Power dispatching operation ticket auditing model dynamic fine tuning method, system and device based on gradient enhancement strategy optimization and storage medium

Abstract

The invention discloses a dynamic fine tuning method, system and equipment for a power dispatching operation ticket auditing model based on gradient enhancement strategy optimization, and a storage medium. The method comprises the following steps: converting a power dispatching rule base into a differentiable loss function, and fusing the differentiable loss function into a GRPO training framework; text semantics and power grid topological graph features are extracted, an improved GRPO training framework is utilized, pre-training is carried out based on a Monte Carlo search tree MCTS, and an adversarial sample model conforming to power grid physical constraints is generated; updating rules in real time and importing the rules into the confrontation sample model online; the method focuses on rule dynamic embedding and sample imbalance joint optimization.

Description

Technical Field

[0001] The present invention relates to the technical field of intersection of power system automation and artificial intelligence, and in particular to a method, system, device and storage medium for dynamically fine-tuning a power dispatching operation ticket review model based on gradient reinforcement strategy optimization. Background Art

[0002] Traditional power dispatch operation ticket review relies on manual experience, resulting in low efficiency and susceptible to subjective factors. Existing rule-based automated review systems lack flexibility and are unable to adapt to changes in grid topology or new operational scenarios. Deep learning models suffer from overfitting during review due to insufficient generalization and imbalanced training data (few error samples). For example, review systems based on expert rule bases require manual rule updates, resulting in high maintenance costs. CNN / RNN-based review models lack dynamic optimization capabilities and are ineffective in detecting new types of errors. Error samples account for less than 0.5% of all ticket operations (according to 2022 statistics from a certain provincial dispatch). Adversarial examples generated by traditional data augmentation methods can easily violate device association constraints (e.g., violating busbar topology connections). Operation ticket review requires simultaneous analysis of text instructions (e.g., "pull switch 101") and real-time grid topology. Existing single-modal models cannot capture cross-modal risks (e.g., operating a faulty device without isolating it). Improving model review accuracy with limited error samples while dynamically adapting to changing grid rules is a pressing issue. Summary of the Invention

[0003] Purpose of the invention: The purpose of the present invention is to provide a method, system, device and storage medium for dynamic fine-tuning of the power dispatching operation ticket review model based on gradient reinforcement strategy optimization to solve the problems existing in the background technology.

[0004] Technical solution: The present invention describes a method for dynamically fine-tuning the power dispatching operation ticket review model based on gradient reinforcement strategy optimization, comprising the following steps: collecting power dispatching rules from the power dispatching system regulations, converting each rule in the power dispatching rules into a differentiable loss function, integrating the loss function into the gradient reinforcement strategy through the back-propagation algorithm, and optimizing the loss function of the GRPO training framework; extracting the semantic features of the operation ticket text and the equipment connection relationship features in the power grid topology diagram; fusing the text semantic features with the power grid topology diagram features, and inputting them into the GRPO training framework, performing pre-training based on the Monte Carlo search tree MCTS algorithm, and generating an adversarial sample model that conforms to the physical constraints of the power grid.

[0005] Furthermore, the power dispatching rules include: anti-error blocking logic, grid topology constraints and operation timing restrictions.

[0006] Furthermore, the loss function formula is as follows:

[0007]

[0008] Among them, C k (x) is the kth rule; λ k is the penalty weight of the learnable rule; θ k is the safety threshold.

[0009] Furthermore, the language representation model BERT is used to extract text features; and the graph neural network is used to learn the electrical relationship between devices to extract the power grid topology map features.

[0010] Furthermore, the state space defined in GRPO training is the set of device states of the current operation ticket; the action space is the set of compliant operation replacement actions; and the reward function is the weighted sum of the reduction in rule violation and the semantic rationality score.

[0011] Furthermore, it also includes: when a scheduling rule update is detected, the new rule is dynamically encoded into an incremental loss function, and the incremental loss function is input into the adversarial sample model through an online gradient descent algorithm to update the parameters.

[0012] The present invention provides a system for dynamically fine-tuning a power dispatching operation ticket review model based on gradient reinforcement strategy optimization, comprising:

[0013] Conversion module: used to collect power dispatch rules from the power dispatch system regulations and convert each rule in the power dispatch rules into a differentiable loss function; the loss function is integrated into the gradient reinforcement strategy through the backpropagation algorithm to optimize the loss function of the GRPO training framework;

[0014] Extraction module: used to extract the semantic features of the operation ticket text and the equipment connection relationship features in the power grid topology diagram;

[0015] Training module: It is used to fuse text semantic features with grid topology features, input them into the GRPO training framework, and perform pre-training based on the Monte Carlo search tree (MCTS) algorithm to generate an adversarial sample model that conforms to the physical constraints of the grid.

[0016] Furthermore, in the conversion module, the power dispatching rules include: anti-error blocking logic, grid topology constraints and operation timing restrictions.

[0017] Furthermore, in the conversion module, the loss function formula is as follows:

[0018]

[0019] Among them, C k (x) is the kth rule; λ k is the penalty weight of the learnable rule; θ k is the safety threshold.

[0020] Furthermore, in the extraction module, BERT is used to extract text features; and graph neural networks are used to learn the electrical relationship between devices and the grid topology features.

[0021] Furthermore, in the training module, the state space defined in GRPO training is the set of device states of the current operation ticket; the action space is the set of compliant operation replacement actions; and the reward function is the weighted sum of the reduction in rule violation and the semantic rationality score.

[0022] An electronic device described in the present invention includes a memory, a processor, and a computer program stored in the memory, and the processor implements the steps of any one of the methods when executing the program.

[0023] The computer-readable storage medium of the present invention stores a computer program, which implements the steps of any one of the methods when executed by a processor.

[0024] Beneficial Effects: Compared with existing technologies, this invention offers the following significant advantages: It proposes a tensor representation method for power equipment state transition matrices, ensuring that the GRPO optimization process strictly adheres to Kirchhoff's laws and other physical laws of the power grid. It supports the absorption of new rules through incremental learning without downtime, ensuring 24 / 7 continuous audit service operation. It is suitable for automated auditing of operation tickets at provincial and prefectural power dispatch centers and new energy stations. It also focuses on the joint optimization of dynamic rule embedding and sample imbalance. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Flowchart of the present invention. DETAILED DESCRIPTION

[0026] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0027] like Figure 1 As shown, an embodiment of the present invention provides a method for dynamically fine-tuning a power dispatching operation ticket review model based on gradient enhancement strategy optimization, comprising the following steps:

[0028] (1) Collect power dispatching rules and encode them into a differentiable loss function, where the power dispatching rules include anti-error blocking logic, grid topology constraints, and operation timing restrictions. The loss function formula is as follows:

[0029] The formula is as follows:

[0030]

[0031] Among them, C k (x) is the kth rule; λ k is the penalty weight of the learnable rule; θ kis the safety threshold.

[0032] (2) The GRPO training framework is used to integrate the differentiable loss function with the model classification loss, and the model parameters and rule penalty weights are jointly iteratively updated through the gradient optimization algorithm to achieve coordinated tuning of rule constraints and model parameters; the BERT model is used to extract text features; and the graph neural network is used to learn the electrical relationship between devices to obtain the grid topology map features.

[0033] (3) Generate adversarial samples that comply with the physical constraints of the power grid based on the Monte Carlo search algorithm; the state space is the set of device states of the current operation ticket, the action space is the compliant operation replacement, and the reward function is the weighted sum of the reduction in rule violation and the semantic rationality score.

[0034] (4) Import the updated scheduling rules through the online interface, dynamically encode the new rules into a differentiable loss function and inject it into the optimization process of step (2) to complete the incremental fine-tuning of the model parameters.

[0035] A dynamic fine-tuning system for the power dispatching operation ticket review model based on gradient reinforcement strategy optimization includes:

[0036] Conversion module: This module collects power dispatch rules from the power dispatch system regulations and converts each rule into a differentiable loss function. The loss function is integrated into the gradient reinforcement strategy through the back-propagation algorithm to optimize the loss function of the GRPO training framework. Power dispatch rules include: anti-error blocking logic, grid topology constraints, and operation timing restrictions. The loss function formula is as follows:

[0037]

[0038] Among them, C k (x) is the kth rule; λ k is the penalty weight of the learnable rule; θ k is the safety threshold.

[0039] Extraction module: used to extract the semantic features of the operation ticket text and the equipment connection relationship features in the power grid topology diagram; use BERT to extract text features; and use graph neural networks to learn the electrical relationship between devices and the power grid topology features.

[0040] The training module fuses text semantic features with grid topology features, then feeds them into the GRPO training framework. Pre-training is performed using the Monte Carlo Search Tree (MCTS) algorithm to generate adversarial model models that conform to the physical constraints of the grid. The state space defined in GRPO training is the set of device states in the current operation ticket; the action space is the set of compliant operation replacement actions; and the reward function is a weighted sum of the reduction in rule violation and the semantic rationality score.

[0041] An electronic device described in the present invention includes a memory, a processor, and a computer program stored in the memory, and the processor implements the steps of any one of the methods when executing the program.

[0042] The computer-readable storage medium of the present invention stores a computer program, which implements the steps of any one of the methods when executed by a processor.

Claims

1. A dynamic fine-tuning method for the power dispatching operation ticket review model based on gradient reinforcement strategy optimization, characterized in that: The following steps are involved: The power dispatching rules from the power dispatching system regulations are collected, and each rule in the power dispatching rules is converted into a differentiable loss function. The loss function is integrated into the gradient reinforcement strategy through the back-propagation algorithm to optimize the loss function of the GRPO training framework. The semantic features of the operation ticket text and the equipment connection relationship features in the power grid topology are extracted. After feature fusion of the text semantic features and the power grid topology features, the features are input into the GRPO training framework and pre-trained based on the Monte Carlo search tree (MCTS) algorithm to generate an adversarial sample model that conforms to the physical constraints of the power grid.

2. The method for dynamic fine-tuning of a power dispatching operation ticket audit model based on gradient reinforcement strategy optimization according to claim 1 is characterized in that: Power dispatching rules include: anti-error blocking logic, grid topology constraints and operation timing restrictions.

3. The method for dynamic fine-tuning of the power dispatching operation ticket audit model based on gradient reinforcement strategy optimization according to claim 1 is characterized in that: The loss function formula is as follows: Among them, C k (x) is the kth rule; λ k is the penalty weight of the learnable rule; θ k is the safety threshold.

4. The method for dynamic fine-tuning of a power dispatching operation ticket audit model based on gradient reinforcement strategy optimization according to claim 1 is characterized in that: The language representation model BERT is used to extract text features; the graph neural network is used to learn the electrical relationship between devices and extract the power grid topology map features.

5. The method for dynamic fine-tuning of a power dispatching operation ticket audit model based on gradient reinforcement strategy optimization according to claim 1 is characterized in that: The state space defined in GRPO training is the set of device states of the current operation ticket; the action space is the set of compliant operation replacement actions; and the reward function is the weighted sum of the reduction in rule violation and the semantic rationality score.

6. The method for dynamic fine-tuning of a power dispatching operation ticket audit model based on gradient reinforcement strategy optimization according to claim 1 is characterized in that: Also includes: When a scheduling rule update is detected, the new rule is dynamically encoded into an incremental loss function, and the incremental loss function is input into the adversarial sample model through the online gradient descent algorithm for parameter update.

7. A dynamic fine-tuning system for the power dispatching operation ticket review model based on gradient reinforcement strategy optimization, characterized in that: include: Conversion module: used to collect power dispatch rules from the power dispatch system regulations and convert each rule in the power dispatch rules into a differentiable loss function; the loss function is integrated into the gradient reinforcement strategy through the backpropagation algorithm to optimize the loss function of the GRPO training framework; Extraction module: used to extract the semantic features of the operation ticket text and the equipment connection relationship features in the power grid topology diagram; Training module: It is used to fuse text semantic features with grid topology features, input them into the GRPO training framework, and perform pre-training based on the Monte Carlo search tree (MCTS) algorithm to generate an adversarial sample model that conforms to the physical constraints of the grid.

8. The power dispatching operation ticket audit model dynamic fine-tuning system based on gradient reinforcement strategy optimization according to claim 7 is characterized in that: In the conversion module, the power dispatching rules include: anti-error blocking logic, grid topology constraints and operation timing restrictions.

9. The power dispatching operation ticket audit model dynamic fine-tuning system based on gradient reinforcement strategy optimization according to claim 6 is characterized in that: In the conversion module, the loss function formula is as follows: Among them, C k (x) is the kth rule; λ k is the penalty weight of the learnable rule; θ k is the safety threshold.

10. The power dispatching operation ticket audit model dynamic fine-tuning system based on gradient reinforcement strategy optimization according to claim 7 is characterized in that: In the extraction module, BERT is used to extract text features; graph neural networks are used to learn the electrical relationships between devices and extract power grid topology features.

11. The power dispatching operation ticket review model dynamic fine-tuning system based on gradient reinforcement strategy optimization according to claim 7 is characterized in that: In the training module, the state space defined in GRPO training is the set of device states of the current operation ticket; the action space is the set of compliant operation replacement actions; and the reward function is the weighted sum of the reduction in rule violation and the semantic rationality score.

12. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the method according to any one of claims 1 to 6 when executing the program.

13. A computer-readable storage medium, characterized in that A computer program is stored, and when the program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images