Power grid section limit evaluation method and system

Abstract

The invention discloses a power grid section limit evaluation method and system, and the method comprises the steps: obtaining typical operation mode data of a power grid, and generating a sample set; establishing a safety evaluation model, calculating a sample score based on section transmission power and a safety constraint penalty index, and dividing an experience pool; constructing a deep reinforcement learning agent, wherein the action space of the deep reinforcement learning agent adopts a unit sensitivity sorting strategy, and a reward function synchronizes security constraints and section power indexes; after pre-training through a priority experience pool, introducing a new energy output constraint and a wind curtailment and light curtailment penalty term to carry out secondary training; and finally fusing a strategy distribution entropy optimization mechanism, and freezing network shallow layer parameters to complete migration training. The problem of low training efficiency is solved through combination of a priority experience pool screening mechanism and dynamic weight distribution; through collaborative design of strategy distribution entropy optimization and shallow parameter freezing, the problem of poor model migration adaptability is solved, and accurate evaluation of the section limit under the high-proportion new energy power grid is realized.

Description

Technical Field

[0001] The present invention relates to a power system, and in particular to a power grid section limit assessment method and system. Background Art

[0002] As the load rate of power transmission sections continues to rise, the pressure to ensure power supply is increasing, which places higher demands on the refinement and leanness of power system security and stability analysis. At the same time, the installed capacity of renewable energy power generation, represented by wind power and photovoltaics, has grown rapidly, and its proportion in the power system has continued to increase, making the boundary conditions that need to be considered in the grid section limit assessment more complex and diverse. Traditional grid section limit assessment methods mainly rely on deterministic power flow calculations and static safety analysis, and perform calculations based on preset operating modes and fixed boundary conditions. This method often exhibits obvious defects such as insufficient adaptability and limited calculation accuracy when faced with fluctuations in renewable energy output, random changes in load, and complex and changeable operating scenarios. In recent years, deep reinforcement learning, as an important algorithm in the field of artificial intelligence, has shown great potential in the field of power system security and stability analysis due to its outstanding high-dimensional data processing capabilities, rapid feature extraction, and policy learning advantages. However, existing methods for assessing power grid section limits based on deep reinforcement learning (DRL) still have significant shortcomings. First, existing models fail to adequately account for random variations in operating scenarios, resulting in poor model stability in practical applications. Second, existing methods perform poorly in terms of state space exploration efficiency, model training speed, and cross-scenario migration capabilities, making them unable to meet the stringent real-time, accuracy, and adaptability requirements of modern power systems. These technical shortcomings severely restrict the practical application of DRL in power grid security analysis, making it ineffective in addressing the dynamic changes and uncertainties faced by power grids with a high proportion of renewable energy. Summary of the Invention

[0003] Purpose of the invention: The purpose of the present invention is to provide a power grid section limit assessment method that can adapt to new energy fluctuations, has high-precision calculation and strong cross-scenario migration capabilities. Another purpose of the present invention is to provide a power grid section limit assessment system.

[0004] Technical solution: The power grid section limit assessment method of the present invention includes the following contents:

[0005] Obtain initial power flow data for typical grid operation modes; generate a power flow sample dataset through parameter adjustment; calculate sample scores by establishing a safety assessment model and sample evaluation function, and divide the samples into priority experience pools and general experience pools based on the scores;

[0006] A deep reinforcement learning-based intelligent agent model for power section limit assessment was constructed. Its components included a state space, an action space, and a reward function identical to the sample evaluation function. Using a Markov decision process framework, a deep reinforcement learning algorithm was used to pre-train the model using samples from a prioritized experience pool.

[0007] Save the parameters of the pre-trained agent and clear the priority experience pool and the general experience pool. Introduce the uncertainty constraint of renewable energy output and the penalty term for wind and solar power abandonment, modify the reward function, generate extreme load fluctuation sample data and store it in the experience pool, perform secondary training on the model, and obtain new model parameters.

[0008] The fusion strategy distributes entropy optimization mechanism, freezes the shallow parameters of the model's neural network, performs migration optimization training on the model through new scene samples, obtains the final model and saves the parameters.

[0009] When the model is determined to have converged, the policy and value network parameters are saved. The policy distribution entropy weight is incorporated into the policy network loss update iteration. Transfer fine-tuning is then used to freeze the shallow layers of the agent's policy and value networks, preserving the agent's experience extracting underlying features.

[0010] Multiple new scenarios that significantly differ from typical grid operation are generated, such as randomly selecting generator nodes to expand their output limits, randomly selecting generator nodes to deploy new energy units, and randomly selecting generators to be removed from maintenance. These new scenarios are then fed into the agent as initial states. The agent then generates actions and performs power flow calculations, storing the actions and power flow calculation results as new samples in the priority experience pool. The agent then traverses the updated sample data in the priority experience pool, calculates loss values ​​through an experience replay mechanism, and then updates the policy network and value network for training. This ultimately results in a grid section limit calculation model that can adapt to diverse scenarios.

[0011] Preferably, the convergence judgment of the pre-training, secondary training and transfer optimization training depends on the stability of the model loss curve and reward curve in each training stage, specifically: monitoring the fluctuation range of the loss curve and reward curve in the latest consecutive N×30% rounds of training, and judging the convergence and stopping the training when the difference between the maximum value and the minimum value relative to the mean does not exceed the preset threshold, where N is the number of training rounds and N≥100.

[0012] Preferably, the parameterized adjustment includes: using Latin hypercube sampling to generate the generator set output adjustment amount and obtain the adjusted power flow calculation result; merging the initial state data, the adjustment amount and the adjusted power flow calculation result and storing them as samples; screening out redundant samples through similarity analysis and retaining the sample set that meets the preset distribution requirements.

[0013] Preferably, the sample score is calculated as follows:

[0014] The objective function of the security assessment model is defined as: maximizing the transmission power of the selected section while satisfying the grid security constraints; the model's security constraints include power flow convergence, steady-state constraints, and transient constraints;

[0015] The sample evaluation is scored based on the selected section transmission power and safety constraint penalty index. The sample evaluation function is defined as:

[0016]

[0017]

[0018] K=ω1∑ΔP S +ω2∑ΔU+ω3∑ΔI

[0019] Among them, C is the sample score, W is a large positive value, and the maximum penalty is given when the trend does not converge; P S is the cross-section transmission power, C1 is the safety constraint penalty index; K is the severity of violation of the steady-state constraint, ΔP S is the generator active output limit, ΔU is the node voltage limit, ΔI is the line thermal stability limit, and the value is 0 when the above constraints are not violated. ω1, ω2, and ω3 are weight coefficients; η is the transient power angle margin, and λ1, λ2, and λ3 are weight coefficients. When the sample meets all safety constraints, the sample score is determined by the section transmission power. When the sample does not meet the steady-state constraints, a penalty is given according to the amount of limit violation. When the sample meets the steady-state constraints but the transient power angle margin is insufficient, a penalty is given according to the size of the transient power angle margin.

[0020] The threshold value θ is defined as the mean score of the samples in the sample set. The priority experience pool stores samples with scores satisfying C≥θ, and the common experience pool stores samples with scores C<θ, and gives high replay weights to samples in the priority experience pool.

[0021] Preferably, the steady-state constraints are defined as upper and lower limit constraints on the active output of the generator, upper and lower limit constraints on the node voltage, and line thermal stability constraints; the transient constraints are defined as calculating the transient power angle margin η when N-1 faults occur in the power grid through the extended equal area criterion, and when η≥0, it is determined that the sample meets transient stability.

[0022] Preferably, the state space of the power section limit assessment agent model of the deep reinforcement learning is established by the sample state features adjusted in the flow sample data set, and the state features include the active output of the generator, the voltage of each node, the active power of the load, the reactive power of the load, and the transmission flow value of the selected section;

[0023] The action space is established by the generator adjustment amount in the power flow sample data set. The model is trained for several rounds and a fixed number of action adjustments are made in each round. The range of each step action adjustment is P max is the upper limit of generator output, P min is the lower limit of generator output, P0 is the initial generator output for that round, and n is the number of action steps per round. The sensitivity is calculated based on the generator output adjustment and the adjusted cross-sectional power. Generators are ranked by sensitivity, prioritizing the units in the top φ% and bottom φ% of the sensitivity ranking. The remaining units are then adjusted to balance the overall grid output. If any power difference remains after adjustment, the balancing machine assumes the responsibility. If the balancing machine cannot assume the responsibility, a new action is generated. The agent is given the maximum reward when the safety constraints are met and the obtained cross-sectional power is maximized.

[0024] Preferably, the reward function is updated as follows:

[0025]

[0026] K=ω1∑ΔP S +ω2∑ΔU+ω3∑ΔI

[0027] Among them, R is the reward function, W is a large positive value, and the maximum penalty is given when the trend does not converge; P S is the cross-section transmission power, R1 is the safety constraint judgment index, C qfg is the penalty term for wind and solar curtailment, K is the degree to which the sample deviates from the steady-state constraint safety boundary, ΔP S is the limit of the active output of the generator, ΔU is the limit of the node voltage, ΔI is the limit of the line thermal stability, and the value is 0 when the above constraints are not violated. ω1, ω2, and ω3 are weight coefficients; η is the transient power angle margin, are the abandoned wind and solar power, γ qf , γ qg are the penalty coefficients for wind and solar power curtailment, and λ1, λ2, λ3, and λ4 are weight coefficients. The maximum reward is given to the intelligent agent when the safety constraints are met, the obtained section power is maximized, and the amount of wind and solar power curtailment by renewable energy is minimized.

[0028] Preferably, the classification principle of the extreme sample data is: through the real-time interaction between the pre-trained intelligent agent and the environment, the initial state of each training round is adjusted, and the Latin hypercube sampling method is used to generate the load in the typical operation mode. For new samples that fluctuate within the range of [1-ω%, 1+ω%], for rounds whose load is not within the range of [1-ω%, 1+ω%], additional weights are assigned to the samples in the round and stored in the priority experience pool, where

[0029] A power grid section limit assessment system of the present invention comprises:

[0030] The sample generation module is used to generate unit output adjustments through Latin hypercube sampling, perform power flow calculations, and construct sample data sets;

[0031] The evaluation and screening module is equipped with a safety constraint penalty index calculation unit and a sample scoring unit to screen high-value samples based on the cross-section transmission power and the degree of constraint violation;

[0032] The experience pool management module realizes the dynamic storage of the priority experience pool and the common experience pool, and assigns high replay weights to samples in the priority experience pool;

[0033] The agent training module, which includes a state space construction unit, an action strategy generation unit, and a reward function calculation unit, trains deep reinforcement learning models using the Markov decision process framework;

[0034] The migration optimization module is used to freeze the shallow parameters of the neural network and inject the policy distribution entropy weights to perform multi-scenario migration training.

[0035] Preferably, the experience pool management module includes:

[0036] Priority calculation unit, used to dynamically assign playback weights based on sample scores;

[0037] The storage partition unit stores samples with scores higher than the mean into the priority storage area, and the remaining samples into the common storage area; the sampling control unit gives priority to samples in the priority storage area for training, and the sampling probability is positively correlated with the sample score.

[0038] A computer-readable storage medium of the present invention stores a computer program thereon, and when the computer program is executed by a processor, the method for evaluating power grid section limits based on deep reinforcement learning as claimed in the claim above is implemented.

[0039] A computer device of the present invention includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned deep reinforcement learning power grid section limit assessment method are implemented.

[0040] Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: (1) By combining artificial experience to screen high-value samples for pre-training the model, the blind exploration and repeated calculation problems of the traditional deep reinforcement learning model in the initial training stage are effectively avoided, thereby shortening the feature extraction cycle and reducing the overall training time of the model; (2) By introducing the uncertainty constraints of new energy output, extreme samples of load fluctuations and the wind and solar power abandonment penalty mechanism, it is ensured that the evaluation results meet the safety and stability constraints while taking into account the new energy consumption needs, thereby improving the engineering applicability of the decision-making scheme; (3) By adopting the strategy distribution entropy optimization mechanism and the hierarchical migration training strategy, the trained intelligent agent can quickly adapt to complex scene changes such as new energy station switching and load mutation, thereby improving the evaluation accuracy of the model in untrained scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is the overall flow chart of the power grid section limit assessment method of the present invention.

[0042] Figure 2 This is a flow chart of the intelligent agent training in the power grid section limit assessment method of the present invention. DETAILED DESCRIPTION

[0043] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0044] like Figure 1-2 As shown, this embodiment provides a method for evaluating the section limit of a high-proportion new energy power grid based on deep reinforcement learning, including the following contents:

[0045] Obtain initial power flow data of typical power grid operation modes; generate power flow sample data sets through parameter adjustment; calculate sample scores by establishing a safety assessment model and sample evaluation function, and divide the samples into priority experience pool and general experience pool according to the sample scores.

[0046] Sample data from a typical grid operation mode is obtained for power flow calculations, using this typical operation mode as the initial state for subsequent adjustments. Generator output parameters are adjusted and the adjusted power flow calculation results are obtained. The power flow calculation results for the initial state data, the adjustments, and the adjusted power flow calculation results are combined and stored as a single sample. Generator output adjustments are generated using the Latin hypercube sampling method, resulting in a massive sample set. Similarity analysis is used to filter out redundant samples, resulting in a filtered power flow sample dataset.

[0047] A power grid section limit safety assessment model was established. The model's objective function was defined as maximizing the transmission power of a selected section while satisfying grid safety constraints. The model's safety constraints included power flow convergence, steady-state constraints, and transient constraints. Steady-state constraints were defined as upper and lower limits on generator active output, upper and lower limits on node voltage, and line thermal stability constraints. Transient constraints were defined as calculating the transient power angle margin η when N-1 faults occurred in the grid using the extended equal-area criterion. A sample was considered transiently stable when η ≥ 0.

[0048] The sample evaluation is scored based on the selected section transmission power and safety constraint penalty index. The sample evaluation function is defined as:

[0049]

[0050] K=ω1∑ΔP S +ω2∑ΔU+ω3∑ΔI

[0051] Among them, C is the sample score, W is a large positive value, and the maximum penalty is given when the trend does not converge; P S is the cross-section transmission power, C1 is the safety constraint penalty index; K is the severity of violation of the steady-state constraint, ΔP S =(ΔU) represents the generator active output limit, ΔU represents the node voltage limit, and ΔI represents the line thermal stability limit. This value is 0 when the above constraints are not violated. ω1, ω2, and ω3 are weight coefficients. η represents the transient power angle margin, and λ1, λ2, and λ3 are weight coefficients. When a sample meets all safety constraints, its score is determined by the cross-section transmission power. When a sample fails to meet steady-state constraints, a penalty is applied based on the degree of violation. When a sample meets steady-state constraints but has insufficient transient power angle margin, a penalty is applied based on the size of the transient power angle margin.

[0052] Create a priority experience pool and a normal experience pool to store the sample data. The threshold value θ is defined as the mean score of the samples in the sample set. The priority experience pool stores samples with scores satisfying C≥θ, and the normal experience pool stores samples with scores C<θ. Samples in the priority experience pool are given a high replay weight.

[0053] A deep reinforcement learning-based intelligent agent model for power grid section limit assessment is constructed. Its composition includes state space, action space, and a reward function that is the same as the sample evaluation function. Through the Markov decision process framework, a deep reinforcement learning algorithm is used to pre-train the model using samples in the priority experience pool.

[0054] The state space is established by adjusting the sample state characteristics of the flow sample data set, and the state characteristics include the active output of the generator, the voltage of each node, the active power of the load, the reactive power of the load and the transmission flow value of the selected section.

[0055] The action space is established by the generator adjustment amount in the power flow sample data set. The model is trained for several rounds and a fixed number of action adjustments are performed in each round. The range of each action adjustment step is P max is the upper limit of generator output, P min is the lower limit of generator output, P0 is the initial generator output for that round, and n is the number of action steps in each round. The sensitivity is calculated based on the generator output adjustment and the adjusted cross-sectional power. Generators are ranked by sensitivity. During adjustment, the units ranked in the top φ% and bottom φ% of the sensitivity ranking are prioritized for adjustment. The remaining units are then adjusted to balance the overall grid output. If any power difference remains after adjustment, it is assumed by the balancing machine. If the balancing machine cannot assume the responsibility, a new action is generated.

[0056] The reward function is defined as:

[0057]

[0058] K=ω1∑ΔP S +ω2∑ΔU+ω3∑ΔI

[0059] Among them, R is the reward function, W is a large positive value, and the maximum penalty is given when the trend does not converge; P S is the cross-section transmission power, R1 is the safety constraint judgment index; K is the degree of deviation of the sample from the steady-state constraint safety boundary, ΔP S The value is 0 when the above constraints are not violated. ω1, ω2, and ω3 are weight coefficients. η is the transient power angle margin, and λ1, λ2, and λ3 are weight coefficients. The agent is given the maximum reward when the safety constraints are met and the resulting section power is maximized.

[0060] Based on the state space, action space, and reward function, a Markov decision process for power grid section limit assessment is constructed. Based on the Markov decision process framework, a deep reinforcement learning algorithm is used to train the intelligent agent. The reward function and the sample evaluation function are defined identically. The deep reinforcement learning algorithm consists of a policy network and a value network, each of which contains several fully connected layers and activation layers. Samples are selected from the priority experience pool, and their state features are extracted and normalized for preprocessing before being input into the intelligent agent's policy and value networks. The loss value is calculated through an experience replay mechanism to update the network for pretraining.

[0061] When the difference between the maximum and minimum values ​​of the model loss and reward curves relative to the mean in the latest N0 consecutive rounds of training does not exceed the preset threshold ξ, the model is considered to have converged and training is stopped. N0 is defined as 30% of the number of training rounds N. When N is less than 100, the model is not considered to have converged.

[0062] Save the parameters of the pre-trained agent, clear the priority experience pool and the normal experience pool; introduce the uncertainty constraint of renewable energy output and the penalty term for wind and solar power abandonment, modify the reward function, generate extreme load fluctuation sample data and store it in the experience pool, perform secondary training on the model, and obtain new model parameters.

[0063] Add upper and lower output limits for renewable energy stations to the steady-state constraints, and add a new reward function term with the goal of minimizing wind and solar curtailment. Adjust it to the same order of magnitude as the original reward function and then add it. The reward function is updated as follows:

[0064]

[0065] K=ω1∑ΔP S +ω2∑ΔU+ω3∑ΔI

[0066] Among them, R is the reward function, W is a large positive value, and the maximum penalty is given when the trend does not converge; P S is the cross-section transmission power, R1 is the safety constraint judgment index, C qfg is the penalty term for wind and solar curtailment, K is the degree to which the sample deviates from the steady-state constraint safety boundary, ΔP S is the limit of the active output of the generator, ΔU is the limit of the node voltage, ΔI is the limit of the line thermal stability, and the value is 0 when the above constraints are not violated. ω1, ω2, and ω3 are weight coefficients; η is the transient power angle margin, are the abandoned wind and solar power, γ qf , γ qg are the penalty coefficients for wind and solar curtailment, and λ1, λ2, λ3, and λ4 are weight coefficients. The maximum reward is given to the agent when the safety constraints are met, the obtained section power is maximized, and the amount of renewable energy curtailment is minimized.

[0067] The pre-trained agent interacts with the environment in real time, adjusts the initial state of each training round, and generates loads in typical operation modes through the Latin hypercube sampling method. For new samples that fluctuate within the range of [1-ω%, 1+ω%], for rounds whose load is not within the range of [1-ω%, 1+ω%], additional weights are assigned to the samples in the round and stored in the priority experience pool, where The agent traverses the samples with high weights in the priority experience pool, recalculates the loss value through the experience replay mechanism, and updates the policy network and value network for training.

[0068] When the difference between the maximum and minimum values ​​of the model loss and reward curves relative to the mean in the latest N0 consecutive rounds of training does not exceed the preset threshold ξ, the model is considered to have converged and training is stopped. N0 is defined as 30% of the number of training rounds N. When N is less than 100, the model is not considered to have converged.

[0069] The fusion strategy distributes entropy optimization mechanism, freezes the shallow parameters of the model's neural network, performs migration optimization training on the model through new scene samples, obtains the final model and saves the parameters.

[0070] When the model is determined to have converged, the policy and value network parameters are saved. The policy distribution entropy weight is incorporated into the policy network loss update iteration. Transfer fine-tuning is then used to freeze the shallow layers of the agent's policy and value networks, preserving the agent's experience extracting underlying features.

[0071] Multiple new scenarios that significantly differ from typical grid operation are generated, such as randomly selecting generator nodes to expand their output limits, randomly selecting generator nodes to deploy new energy units, and randomly selecting generators to be removed from maintenance. These new scenarios are then fed into the agent as initial states. The agent then generates actions and performs power flow calculations, storing the actions and power flow calculation results as new samples in the priority experience pool. The agent then traverses the updated sample data in the priority experience pool, calculates loss values ​​through an experience replay mechanism, and then updates the policy network and value network for training. This ultimately results in a grid section limit calculation model that can adapt to diverse scenarios.

[0072] When the difference between the maximum and minimum values ​​of the model loss and reward curves relative to the mean in the latest N0 consecutive rounds of training does not exceed the preset threshold ξ, the model is considered to have converged and training is stopped. N0 is defined as 30% of the number of training rounds N. When N is less than 100, the model is not considered to have converged.

[0073] The power grid section limit assessment system described in the present invention includes: a sample generation module for generating unit output adjustment through Latin hypercube sampling, performing power flow calculation and constructing a sample data set; an evaluation and screening module, configured with a safety constraint penalty index calculation unit and a sample scoring unit, screening high-value samples according to the section transmission power and the degree of constraint violation; an experience pool management module, realizing dynamic storage of the priority experience pool and the ordinary experience pool, and assigning high replay weights to the priority experience pool samples; an intelligent agent training module, including a state space construction unit, an action strategy generation unit and a reward function calculation unit, training a deep reinforcement learning model through a Markov decision process framework; a transfer optimization module, used to freeze the shallow parameters of the neural network and inject the policy distribution entropy weight to perform multi-scenario transfer training. The experience pool management module includes: a priority calculation unit for dynamically assigning replay weights according to sample scores; a storage partition unit for storing samples with scores higher than the mean into a priority storage area and the remaining samples into a common storage area; a sampling control unit for preferentially selecting samples in the priority storage area for training, and the sampling probability is positively correlated with the sample score.

[0074] A computer-readable storage medium of the present invention stores a computer program thereon, and when the computer program is executed by a processor, the method for evaluating power grid section limits based on deep reinforcement learning as claimed in the claim above is implemented.

[0075] A computer device of the present invention includes a memory and a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the steps of the above-mentioned deep reinforcement learning method for evaluating power grid section limits are implemented.

[0076] The processor is configured to execute the computer program stored in the memory to implement the various steps in the method involved in the above embodiment.

Claims

1. A power grid section limit assessment method, characterized in that: Includes the following: Obtain initial power flow data for typical grid operation modes; generate a power flow sample dataset through parameter adjustment; calculate sample scores by establishing a safety assessment model and sample evaluation function, and divide the samples into priority experience pools and general experience pools based on the scores; A deep reinforcement learning-based intelligent agent model for power section limit assessment was constructed. Its components included a state space, an action space, and a reward function identical to the sample evaluation function. Using a Markov decision process framework, a deep reinforcement learning algorithm was used to pre-train the model using samples from a prioritized experience pool. Save the parameters of the pre-trained agent and clear the priority experience pool and the general experience pool. Introduce the uncertainty constraint of renewable energy output and the penalty term for wind and solar power abandonment, modify the reward function, generate extreme load fluctuation sample data and store it in the experience pool, perform secondary training on the model, and obtain new model parameters. The fusion strategy distributes entropy optimization mechanism, freezes the shallow parameters of the model's neural network, performs migration optimization training on the model through new scene samples, obtains the final model and saves the parameters.

2. The evaluation method according to claim 1, wherein: The convergence judgment conditions for the pre-training, secondary training and transfer optimization training are as follows: monitoring the fluctuation range of the loss curve and the reward curve in the latest consecutive N×30% rounds of training, and determining that convergence has been achieved and stopping training when the difference between the maximum and minimum values ​​relative to the mean does not exceed a preset threshold, where N is the number of training rounds and N≥100.

3. The evaluation method according to claim 1, wherein: The parameterized adjustment includes: using Latin hypercube sampling to generate the generator set output adjustment and obtain the adjusted power flow calculation results; combining the initial state data, the adjustment amount and the adjusted power flow calculation results and storing them as samples; and filtering out redundant samples through similarity analysis to retain the sample set that meets the preset distribution requirements.

4. The evaluation method according to claim 1, wherein: The sample score is calculated as follows: The objective function of the security assessment model is defined as: maximizing the transmission power of the selected section while satisfying the grid security constraints; the model's security constraints include power flow convergence, steady-state constraints, and transient constraints; The sample evaluation is scored based on the selected section transmission power and safety constraint penalty index. The sample evaluation function is defined as: K=ω1∑ΔP S +ω2∑ΔU+ω3∑ΔI Among them, C is the sample score, W is a large positive value, and the maximum penalty is given when the trend does not converge; P S is the cross-section transmission power, C1 is the safety constraint penalty index; K is the severity of violation of the steady-state constraint, ΔP S is the generator active output limit, ΔU is the node voltage limit, ΔI is the line thermal stability limit, and the value is 0 when the above constraints are not violated. ω1, ω2, and ω3 are weight coefficients; η is the transient power angle margin, and λ1, λ2, and λ3 are weight coefficients. When the sample meets all safety constraints, the sample score is determined by the section transmission power. When the sample does not meet the steady-state constraints, a penalty is given according to the amount of limit violation. When the sample meets the steady-state constraints but the transient power angle margin is insufficient, a penalty is given according to the size of the transient power angle margin. The threshold value θ is defined as the mean score of the samples in the sample set. The priority experience pool stores samples with scores satisfying C≥θ, and the common experience pool stores samples with scores C<θ, and gives high replay weights to samples in the priority experience pool. The steady-state constraints are defined as upper and lower limits on the active output of the generator, upper and lower limits on the node voltage, and line thermal stability constraints. The transient constraints are defined as calculating the transient power angle margin η when N-1 faults occur in the power grid through the extended equal-area criterion. When η ≥ 0, the sample is judged to meet transient stability.

5. The evaluation method according to claim 1, wherein: The state space of the power section limit assessment agent model based on deep reinforcement learning is established by adjusting the sample state features of the power flow sample data set, including the active output of the generator, the voltage of each node, the active power of the load, the reactive power of the load, and the transmission power flow value of the selected section; The action space is established by the generator adjustment amount in the power flow sample data set. The model is trained for several rounds and a fixed number of action adjustments are made in each round. The range of each step action adjustment is P max is the upper limit of generator output, P min is the lower limit of generator output, P0 is the initial generator output of this round, and n is the number of action steps in each round; the sensitivity is calculated based on the generator output adjustment amount and the adjusted section power, and the generators are sorted according to the sensitivity. When adjusting, the units with the top φ% and bottom φ% of the sensitivity ranking are given priority for adjustment, and then the remaining units are adjusted to balance the overall output of the power grid. If there is still a power difference after the adjustment, it will be borne by the balancing machine. If the balancing machine cannot bear it, the action will be regenerated; when the safety constraints are met and the obtained section power is the largest, the intelligent agent is given the maximum reward.

6. The evaluation method according to claim 5, wherein: The reward function is updated as follows: K=ω1∑ΔP S +ω2∑ΔU+ω3∑ΔI Among them, R is the reward function, W is a large positive value, and the maximum penalty is given when the trend does not converge; P S is the cross-section transmission power, R1 is the safety constraint judgment index, C qfg is the penalty term for wind and solar curtailment, K is the degree to which the sample deviates from the steady-state constraint safety boundary, ΔP S is the limit of the active output of the generator, ΔU is the limit of the node voltage, ΔI is the limit of the line thermal stability, and the value is 0 when the above constraints are not violated. ω1, ω2, and ω3 are weight coefficients; η is the transient power angle margin, P t qf 、P t qg are the abandoned wind and solar power, γ qf , γ qg are the penalty coefficients for wind and solar power curtailment, and λ1, λ2, λ3, and λ4 are weight coefficients. The maximum reward is given to the intelligent agent when the safety constraints are met, the obtained section power is maximized, and the amount of wind and solar power curtailment by renewable energy is minimized.

7. The evaluation method according to claim 1, wherein: The classification principle of extreme sample data is as follows: through the real-time interaction between the pre-trained intelligent agent and the environment, the initial state of each training round is adjusted, and the Latin hypercube sampling method is used to generate the load in the typical operation mode. For new samples that fluctuate within the range of [1-ω%, 1+ω%], for rounds whose load is not within the range of [1-ω%, 1+ω%], additional weights are assigned to the samples in the round and stored in the priority experience pool, where 8. A power grid section limit assessment system, characterized in that: include: The sample generation module is used to generate unit output adjustments through Latin hypercube sampling, perform power flow calculations, and construct sample data sets; The evaluation and screening module is equipped with a safety constraint penalty index calculation unit and a sample scoring unit to screen high-value samples based on the cross-section transmission power and the degree of constraint violation; The experience pool management module is used to dynamically store the priority experience pool and the common experience pool, and assign high replay weights to samples in the priority experience pool; The agent training module, which includes a state space construction unit, an action strategy generation unit, and a reward function calculation unit, trains deep reinforcement learning models using the Markov decision process framework; The migration optimization module is used to freeze the shallow parameters of the neural network and inject the policy distribution entropy weights to perform multi-scenario migration training.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the power grid section limit assessment method based on deep reinforcement learning described in claims 1-10 is implemented.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the power grid section limit assessment method based on deep reinforcement learning are implemented.

Images