Data fidelity loss for:

[0092] in, The marginal electricity price at nodes predicted by the physical information neural operator; The actual nodal marginal electricity price labels obtained by solving AC-OPF using traditional optimization methods in the training data; For the predicted active power output of the generator; Labels for the actual active power output of generators in the training data; For the predicted line apparent power flow; Labels for the apparent power flow of real lines in the training data; It is an L2 norm.

[0093] Physical residual loss Includes three sub-losses: The power balance equation residual measures the difference between total power generation and load demand plus grid losses, and is calculated as follows: ,in For generator sets, For the set of load nodes, For the predicted generator Those who have made contributions For nodes The load demand, For predicted network loss; The voltage constraint residual uses a soft constraint method to penalize voltage over-limit behavior, and is calculated as follows: ,in For the number of power grid nodes, For the predicted nodes voltage amplitude, pu、 pu represents the upper and lower voltage limits; Power flow consistency constraint residuals ensure that the predicted line power flow satisfies the nodal admittance matrix-based requirement. The power flow calculation pattern is calculated as follows: ,in For the predicted line apparent power flow, This is a power flow calculation function based on the nodal admittance matrix. This is the predicted node voltage magnitude vector. This is the predicted node voltage phase angle vector. Here is the nodal admittance matrix; The overall physical residual loss is .

[0094] Boundary condition loss :

[0095] in, For generator sets; For generator The lower limit of meritorious contribution; For generator The upper limit of effective output; For the predicted generator Those who have made meritorious contributions; This is a function that maximizes the output of a generator and is used to penalize generators that exceed their output limits.

[0096] In this embodiment, the weight coefficients of the loss function are set to , .

[0097] With this configuration, PINO has a verification loss of approximately 10.24 on the test set, a power balance error of approximately 0.076 pu (corresponding to a relative error of approximately 7.6%), and a thermal limit violation rate of 0%.

[0098] In this embodiment, the physical information neural operator supports batch parallel solution of the physical fields corresponding to multiple candidate strategies, that is, for the output of the logic layer... The physical information neural operator inputs a batch of candidate strategies into the model and outputs five sets of physical field prediction results at once:

[0099] in, For strategy The corresponding physics prediction results. When the batch size is 1000, the single inference latency is approximately 55ms; when the batch size is 10000, the single inference latency is approximately 380ms, which is much less than the solution time of more than 10 minutes for traditional numerical optimization methods.

[0100] In this embodiment, the total number of parameters of the physical information neural operator is approximately 1.9 million, and the model size is approximately 40.8 MB. The inference stage employs FP16 mixed precision acceleration, achieving an inference throughput of approximately 2000 scenes / second on a high-performance GPU accelerator card.

[0101] In one embodiment, the decision coordination layer is used for predicting physical fields based on the output of the physical layer. The final policy selection is achieved through multi-objective Pareto optimization and reinforcement learning agents, thereby generating control commands.

[0102] Some implementation schemes, such as Figure 5 As shown, the multi-objective Pareto optimization is based on five optimization objectives: safety, reliability, economy, environment, and response speed. The Pareto front solution set is generated using a non-dominated sorting genetic algorithm.

[0103] security goals The system safety margin is measured by minimizing the voltage deviation, and is calculated as follows: ,in For strategy corresponding nodes Voltage amplitude prediction pu is the rated voltage; Reliability targets The ability of a system to respond to emergencies is measured by maximizing its reserve capacity, and is calculated as follows: ,in For generator The upper limit of effective output, For strategy Corresponding generator For active power output prediction, the objective function is negative to minimize the output. This is equivalent to maximizing reserve capacity; Economic objectives By minimizing power generation costs to reduce operating costs, calculated as follows: ,in For generator The quadratic cost function, Cost coefficient; Environmental goals Promoting low-carbon development by minimizing carbon emissions, calculated as ,in Carbon emission intensity of generator g (unit: g) / MWh), for renewable energy generators For thermal power units Usually ; Response speed target By minimizing changes in control commands and avoiding frequent adjustments, the calculation is as follows: ,in For the generator at the previous moment The contribution of this objective is to improve control stability.

[0104] The Pareto optimization submodule uses the Non-Dominated Sorting Genetic Algorithm (NSGA-II) to generate the Pareto front solution set. The NSGA-II algorithm flow is as follows: Step 1: Initialize the population ,Include A random strategy.

[0105] Step 2: For each strategy Calculate its five objective function values. .

[0106] Step 3: Perform non-dominated ranking, dividing the strategies into different dominance levels. .Strategy Domination strategy If and only if No inferior to in all objectives And strictly superior to at least one objective .

[0107] Step 4: Calculate the crowding distance to measure the density of each policy in the target space. Crowding distance The calculation is as follows:

[0108] in, , Strategy in the target space Adjacency strategy; , For the goal The maximum and minimum values.

[0109] Step 5: Generate offspring population through tournament selection, crossover, and mutation operations. .

[0110] Step 6: Merge parent and child generations Re-perform the non-dominated sorting and crowding distance calculation, and select the previous one. A strategy as the next generation .

[0111] Step 7: Repeat steps 2-6 until the maximum number of iterations is reached. .

[0112] Algorithm outputs first frontier All non-dominated strategies constitute the Pareto front solution set. .

[0113] In some implementations, the reinforcement learning agent selects the optimal control policy from the Pareto front solution set. Specifically, it employs Deep Q-Network (DQN) or Proximal Policy Optimization (PPO) algorithms, learning a state-action value function based on historical operating data and control performance feedback. The optimal strategy is selected under different power grid operating conditions. The final selected control strategy is converted into specific control commands by the control command generation submodule and issued to the power system for execution. In one specific implementation scheme, the learning process includes: State representation: The state of the agent Including current power grid status characteristics (64 dimensions), historical control effect evaluation (including cost, voltage deviation, reserve capacity, etc. of the past 5 control cycles), and Pareto front solution set feature statistics (including solution set size, objective function mean and variance, etc.), totaling approximately 100 features.

[0114] Action Space: The action space of the agent is a Pareto front solution set. The strategy index in, i.e. ,in The size of the Pareto front solution set (usually 10-20).

[0115] Reward function: The agent's reward function A weighted combination of multiple objectives is taken into account:

[0116] in, The target weights are dynamically adjusted based on the power grid's operating conditions. Under normal operating conditions, economic and environmental factors have higher weights. , Under abnormal operating conditions, safety and reliability have higher weightings. , ).

[0117] Q-network architecture: The Q-network uses a multilayer perceptron (MLP) architecture, including 3 hidden layers, each with 256 neurons, and the activation function is ReLU.

[0118] in, These are the Q-network parameters.

[0119] Training process: DQN employs an experience replay and target network mechanism for training. The agent interacts with the power system environment, transmitting experience tuples... The data is stored in an experience pool, and batches of data are randomly sampled from the experience pool to update the Q-network parameters. The loss function of the Q-network is:

[0120] in, For experience pool; Discount factor; The target network parameters are updated every 100 training steps.

[0121] After training, the agent can determine the current state. Choose the action that maximizes the Q value:

[0122] As a preferred implementation, it simultaneously generates interpretable reports, including SHAP value feature importance analysis, rule trigger logs, inference path visualization, and confidence scores, for operators to review the decision-making process. This implementation plan specifically includes, for example... Figure 6 As shown.

[0123] SHAP Value Feature Importance Analysis: The SHAP (SHapley Additive exPlanations) method is used to calculate the contribution of each input feature to the control decision. SHAP Value Defined as:

[0124] in, For the entire feature set; For features not included A subset of features; To use only a subset of features The model prediction value at that time. SHAP value Reflects characteristics The marginal contribution to the model's predictions is indicated by a positive value for a positive impact and a negative value for a negative impact.

[0125] Rule Trigger Log: Records the triggering status of each constraint rule in the logic layer module, including the triggered rule ID, rule description, trigger time, and related parameter values. For example, if the voltage constraint rule is triggered, the log records "Rule V001: Node 5 voltage 0.93 pu is lower than the lower limit 0.95 pu, trigger time: 2026-03-09 14:23:15".

[0126] Visualization of the reasoning path: The reasoning path of control decision is displayed in the form of a flowchart, including the complete link of state estimation at the perception layer → policy generation at the intuition layer → constraint verification at the logic layer → field prediction at the physical layer → decision coordination and policy selection, and the inputs, outputs and key parameters of each link are marked.

[0127] Confidence Score: Based on the quantification of uncertainty in model predictions, a confidence score is calculated for control decisions. The confidence score is calculated using an ensemble learning method:

[0128] in, The mean of the predictions from multiple models; The standard deviation of the prediction results from multiple models. When When the decision confidence level is high, the decision confidence level is relatively high; when When this occurs, the system issues a low-confidence warning and recommends manual review.

[0129] The complete control cycle of the power system based on the multi-layer cognitive architecture described in this application is approximately 1-2 seconds, which meets the timeliness requirements of real-time control of the power system.

[0130] To meet the requirement of sufficient disclosure in the invention specification, the following typical parameter configurations for a multi-layered cognitive control system are provided for reference during implementation. System parameters are divided into four categories: model parameters, training parameters, operational parameters, and hardware parameters.

[0131] (1) Model parameters: Perception layer extended Kalman filter algorithm parameters: sampling period ms; process noise covariance diagonal elements are pu²; Standard deviation of PMU voltage measurement noise pu; PMU phase angle measurement noise standard deviation rad; Standard deviation of SCADA power measurement noise pu; Mahalanobis distance anomaly detection threshold .

[0132] Intuitive Layer Liquid Neural Network Parameters: Input Dimension Hidden Dimensions Number of layers Attention count Total number of parameters K; Minimum value of liquid time constant Maximum value of liquid time constant Power grid pressure index threshold Voltage deviation weighting Frequency deviation weighting .

[0133] Logic Layer SMT Solver Parameters: Upper Voltage Limit pu; lower voltage limit pu; Rotating Reserve Capacity Ratio Control cycle min; output the number of feasible strategies. Voltage constraint rule priority Priority of thermal limitation rules Economic constraints take priority .

[0134] Physical layer physical information neural operator parameters: input feature dimension (Voltage, phase angle, frequency, ROC / F, load, current); Number of hidden channels Fourier block number Frequency mode number Total number of parameters M; Weight of physical residual loss Boundary condition loss weights .

[0135] Decision-making and coordination layer parameters: Pareto optimal population size Pareto optimization iteration count DQN hidden layer dimensions DQN Hidden Layers Experience pool size Discount factor Target network update frequency Step; Economic weight under normal operating conditions Normal operating condition environmental weights Safety weights under abnormal operating conditions Reliability weight under abnormal operating conditions .

[0136] (2) Training parameter configuration: Liquid Neural Network Training Parameters: Learning Rate Weight decay Batch size Sequence length Number of training rounds Optimizer: AdamW; Learning rate scheduler: CosineAnnealingWarmRestarts (period) Gradient clipping threshold .

[0137] Training parameters for physical information neural operators: learning rate Weight decay Batch size Number of training rounds Optimizer: AdamW; Learning rate scheduler: StepLR (decay factor) decay step size Gradient clipping threshold Data augmentation: Add Gaussian noise (standard deviation) pu).

[0138] DQN training parameters: learning rate Batch size Training steps Optimizer: Adam; -greedy policy initial value ; -Greedy policy final value ; decay steps .

[0139] (3) Running parameter configuration: System control cycle Seconds; State estimation delay of the perception layer ms; intuitionistic reasoning delay ms; logic layer constraint solving delay ms; Physical layer batch inference latency Seconds (approximately 55ms for 1000 scenarios, approximately 380ms for 10000 scenarios); Decision coordination module latency optimization ms; Interpretability report generation delay ms.

[0140] Data acquisition frequency: PMU sampling rate is 50 Hz or 60 Hz; SCADA sampling period is 2-4 seconds; weather forecast update period is 15 minutes.

[0141] Based on the same inventive concept, embodiments of the present invention also provide an intelligent control and decision-making system for power systems based on a multi-layer cognitive architecture; such as Figure 7 As shown, the system includes: The data acquisition and interface module is used to acquire real-time measurement data from the power system. The perception layer module includes an extended Kalman filter unit for receiving real-time measurement data from the power system, performing state estimation using the extended Kalman filter algorithm, and generating a system state vector. In a preferred embodiment, the module also includes an anomaly detection unit for calculating the Mahalanobis distance of the measurement residuals based on the covariance matrix output during state estimation, detecting measurement anomalies based on the Mahalanobis distance, and outputting anomaly alarm information for operators to decide whether to trigger a degraded control mode.

[0142] The intuition layer module is used to generate a set of control strategy candidates based on the system state vector using a liquid neural network; The logic layer module is used to verify the feasibility of the control strategy candidate set based on neural symbolic constraints, and to prioritize the verified control strategies to obtain the candidate control strategies. In a preferred embodiment, this module includes a neural symbolic constraint engine, a rule buffer, and a symbolic reasoning module. The neural symbolic constraint engine is used to express the operating constraints of the power system as first-order logic rules and to verify the logical satisfiability of the control strategy candidate set. The rule buffer stores programmable rules including voltage constraints, thermal limits, reserve capacity constraints, and ramp rate constraints, supporting operators to dynamically modify rule priorities. The symbolic reasoning module is used to rank the candidate strategies based on the verification results of the neural symbolic constraint engine and the rule priorities in the rule buffer, and output the top few feasible strategies.

[0143] The physical layer module is used to quickly solve the AC optimal power flow problem using physical information neural operators for the candidate control strategies, and outputs physical field prediction results including node marginal electricity price, generation scheduling, line power flow, and congestion probability. As a preferred embodiment, this module is equipped with a node marginal electricity price calculation submodule, which is used to calculate the marginal electricity price of each node based on the Lagrange multipliers in the physical field prediction results; and a congestion assessment submodule, which is used to identify congested lines and calculate the congestion probability based on the line power flow in the physical field prediction results.

[0144] The decision coordination layer module is used to select strategies and generate control commands based on the physical field prediction results through multi-objective Pareto optimization and reinforcement learning agents. In this embodiment, this module is configured with an independent control command generation submodule, which is used to convert the control strategy selected by the reinforcement learning agent into specific control commands including generator active power output, reactive power output, and transformer tap position. Optionally, an interpretability output submodule is also configured to generate interpretable reports including SHAP value feature importance analysis, rule trigger logs, inference path visualization, and confidence scores, for operators to view and review.

[0145] In this embodiment, the intelligent control and decision-making system for power systems based on a multi-layer cognitive architecture adopts an asynchronous parallel architecture. The modules of the perception layer, intuition layer, logic layer, and physical layer can operate and be upgraded independently, exhibiting good modularity and scalability. The execution process of each module is the same as the execution principle of each layer in the aforementioned intelligent control and decision-making method for power systems based on a multi-layer cognitive architecture, and therefore will not be repeated here.

[0146] Furthermore, to verify the effectiveness of this application, the following embodiments are provided for verification: This embodiment uses the IEEE 118-node test system and historical operating data of the actual power grid to verify the comprehensive performance of the multi-layer cognitive control system under various operating conditions.

[0147] 1) Experimental configuration parameters Test system: IEEE 118-node system, comprising 118 nodes, 54 generators, and 186 lines. The total system load ranges from 3000 to 5000 MW, with wind and solar power penetration rates of 30% to 50%.

[0148] Experimental scenarios: Three typical scenarios were set up for the experiment: ① Normal economic dispatch scenario (stable load and stable output of renewable energy); ② Rapid response scenario for abnormal operating conditions (simulating line N-1 fault, sudden load increase, etc.); ③ High renewable energy penetration scenario (significant fluctuations in wind power and photovoltaic output).

[0149] Baseline Comparison: To verify the technical advantages of the present invention, three baseline methods were compared experimentally: ① traditional AC-OPF optimization method (based on interior point method solver); ② simplified DC power flow model (DC-OPF); ③ pure data-driven method (neural network without physical constraints).

[0150] Performance metrics: The experiment evaluates system performance from five dimensions: ① Solution time (seconds); ② Power balance error (pu); ③ Voltage prediction error (%); ④ Constraint violation rate (%); ⑤ Control cost ($).

[0151] 2) Experimental Results and Figure Descriptions Scenario 1: Normal economic scheduling Under normal economic scheduling scenarios, the multi-layer cognitive control system of this invention achieves second-level real-time response and low-cost control. For example... Figure 8 As shown, the complete control cycle is approximately 1.2 seconds, with the following delays: perception layer state estimation (45ms), intuition layer inference (15ms), logic layer constraint solving (75ms), physical layer batch inference (580ms, solving for 5 candidate policies), decision coordination and optimization (385ms), and interpretability report generation (100ms). Compared to the traditional AC-OPF method (approximately 12 minutes), this invention improves the solution speed by approximately 600 times.

[0152] In terms of economics, the average control cost of this invention is $4850 / h, which is only 1.5% higher than the traditional AC-OPF method ($4780 / h), but the solution time is significantly reduced; compared with the DC-OPF method ($5120 / h), the cost is reduced by 5.3%. This shows that this invention significantly improves computational efficiency while maintaining near-optimal economics.

[0153] Regarding the physical constraint satisfaction rate, such as Figure 9 As shown, the power balance error of this invention is approximately 0.076 pu (corresponding to approximately 228 MW, accounting for 5.7% of the total load), with a voltage constraint violation rate of 0% and a thermal constraint violation rate of 0%. Compared to the pure data-driven method (power balance error 0.26 pu, voltage constraint violation rate 12%), this invention significantly improves the physical constraint satisfaction rate through the physical residual loss and boundary condition loss of the physical information neural operator.

[0154] Scenario 2: Rapid Response to Abnormal Operating Conditions In abnormal operating conditions (simulating an N-1 fault in line L54-L55), the liquid neural network of this invention achieves rapid response through an adaptive time constant mechanism. For example... Figure 10 As shown, the power grid pressure indicators before the fault occurred... Liquid time constant The system maintains stable output; after a fault occurs, the power grid pressure index... The liquid time constant rapidly rises to 0.68. The voltage deviation was automatically reduced to 0.12, improving the system response speed by approximately 7 times. Within 1.5 seconds of the fault, the system completed fault identification, replanning, and control command issuance, successfully reducing the node voltage deviation from 12% to less than 3%, thus preventing voltage collapse.

[0155] Compared with the traditional fixed time constant method, the present invention improves the response speed under abnormal operating conditions by about 5-9 times, while avoiding the control oscillation problem caused by excessively fast response (the control command change amplitude of the traditional fast response method is about 2.3 times that of the present invention).

[0156] Scenario 3: High Renewable Energy Penetration In scenarios with high renewable energy penetration (wind and solar power accounting for 50%, with power output fluctuations ranging from 20% to 80% of the rated value), the PINO model of this invention has significant advantages over the DC-OPF model in voltage and power flow prediction. Experimental results show that the DC-OPF model has a voltage prediction error of approximately 7.2% and a line power flow prediction error of approximately 15.8%; while the PINO model of this invention has a voltage prediction error of approximately 1.8% and a line power flow prediction error of approximately 2.9%, representing improvements in prediction accuracy of approximately 4 times and 5.4 times, respectively.

[0157] In terms of environmental benefits, this invention employs multi-objective Pareto optimization to prioritize the utilization of renewable energy and reduce carbon emissions while ensuring safety and reliability. During the experiment (24 hours), the carbon emissions of this invention were approximately 32.5 tons of CO2, a reduction of about 15% compared to the traditional single-objective optimization method (38.2 tons of CO2).

[0158] 3) Analysis of experimental results Based on the experimental results of the three scenarios described above, the multi-layer cognitive control system of the present invention exhibits significant advantages in the following aspects: ① Real-time performance: The complete control cycle is 1-2 seconds, which is about 300-600 times faster than the traditional AC-OPF method (more than 10 minutes), meeting the second-level real-time control requirements of power systems.

[0159] ② Physical constraints are met: power balance error 0.076 pu, voltage constraint violation rate 0%, thermal constraint violation rate 0%, which is a significant improvement compared to the pure data-driven method (power balance error 0.26 pu, voltage constraint violation rate 12%).

[0160] ③ High adaptability to renewable energy: In a scenario with 50% renewable energy penetration, the voltage prediction error is 1.8% and the line power flow prediction error is 2.9%, which is about 4-5 times better than the DC-OPF method (voltage error 7.2% and power flow error 15.8%).

[0161] ④ Multi-objective collaborative optimization: Under the premise of ensuring safety and reliability, the economic efficiency is improved by 12-18% and carbon emissions are reduced by 8-15%, which has significant advantages over traditional single-objective optimization methods.

[0162] ⑤ Rapid response to abnormal operating conditions: Through the adaptive time constant mechanism of the liquid neural network, the response speed is improved by 5-9 times under abnormal operating conditions, while avoiding control oscillation.

[0163] ⑥ Interpretability: Generate interpretable reports such as SHAP value feature importance analysis, rule trigger logs, inference path visualization, and confidence scores to meet the traceability requirements for safe production in power systems.

[0164] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0165] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent control and decision-making in power systems based on a multi-layered cognitive architecture, characterized in that, include: The system receives real-time measurement data from the power system through the sensing layer and performs state estimation based on the extended Kalman filter algorithm to generate a system state vector. Based on the system state vector, the intuitive layer uses a liquid neural network to generate a set of control strategy candidates. The feasibility of the control strategy candidate set is verified by the logic layer based on neural symbol constraints, and the candidate control strategies that pass the verification are prioritized to obtain the candidate control strategies. The physical layer utilizes physical information neural operators to quickly solve the AC optimal power flow problem for the candidate control strategies, and outputs physical field prediction results including node marginal electricity price, generation scheduling, line power flow and congestion probability. Furthermore, based on the prediction results of the physical field, the decision coordination layer utilizes multi-objective Pareto optimization and reinforcement learning agents to select policies and generate control commands.

2. The intelligent control and decision-making method for power systems based on a multi-layer cognitive architecture as described in claim 1, characterized in that, When performing state estimation based on the extended Kalman filter algorithm, a covariance matrix is ​​output synchronously. This covariance matrix is ​​used to calculate the Mahalanobis distance of the measurement residuals, and to generate an alarm for abnormal measurement data based on the Mahalanobis distance and a preset threshold.

3. The intelligent control and decision-making method for power systems based on a multi-layer cognitive architecture as described in claim 1, characterized in that, The liquid neural network includes multiple liquid time constant units, and the time constant of each liquid time constant unit is dynamically adjusted according to the power grid pressure index; the power grid pressure index is obtained by a weighted combination of node voltage deviation and system frequency deviation.

4. The intelligent control and decision-making method for power systems based on a multi-layer cognitive architecture as described in claim 1, characterized in that, The neural symbol constraints are first-order logic rules, including one or more of the following: voltage constraints, thermal limit constraints, reserve capacity constraints, ramp rate constraints, and power balance constraints.

5. The intelligent control and decision-making method for power systems based on a multi-layer cognitive architecture as described in claim 1, characterized in that, The feasibility of the candidate control strategies is verified, and the verified control strategies are prioritized, including: The neural symbolic constraints are converted into SMT formulas, and the SMT solver is used to verify whether the candidate control strategy satisfies all the constraint rules. For the candidate control strategies that pass the verification, a comprehensive score is calculated based on the priority of the constraint rules, and the strategies are then ranked according to the priority of the comprehensive scores.

6. The intelligent control and decision-making method for power systems based on a multi-layer cognitive architecture as described in claim 1, characterized in that, The physical information neural operator, from input to output, includes an encoder, a Fourier block layer, and three parallel prediction heads connected to it. The three parallel prediction heads are used to output the node marginal electricity price, generation scheduling, and line power flow, respectively.

7. The intelligent control and decision-making method for power systems based on a multi-layer cognitive architecture as described in claim 6, characterized in that, The physical layer is also used to identify congested lines and calculate congestion probabilities based on the line power flow predicted by the neural operator based on the physical information.

8. The intelligent control and decision-making method for power systems based on a multi-layer cognitive architecture as described in claim 1 or 6, characterized in that, The physical information neural operator is trained using a joint loss function that includes data fidelity loss, physical residual loss, and boundary condition loss. The physical residual loss includes the power balance equation residual sub-loss, voltage constraint residual sub-loss, and power flow consistency constraint residual loss.

9. The intelligent control and decision-making method for power systems based on a multi-layer cognitive architecture as described in claim 1, characterized in that, The multi-objective Pareto optimization includes five optimization objectives: safety, reliability, economy, environment, and response speed. A non-dominated sorting genetic algorithm is used to generate the Pareto front solution set. The reinforcement learning agent is used to learn to select the optimal control strategy from the Pareto front solution set under different operating conditions based on historical operating data and control effect feedback.

10. A power system intelligent control and decision-making system based on a multi-layer cognitive architecture, characterized in that, The method for intelligent control and decision-making of power systems based on a multi-layer cognitive architecture as described in any one of claims 1-9 includes: The perception layer module is used to receive real-time measurement data of the power system and perform state estimation based on the extended Kalman filter algorithm to generate a system state vector. The intuition layer module is used to generate a set of control strategy candidates based on the system state vector using a liquid neural network; The logic layer module is used to perform feasibility verification on the candidate set of control strategies based on neural symbol constraints, and to prioritize the candidate control strategies that pass the verification to obtain the candidate control strategies. The physical layer module is used to quickly solve the AC optimal power flow problem for the candidate control strategy using physical information neural operators, and outputs physical field prediction results including node marginal electricity price, generation scheduling, line power flow and congestion probability. The decision coordination layer module is used to select policies and generate control commands based on the physical field prediction results, using multi-objective Pareto optimization and reinforcement learning agents.

Images