Power system operation simulation method based on physical information neural network

Abstract

The application discloses a kind of power system operation simulation methods based on physical information neural network, comprising the following steps: obtaining the operation monitoring data under each operation section of power system, constructs operation state data set;According to historical operation data and transient simulation data, construct operation simulation training sample, set the state space, action space and constraint condition of operation simulation;Build operation simulation model based on physical information neural network, combine prior knowledge and embed operation constraint, carry out operation simulation machine learning training to model;Set operation simulation correction evaluation index, carry out fitting evaluation to search result, adjust and update network parameter, obtain pre-training model;According to pre-training model and real-time monitoring data, carry out operation simulation deduction, and evaluate power system state risk.Improves the solving efficiency and accuracy of complex operation simulation of power system, guarantees the accurate perception of operation trend, assists the decision optimization of dispatching scheme, and can be widely applied to power grid dispatching.

Description

Technical Field

[0001] This invention relates to the field of power grid operation and dispatching technology, and more specifically, to a power system operation simulation method based on a physical information neural network. Background Technology

[0002] With the increasing demands for observable and measurable power system operations, efficient and accurate simulation methods are urgently needed for comprehensive perception and decision support of power system operation status. Current power system operation simulations primarily rely on decoupled simulation calculations across multiple time scales, exhibiting a high degree of dependence on simulation modeling and facing difficulties in sequentially solving time-series states. However, with the widespread deployment of massive terminal measurement equipment, power system operation simulation problems can be addressed by combining data-driven machine learning methods to solve the problems of low solution efficiency, insufficient accuracy in multi-time-scale simulations, and inadequate scenario adaptability in traditional simulation and optimization methods. This allows for efficient solution of system operation status deductions and effective applications.

[0003] Existing research is mainly based on simulation models and optimization algorithms. For example, Chinese patent application CN107834543A proposes a power system operation simulation method based on two-stage mixed integer programming. It uses a two-stage optimization algorithm to solve medium- and long-term operation simulation problems. However, under the condition of increasing power system scenario scale, it is difficult to guarantee computational efficiency and algorithm solvability. It lacks sequential solution of multi-timescale coupled processes, and the feasibility of online application of the method needs to be improved.

[0004] Therefore, how to improve power system operation simulation methods to meet the timeliness and accuracy requirements of online system operation solutions has become an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of power grid recovery control and efficient reconfiguration after a power grid fault, and to provide a power system operation simulation method based on physical information neural networks, thereby improving the feasibility and timeliness of solving power grid reconfiguration recovery control.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A power system operation simulation method based on physical information neural networks includes the following steps:

[0008] S1. Obtain operational monitoring data at various operating sections of the power system and construct a power system operational status dataset;

[0009] S2. Based on the historical operating data and transient simulation data of the power system, construct operating simulation training samples and set the state space, action space and constraints of the operating simulation.

[0010] S3. Construct a power system operation simulation model based on physical information neural network, combine prior knowledge and embed operational constraints, and train the model for operation simulation machine learning;

[0011] S4. Set up the simulation correction evaluation index, evaluate the physical information neural network search results, adjust and update the network parameters, and obtain the pre-trained simulation model.

[0012] S5. During the online operation phase, based on the pre-trained operation simulation model and real-time monitoring data of the power system, power system operation simulation is performed to assess the power system status risk.

[0013] Preferably, in step S1, the operating status dataset includes power flow data, voltage curve information, and power angle curve information. The power grid state matrix and the operating feature matrix constitute the operating status dataset input vector, represented as:

[0014] F = [G, N]

[0015] G ij =[δ ij ,r ij ,x ij ]

[0016] N i =[U i ,P ij Q ij ,θ i ,P gen,i Q gen,i ]

[0017] In the formula, F is the input matrix of the operating state dataset; G is the power grid state matrix; N is the operating feature matrix; G ij Let δ be the path state matrix from node i to node j; ij This represents the switching state of the line flowing from node i to node j; the value is 1 when the line is closed and 0 otherwise. ij x is the resistance of the line flowing from node i to node j; ij N is the reactance of the line flowing from node i to node j; i The characteristic matrix for running at power grid node i; U i P represents the voltage value at node i. ij Q represents the active power of the line flowing from node i to node j; ij θ represents the reactive power of the line flowing from node i to node j; i P represents the phase angle value at node i; gen,i Q represents the active power of the power source connected to node i in the power grid; gen,i Connect the reactive power of the power source to grid node i.

[0018] Preferably, in step S2, the historical operating sequence data of the power system and the transient simulation data are randomly sampled to form the operating simulation training sample. The transient simulation generates simulation data using Python software or PSD-BPA software.

[0019] Preferably, in step S2, the state space is the input matrix of the operating state dataset, the action space is the power output of the generator connected to the node, and the constraints include power balance constraints, node voltage constraints, line power constraints, and unit ramping constraints.

[0020] Power balance constraint means that the active power generated by generator units in each section of the power system is balanced with the total power demand of the node load. Node voltage constraint means that the voltage of each node in each section of the power system meets the upper and lower voltage limits. Line power constraint means that the power of each line in the power system meets the line capacity range. Generator unit ramping constraint means that the regulation of increasing and decreasing the output of generator units meets the power adjustment constraint.

[0021] The power balance constraint is expressed as:

[0022]

[0023] In the formula, P k,gen,t P represents the active power of generator unit k in the system at each cross section at time t; i,load,t N represents the load power of node i in the system at each cross section at time t; gen N represents the total number of generator sets. bus This represents the total number of power grid nodes.

[0024] The node voltage constraint is expressed as:

[0025] U i min ≤U i ≤U i max

[0026] In the formula, U i max U represents the upper limit of the voltage at node i; i min U represents the lower limit of the voltage at node i; i The voltage value at node i;

[0027] Line power constraints are expressed as follows:

[0028]

[0029] In the formula, P ij Q represents the active power of the line flowing from node i to node j; ij S represents the reactive power of the line flowing from node i to node j; i,j,maxThe rated capacity of the line flowing from node i to node j;

[0030] The unit ramp-up constraint is expressed as:

[0031] P gen,k,t -P gen,k,t-1 ≤P gen,up

[0032] P gen,k,t-1 -P gen,k,t ≤P gen,down

[0033] In the formula, P gen,k,t-1 P represents the output power of generator unit k at time t-1. gen,k,t P represents the output power of generator unit k at time t. gen,up To increase the upper limit of output power within a single section of the unit, P gen,down To reduce the lower limit of the output power within a single section of the unit.

[0034] Preferably, step S3 includes:

[0035] S3.1. Construct a mechanism model and labeled training sample set for power system operation simulation, transform the operation simulation optimization problem into KKT conditional form, and extract the undetermined optimization parameters;

[0036] S3.2. Set the training data points and error configuration points of the label training sample set, write the KKT form constraints and mechanism equations into the loss function of the physical information neural network, and construct an improved loss function that considers physical information and data errors.

[0037] S3.3. Evaluate the minimum value of the improved loss function, extract the corresponding training samples, and set up an experience replay pool to store the training samples and neural network parameter set;

[0038] S3.4. During the offline training phase of the simulation model, perform sampling training and update the error factor until the training meets the constraints and reaches convergence.

[0039] Preferably, in step S3.1, the mechanism model is a time-series production model with optimal economic cost as the optimization objective, taking into account source-load balance constraints, ramp-up constraints, output constraints, and grid operation constraints. Based on the KKT conditions, the mechanism model is converted into an equation-constrained and optimization model with undetermined parameters, expressed as:

[0040]

[0041] In the formula, Let P be the power generation cost of each unit at time t; k,gen P is the output power of generator set k; k,down P represents the lower limit of the output power of generator set k;k,up This represents the upper limit of the output power of generator set k.

[0042] Preferably, in step S3.2, the labeled training sample set is divided into N t Data points and N c For each configuration point without a label indicating the unit's output, the constraints in KKT form and the fitting error of the mechanism equation are incorporated into the loss function of the physical information neural network. The improved loss function is expressed as follows:

[0043]

[0044] MAE KKT =MSE PDE +MSE BC

[0045] In the formula, MAE D Consistency loss error in data prediction; MAE KKT Correction errors for constraints and mechanistic equations under KKT boundary conditions, including the running mechanism equation matching loss (MSE). PDE Matching loss MSE with boundary conditions BC .

[0046] Preferably, in step S3.3, the improved loss function is evaluated, and the training samples and neural network hyperparameter set corresponding to the minimum loss value points in offline training are identified. During the training process, the sigmoid function is used as the network activation function in the experience replay pool.

[0047] Preferably, in step S4, the simulation correction evaluation index includes an evaluation penalty item, which is calculated from the violation penalty constraint evaluation table and expressed as follows:

[0048] ∑V pen =V pen,pf +V pen,op +V pen,gen

[0049] In the formula, V pen,pf Penalty for source-load balance violation; V pen,op Penalties for violating operational constraints; V pen,gen Constraints and penalties for frequent start-ups and shutdowns of generating units;

[0050] The evaluation index for simulation calibration is expressed as follows:

[0051]

[0052] In the formula, T is the timing step size; f(v pre f(v) represents the evaluation value of the time-series prediction state result of the physical information neural network; actV represents the evaluation value of the actual time-series sample state results; pen,c γ is the penalty value for violating constraint c; γ is a constant coefficient.

[0053] Preferably, in step S5, the pre-trained operation simulation model is invoked during the online operation phase. The model is fine-tuned by accessing real-time monitoring data of the power system, and the online solution of the operation simulation is performed to assess the power system state risks, including cross-section over-limit risk, load loss risk, and voltage over-limit risk.

[0054] Cross-section exceeding limit risk R s Represented as:

[0055]

[0056] In the formula, K represents the number of simulated states; p(E) k (E) represents the power system state. k The probability of S; s (E k (E) represents the power system state. k The severity of consequences of exceeding the limit at the following cross-section; NL is the total number of operating cross-sections of the power system; P l Active power flowing through section l; P l，max The maximum active power at section l;

[0057] Risk of loss of load R C Represented as:

[0058]

[0059] In the formula, S C (E k (E) represents the power system state. k Severity of the consequences of loss of load; P i，cut N represents the load shedding amount at node i; bus The number of nodes;

[0060] Voltage over-limit risk R v Represented as:

[0061]

[0062] In the formula, S v (E k (E) represents the power system state. k Severity of consequences of exceeding voltage limits; V i Let i be the per-unit voltage value of node i.

[0063] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0064] 1. The method of this invention is used to simulate the operation of power systems, which reduces the calculation time for solving the real-time operating status. It adopts a data-driven machine learning algorithm to achieve accurate solution of the operation simulation under small sample conditions, improves the efficiency and optimization of the online state solution and inference function of power systems, and has the adaptability to solve the operation in massive scenarios.

[0065] 2. This invention employs a physical information neural network to simulate and solve power system operation problems. It embeds the mechanistic equations and constraint equations into the neural network architecture to achieve adaptive training and learning of the machine model, ensuring the accuracy and effectiveness of the simulation results. Through offline training and data storage, it avoids the complex simulation calculations and verifications in the online stage, effectively improving the efficiency of online operation simulation applications. Attached Figure Description

[0066] Figure 1 This is a flowchart of the power system operation simulation method based on physical information neural network of the present invention.

[0067] Figure 2 This is a flowchart of the training process for a power system operation simulation model based on a physical information neural network.

[0068] Figure 3 This is a schematic diagram of a power system operation simulation model based on a physical information neural network.

[0069] Figure 4 This is a schematic diagram of a power system operation simulation test based on IEEE 39 nodes.

[0070] Figure 5 A schematic diagram showing the results of power system operation simulation tests based on the IEEE 39-node architecture.

[0071] Figure 6 This is a diagram of the power system operation simulation system based on physical information neural networks. Detailed Implementation

[0072] The power system operation simulation method based on physical information neural network of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0073] Please see Figure 1 This invention discloses a power system operation simulation method based on a physical information neural network, comprising the following steps:

[0074] S1. Obtain operational monitoring data at various operating sections of the power system and construct a power system operational status dataset;

[0075] S2. Based on the historical operating data and transient simulation data of the power system, construct operating simulation training samples and set the state space, action space and constraints of the operating simulation.

[0076] S3. Construct a power system operation simulation model based on physical information neural network, combine prior knowledge and embed operational constraints, and train the model for operation simulation machine learning;

[0077] S4. Set up the simulation correction evaluation index, evaluate the physical information neural network search results, adjust and update the network parameters, and obtain the pre-trained simulation model.

[0078] S5. During the online operation phase, based on the pre-trained operation simulation model and real-time monitoring data of the power system, power system operation simulation is performed to assess the power system status risk.

[0079] Specifically, in step S1, the power system operating status dataset includes power flow data, voltage curve information, and power angle curve information. The power grid state matrix and the operating feature matrix constitute the input vector, represented as:

[0080] F = [G, N]

[0081] G ij =[δ ij ,r ij ,x ij ]

[0082] N i =[U i ,P ij Q ij ,θ i ,P gen,i Q gen,i ]

[0083] In the formula, F is the input matrix of the operating state dataset; G is the power grid state matrix; N is the operating feature matrix; G ij Let δ be the path state matrix from node i to node j; ij This represents the switching state of the line flowing from node i to node j; the value is 1 when the line is closed and 0 otherwise. ij x is the resistance of the line flowing from node i to node j; ij N is the reactance of the line flowing from node i to node j; i The characteristic matrix for running at power grid node i; U i P represents the voltage value at node i. ij Q represents the active power of the line flowing from node i to node j; ij θ represents the reactive power of the line flowing from node i to node j; i P represents the phase angle value at node i; gen,i Q represents the active power of the power source connected to node i in the power grid; gen,i Connect the reactive power of the power source to grid node i.

[0084] Specifically, in step S2, historical operating sequence data of the power system and transient simulation data are randomly sampled to form operating simulation training samples. The transient simulation generates simulation data using Python software or PSD-BPA software.

[0085] In step S2, the state space is the input matrix of the operating state dataset, the action space is the power output of the generators connected to the nodes, and the constraints include power balance constraints, node voltage constraints, line power constraints, and unit ramping constraints.

[0086] Power balance constraint means that the active power generated by generators in each section of the power system is balanced with the total power demand of the node load. Node voltage constraint means that the voltage of each node in each section of the power system meets the upper and lower voltage limits. Line power constraint means that the power of each line in the power system meets the line capacity range. Generator ramp constraint means that the regulation of increasing and decreasing the output of generators meets the power adjustment constraint.

[0087] The power balance constraint is expressed as:

[0088]

[0089] In the formula, P k,gen,t P represents the active power of generator unit k in the system at each cross section at time t; i,load,t N represents the load power of node i in the system at each cross section at time t; gen N represents the total number of generator sets. bus This represents the total number of nodes in the power grid.

[0090] The node voltage constraint is expressed as:

[0091] U i min ≤U i ≤U i max

[0092] In the formula, U i max U represents the upper limit of the voltage at node i; i min This represents the lower limit of the voltage at node i.

[0093] Line power constraints are expressed as follows:

[0094]

[0095] In the formula, P ij Q represents the active power of the line flowing from node i to node j; ij S represents the reactive power of the line flowing from node i to node j; i,j,maxThe rated capacity of the line flowing from node i to node j.

[0096] The unit ramp-up constraint is expressed as:

[0097] P gen,k,t -P gen,k,t-1 ≤P gen,up

[0098] P gen,k,t-1 -P gen,k,t ≤P gen,down

[0099] In the formula, P gen,k,t-1 P represents the output power of generator unit k at time t-1. gen,k,t P represents the output power of generator unit k at time t. gen,up To increase the upper limit of output power within a single section of the unit, P gen,down To reduce the lower limit of the output power within a single section of the unit.

[0100] like Figure 2 As shown, step S3 includes:

[0101] S3.1. Construct a mechanism model and labeled training sample set for power system operation simulation, transform the operation simulation optimization problem into KKT conditional form, and extract the undetermined optimization parameters;

[0102] S3.2. Set the training data points and error configuration points of the label training sample set, write the KKT form constraints and mechanism equations into the loss function of the physical information neural network, and construct an improved loss function that considers physical information and data errors.

[0103] S3.3. Evaluate the minimum value of the improved loss function, extract the corresponding training samples, and set up an experience replay pool to store the training samples and neural network parameter set;

[0104] S3.4. During the offline training phase of the simulation model, perform sampling training and update the error factor until the training meets the constraints and reaches convergence.

[0105] In step S3.1, the mechanism model is a time-series production model with optimal economic cost as the optimization objective, taking into account source-load balance constraints, ramp-up constraints, output constraints, and grid operation constraints. Based on the KKT conditions, the mechanism model is transformed into an equality constraint and optimization model with undetermined parameters, expressed as:

[0106]

[0107] In the formula, Let P be the power generation cost of each unit at time t; k,gen P is the output power of generator set k; k,downP represents the lower limit of the output power of generator set k; k,up This represents the upper limit of the output power of generator set k.

[0108] like Figure 3 As shown, in step S3.2, the labeled training sample set is divided into N t Data points and N c For each configuration point without a label indicating the unit's output, the constraints in KKT form and the fitting error of the mechanism equation are incorporated into the loss function of the physical information neural network. The improved loss function is expressed as follows:

[0109]

[0110] MAE KKT =MSE PDE +MSE BC

[0111] In the formula, MAE D For the consistency loss error of data prediction, MAE KKT Correction errors for constraints and mechanistic equations under KKT boundary conditions, including the running mechanism equation matching loss (MSE). PDE Matching loss MSE with boundary conditions BC .

[0112] In step S3.3, the improved loss function is evaluated, and the training samples and neural network hyperparameter set corresponding to the minimum loss value points in offline training are identified. During training, the empirical replay pool uses the sigmoid function as the network activation function, expressed as:

[0113] Sig(x)=(1+e -x ) -1 .

[0114] Specifically, in step S4, the simulation correction evaluation index includes an evaluation penalty item, which is calculated from the violation penalty constraint evaluation table and is expressed as follows:

[0115] ∑V pen =V pen,pf +V pen,op +V pen,gen

[0116] In the formula, V pen,pf Penalty for source-load balance violation; V pen,op Penalties for violating operational constraints; V pen,gen Constraints and penalties for frequent start-ups and shutdowns of generating units;

[0117] The evaluation index for simulation calibration is expressed as follows:

[0118]

[0119] In the formula, T is the timing step size; f(v pre f(v) represents the evaluation value of the time-series prediction state result of the physical information neural network; act V represents the evaluation value of the actual time-series sample state results; pen,c γ is the penalty value for violating constraint c; γ is a constant coefficient.

[0120] Based on the above evaluation results, the physical information neural network parameters are adjusted for optimization, and the network parameters of the running simulation model are updated. A pre-trained running simulation model is formed through offline training.

[0121] Specifically, in step S5, the pre-trained operation simulation model is invoked during the online operation phase. The model is fine-tuned by accessing real-time monitoring data of the power system, and the online solution of the operation simulation is performed to assess the power system state risks, including cross-section over-limit risks, load loss risks, and voltage over-limit risks.

[0122] Cross-section exceeding limit risk R s Represented as:

[0123]

[0124] In the formula, K represents the number of simulated states; p(E) k (E) represents the power system state. k The probability of S; s (E k (E) represents the power system state. k The severity of consequences of exceeding the limit at the following cross-section; NL is the total number of operating cross-sections of the power system; P l Active power flowing through section l; P l，max The maximum active power at section l;

[0125] Risk of loss of load R C Represented as:

[0126]

[0127] In the formula, S C (E k (E) represents the power system state. k Severity of the consequences of loss of load; P i，cut N represents the load shedding amount at node i; bus The number of nodes;

[0128] Voltage over-limit risk R v Represented as:

[0129]

[0130] In the formula, S v (E k(E) represents the power system state. k Severity of consequences of exceeding voltage limits; V i Let i be the per-unit voltage value of node i.

[0131] This embodiment uses the IEEE 39-node system for power system operation simulation testing. A schematic diagram of the IEEE 39-node system is shown below. Figure 4 As shown, the test parameters included: 39 nodes, 3 fully connected layers in the neural network, a learning rate of 0.0001, 10,000 training samples, and a convergence threshold of 0.01. The number of training epochs for convergence during the offline training phase was 4238.

[0132] Test results are as follows Figure 5 As shown, the proposed physical information neural network algorithm solves the daily operation simulation problem of the power system with a time resolution of 5 minutes. It solves the unit output curves at each time section within a day. The test sample size is 200, and the average accuracy is 96.3%, which is 34.1% higher than the prediction accuracy of traditional neural networks. The total calculation time for the daily test is 26.15s, and the average time for a single-step operation simulation is 0.091s.

[0133] The test results of the embodiments show that the power system operation simulation method based on physical information neural network provided by the present invention can quickly solve the time-series state of the power system, and the solution accuracy reaches the approximate optimal solution of the optimization solution. Compared with the traditional neural network prediction algorithm, the accuracy is significantly improved, and the solution efficiency is significantly improved compared with the optimization algorithm. It meets the timeliness and optimization of dispatching business, and can quickly output key state indicators and risk assessments such as unit curves, node voltage curves and cross-sectional power flow. It allows dispatchers to intuitively obtain the system operation status, identify the dynamic changes of the system, and assist in the optimization of operation and dispatching decisions.

[0134] like Figure 6 As shown, the present invention also discloses a power system operation simulation system based on a physical information neural network, comprising:

[0135] The data acquisition module is used to acquire real-time monitoring data of system operation; the data storage module is used to store time-series historical data and offline training samples; the offline training module is used to run the simulated intelligent agent model construction and offline training, and to store and update the intelligent agent network parameters; the online inference module is used to call real-time data and run the simulated intelligent agent output state inference results in the online stage.

[0136] The data acquisition module interfaces with the operation monitoring system, equipment measurement system, and external information system to acquire real-time operation status data. The offline training module provides configurable network model construction, physical information neural network adjustment, and model import, integrates intelligent agents for offline training, and is connected to the data storage module for data retrieval and storage. The data storage module is connected to the system storage database, integrates multi-source data from the data acquisition module, performs offline sample storage, and is connected to the offline training module to provide sample data verification. The online simulation module calls the running simulation intelligent agent and accesses real-time data to perform online simulation simulation, outputting power system risk assessment results and displaying them visually.

[0137] The present invention also discloses an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the power system operation simulation method based on a physical information neural network as described in any of the preceding claims. One or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module.

[0138] Various parts of this invention can be implemented using hardware, software, firmware, or a combination thereof. The methods and steps described above are stored in memory and executed or implemented by system software or firmware driven by instructions issued by a processor, for example, using discrete logic circuits with logic gates for implementing logical functions on data signals, application-specific integrated circuits with suitable combinational logic gates, programmable gate arrays (PGAs), etc.

[0139] This invention also discloses a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the power system operation simulation method based on a physical information neural network as described in any of the preceding claims. The storage medium, including portable hard drives, read-only memory (ROM), random access memory (RAM), and other media suitable for storing program code, is used to establish an embedded system database. This allows for editing, decoding, and other electronic means to obtain system data and programs, which are then stored in the device memory.

[0140] The above description is a detailed description of the preferred embodiments of the present invention. However, the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modifications made under the technical spirit disclosed in the present invention should fall within the patent scope covered by the present invention.

Claims

1. A method for power system operation simulation based on physical information neural network, characterized in that, The method comprises the following steps: S1. Obtain operation monitoring data of each operation section of the power system, and construct a power system operation state data set; S2. According to the historical operation data and transient simulation data of the power system, construct an operation simulation training sample, and set a state space, an action space and a constraint condition of the operation simulation; S3. Construct a power system operation simulation model based on a physical information neural network, combine prior knowledge and embed operation constraints, and perform operation simulation machine learning training on the model; S4. Set an operation simulation correction evaluation index, fit and evaluate the search result of the physical information neural network, adjust and update the network parameters, and obtain a pre-trained operation simulation model; S5. In an online operation stage, according to the pre-trained operation simulation model and real-time monitoring data of the power system, perform power system operation simulation deduction, and evaluate the state risk of the power system; Step S3 comprises: S3.

1. Construct a mechanism model of the power system operation simulation and a label training sample set, convert the operation simulation optimization problem into a KKT condition form, and extract optimization undetermined parameters; S3.

2. Set the training data points and error configuration points of the label training sample set, write the constraint conditions in the KKT form and the mechanism equation into a loss function of the physical information neural network, and construct an improved loss function considering physical information and data error; S3.

3. Evaluate the minimum point of the improved loss function value, extract the corresponding training sample, and set an experience replay pool to store the training sample and the neural network parameter set; S3.

4. In the offline training stage of the operation simulation model, perform sampling training and update the error factor until the training meets the constraint condition and reaches convergence; In step S3.1, the mechanism model is a time sequence production model taking the optimal economic cost as an optimization target, considering source and load balance constraints, ramp constraints, output constraints and power grid operation constraints, the mechanism model is converted into an equation constraint and an optimization model containing undetermined parameters according to the KKT condition, and is expressed as: s.t.P k,down ≤P k,gen ≤P k,up P gen,k,t -P gen,k,t-1 ≤P gen,up P gen,k,t-1 -P gen,k,t ≤P gen,down In the formula, is the generating cost of each unit at time t; P k,gen is the output power of generator unit k; P k,down is the lower limit of the output power of generator unit k; P k,up is the upper limit of the output power of generator unit k; P k,gen,t Pki(t) represents the active power of generator k in the system at time t under each section; i,load,t Pli(t) represents the load power of node i in the system at time t under each section; gen N represents the total number of generators; bus N represents the total number of nodes in the grid; ij Pij represents the active power of the line from node i to node j; ij Qij represents the reactive power of the line from node i to node j; i,j,max Sij represents the rated capacity of the line from node i to node j; gen,k,t-1 Pki(t-1) represents the output power of generator k at time t-1; gen,k,t Pki(t) represents the output power of generator k at time t; gen,up Pki represents the upper limit of the increased output power within the single section of the unit; gen,down Pki represents the lower limit of the reduced output power within the single section of the unit; In step S3.2, the label training sample set is divided into N t data points and N c configuration points, the configuration points have no unit output result label, the constraint condition in KKT form and the fitting error of the mechanism equation are written into the loss function of the physical information neural network, and the improved loss function is represented as: MAE KKT = MSE PDE + MSE BC where MAE D is the consistency loss error of data prediction; MAE KKT is the constraint and mechanism equation correction error under the KKT boundary condition, including the operation mechanism equation matching loss MSE PDE and the boundary condition matching loss MSE BC . 2.The physical information neural network based power system operation simulation method of claim 1, wherein, In step S1, the operation state data set includes power flow data information, voltage curve information and power angle curve information, a power grid state matrix and an operation characteristic matrix constitute an operation state data set input vector, and are expressed as: F=[G,N] G ij = [δ ij , r ij , x ij ] N i = [U i , P ij , Q ij , θ i , P gen,i , Q gen,i ] In the formula, F is an operation state data set input matrix; G is a power grid state matrix; and N is an operation characteristic matrix; G ij is the line state matrix for the flow from node i to node j; δ ij is the switch state of the line from node i to node j, with value 1 if the line is closed and 0 otherwise; r ij is the resistance of the line from node i to node j; x ij is the reactance of the line from node i to node j; N i is the operating characteristic matrix of the grid node i; U i Vi voltage value for node i; θ i is the phase angle value for node i; P gen,i is the active power connected to the grid node i; Q gen,i is the reactive power connected to the grid node i. 3.The physical information neural network based power system operation simulation method of claim 1, wherein, In step S2, the power system historical operation time sequence data and transient simulation simulation data are randomly sampled to form an operation simulation training sample, and the transient simulation generates simulation data through a python software or a PSD-BPA software. 4.The physical information neural network based power system operation simulation method of claim 1, wherein, In step S2, the state space is an operation state data set input matrix, the action space is the power output of the node connected generator, and the constraint conditions include power balance constraints, node voltage constraints, line power constraints and unit ramp constraints; The power balance constraint indicates that the active power generated by the generator set in each section of the power system is balanced with the total power demand of the node load, the node voltage constraint indicates that the voltage of each node in each section of the power system satisfies the upper and lower voltage range, the line power constraint indicates that the power of each line of the power system satisfies the line capacity range, and the unit ramp constraint indicates that the increase and decrease of the generator output satisfies the power adjustment constraint. The power balance constraint is expressed as: The node voltage constraint is expressed as: U i min ≤U i ≤U i max where U i max represents the upper limit of the voltage at node i; U i min denotes the lower voltage limit of node i; U i is the voltage value of node i; The line power constraint is expressed as: The unit ramping constraint is expressed as: 5.The physical information neural network based power system operation simulation method of claim 1, wherein, In step S3.3, the improved loss function is evaluated, and the training sample corresponding to the minimum value point of the loss value and the neural network hyperparameter set in the offline training are identified, and the experience replay pool in the training process adopts a sigmoid function as the network activation function. 6.The physical information neural network based power system operation simulation method of claim 1, wherein, In step S4, the operation simulation correction evaluation index contains an evaluation penalty term, and the penalty term is calculated by the constraint violation penalty evaluation table and is expressed as: ∑V pen = V pen,pf + V pen,op + V pen,gen wherein V pen,pf is a source load balancing violation penalty; V pen,op is a running constraint violation penalty; V pen,gen is a unit frequent start-stop constraint penalty; The operation simulation correction evaluation index is expressed as: In the formula, T is a time sequence running step; f(v pre ) is a physical information neural network time sequence prediction state result evaluation value; f(v act ) is an actual time sequence sample state result evaluation value; V pen,c is a penalty value for violating the constraint c; and γ is a constant coefficient.

7. The method of claim 6, wherein the physical information neural network-based power system operation simulation method is characterized by, In step S5, the pre-trained operation simulation model is called in the online operation stage, the model is fine-tuned through real-time monitoring data access of the power system, the operation simulation online solution state deduction is executed, the power system state risk is evaluated, including the risk of exceeding the limit, the risk of losing load and the risk of voltage exceeding the limit; Crossing section risk R s is represented as: where K is the number of analog states; p(E k ) is the probability of power system state E k ; S s (E k ) is the severity of the consequences of the violation of the section limit under the power system state E k ; NL is the total number of power system operating sections; P l is the active power flowing through section l; P l，max is the maximum active power of section l; Loss of load risk R C is represented as: where S C (E k ) is the severity of loss-of-load consequences for the power system state E k ; P i，cut is the amount of load shedding at node i; and N bus is the number of nodes. Voltage excursion risk R v is represented as: where S v (E k ) is the severity of the voltage excursion consequence for the power system state E k ; and V i is the voltage per unit at node i.

Images