GIM model electric power data normative auditing method, system and device and medium

Abstract

The invention discloses a GIM model electric power data normative auditing method and system. The method comprises the steps of obtaining GIM model real-time electric power data and historical electric power data; performing first anomaly detection on the acquired historical power data through a variational auto-encoder to obtain normal power data fused with an encoder feature space; constructing a reinforcement learning model, and inputting the normal power data fused with the encoder feature space and the real-time power data into the model; performing active learning iteration on the reinforcement learning model in combination with the partial labeled power data, and training to obtain an optimized reinforcement learning model; and inputting the new real-time power data into the reinforcement learning model to obtain a power data auditing and classifying result. According to the method, VAE is used for reconstructing samples, and abnormity is detected by setting a reconstruction loss threshold value; the DQN is introduced to improve the detection precision and explore a new anomaly category, and the recognition capability of the model on non-standard data is enhanced; the efficient and accurate auditing of the data normalization is realized, and the manual intervention is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of power data detection, and in particular to a GIM model power data standardization audit method, system, equipment and medium. Background Art

[0002] In power systems, the Grid Information Model (GIM) serves as a crucial medium for data exchange, storage, and sharing. Data standardization is crucial. With increasing demands for electricity quantity and quality from key users, such as industry and residents, the expansion of the power grid, and the continuous emergence of power equipment and technologies, the complexity of the power data to be processed is increasing.

[0003] At present, the commonly used GIM model data standardization audit method is based on manual and simple rule script program audit. There are many disadvantages in the manual audit method. For example, manual audit requires a lot of manpower and physical resources, and the audit cost is increased. At the same time, the audit efficiency is low and cannot meet the increasingly complex power grid information data review, affecting the normal operation of the power grid system. In addition, manual audits are highly subjective. Different auditors have different professional levels, work experience, and understanding of national audit standards, so it is difficult to ensure the consistency of audit results. While the efficiency is low, the accuracy also has great room for improvement. Although simple script program audits can improve the audit speed of data standardization to a certain extent, they can only make judgments based on pre-set rules. They lack the ability to understand and analyze the deep semantics and complex logical relationships of the data, which will lead to more errors in the audit results. Summary of the Invention

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

[0005] Therefore, the present invention provides a GIM model power data standardization audit method and system to solve the problems of insufficient standardization of GIM data, inconsistent standards, low efficiency and low accuracy in current script and manual audits.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for reviewing the standardization of power data of a GIM model, comprising: obtaining real-time power data and historical power data of the GIM model;

[0008] Performing a first anomaly detection on the acquired historical power data through a variational autoencoder to obtain normal power data that incorporates the encoder feature space;

[0009] Building a reinforcement learning model, and inputting both normal power data integrated with the encoder feature space and real-time power data into the model;

[0010] Combined with some labeled power data, the reinforcement learning model is actively iterated to obtain an optimized reinforcement learning model after training.

[0011] The new real-time power data is input into the reinforcement learning model to obtain the power data audit and classification results.

[0012] As a preferred solution of the GIM model power data standardization audit method of the present invention, wherein: performing a first anomaly detection on the acquired historical power data by using a variational autoencoder includes:

[0013] Input normal historical power data into the variational autoencoder;

[0014] The normal historical power data is learned through the variational autoencoder, and the sample reconstruction is performed to obtain the reconstruction loss;

[0015] The abnormal supervision signal is obtained based on the reconstruction loss, and normal power data is generated by integrating the encoder feature space.

[0016] The beneficial effect of this preferred technical solution is that by setting a specific reconstruction loss threshold, the indicator can be used as an anomaly score, similar to traditional outlier detection technology, to mark samples that deviate from the established norm as potential anomalies.

[0017] As a preferred solution of the GIM model power data standardization audit method of the present invention, a reinforcement learning model is constructed, and normal power data integrated with the encoder feature space and real-time power data are input into the model, including:

[0018] Combine variational autoencoders with reinforcement learning models consisting of an agent, an environment, and a reward function;

[0019] Based on the abnormal supervision signal of the variational autoencoder, a second anomaly detection is performed on unlabeled real-time data. The anomaly score is generated by comparing the normal power data and the abnormal sequence reconstruction that integrates the encoder feature space.

[0020] Based on the anomaly score, the final abnormal data and normal data are obtained and input into the reinforcement learning model;

[0021] The reinforcement learning model's intelligent agent uses a Q-value function to evaluate the cumulative discounted rewards of the current power dispatch action over the future time span. The extrinsic reward is associated with the cost-effectiveness of the operation, while the intrinsic reward is driven by the reconstruction error of the variational encoder to obtain potential anomalies.

[0022] By balancing the immediate benefits and long-term risks of power dispatching through discount factors, the intelligent agent can achieve the dual goals of cost minimization and active anomaly detection under safety constraints.

[0023] The beneficial effects of this preferred technical solution are: by integrating VAE into a weakly supervised learning environment and selecting DQN as a deep learning method, combined with normal power data with an encoder feature space for anomaly detection, the detection accuracy can be improved, while exploring new anomaly categories in unlabeled data.

[0024] As a preferred solution of the GIM model power data standardization audit method of the present invention, the active learning iteration of the reinforcement learning model is performed in combination with part of the labeled power data, including:

[0025] An active learning mechanism is introduced into the reinforcement learning model, and experience replay is used to store the state transition information of the power grid;

[0026] Processing unlabeled real-time power data through marginal sampling;

[0027] The intelligent agent calculates the Q value according to the real-time status of the power grid and selects quantitatively. It selects part of the real-time power data according to the minimum marginal value for labeling, and the labeled data enters the subsequent training.

[0028] As a preferred solution of the GIM model power data standardization audit method of the present invention, the training includes:

[0029] The preprocessed GIM model data is divided into training set, validation set and test set;

[0030] The model parameters are adjusted through the cross entropy loss function and Adam optimizer, and the model is evaluated through the validation set to obtain the optimized reinforcement learning model after training.

[0031] As a preferred solution of the GIM model power data standardization audit method of the present invention, the evaluation by the Q-value function is specifically expressed as follows:

[0032]

[0033] Among them, E π represents the expectation of the power dispatch strategy; γ represents the discount factor considering the cost-effectiveness of power system operation, which is used to adjust the weight of future rewards in the cumulative rewards; T represents the future time; t represents the current time; s t represents the parameter status of the power grid at time t, including power consumption, transmission capacity, and power generation capacity, a t Represents the decision action of the power grid dispatching center at time t.

[0034] As a preferred solution of the GIM model power data standardization review method of the present invention, preprocessing is performed after the data is acquired, and the preprocessing includes:

[0035] Use Kalman filter to denoise and obtain denoised data;

[0036] The denoised data is combined with mean filling and LSTM model to perform data filling on different types of data;

[0037] After padding, standardization and normalization are performed to obtain preprocessed data for input.

[0038] In a second aspect, the present invention provides a GIM model power data standardization review system, comprising:

[0039] Acquisition module, used to obtain real-time power data and historical power data of GIM model;

[0040] An anomaly detection module, configured to perform a first anomaly detection on the acquired historical power data using a variational autoencoder to obtain normal power data that incorporates the encoder feature space;

[0041] A model building module, configured to input normal power data integrated with the encoder feature space and real-time power data into the model;

[0042] The iterative optimization module is used to combine some labeled power data to perform active learning iterations on the reinforcement learning model, and obtain an optimized reinforcement learning model after training;

[0043] The output module is used to input new real-time power data into the reinforcement learning model to obtain the power data audit classification results.

[0044] In a third aspect, the present invention provides a computer device, comprising:

[0045] memory and processor;

[0046] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the GIM model power data standardization review method are implemented.

[0047] In a fourth aspect, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the GIM model power data standardization review method.

[0048] Compared with the existing technology, the present invention has the following beneficial effects: in terms of data irregularity detection, the present invention constructs a learning model that integrates variational autoencoders (VAE), deep reinforcement learning (DQN) and active learning modules. VAE is used to reconstruct samples, and anomalies are detected by setting a reconstruction loss threshold; DQN is introduced to improve detection accuracy and explore new anomaly categories, and VAE is used to provide signals for weakly supervised learning; active learning technology of marginal sampling is adopted, combined with the experience replay mechanism, and a reward function containing external and internal rewards is used to balance exploration and utilization, efficiently screen unlabeled instances for labeling, and enhance the model's ability to recognize irregular data; efficient and accurate review of the standardization of GIM model data is achieved, and manual intervention is reduced. It can automatically identify potential irregular data structures and logical errors, and can adapt to the ever-changing substation engineering design specifications and industry standards. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 The figure is a schematic diagram of the overall process of the GIM model power data standardization review method according to an embodiment of the present invention.

[0051] Figure 2 This is a GIM model data non-standardization detection method framework based on deep learning in the GIM model power data standardization review method described in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0053] Example 1, reference Figure 1 , as an embodiment of the present invention, provides a GIM model power data standardization review method, comprising:

[0054] S100: Obtaining real-time power data and historical power data of the GIM model;

[0055] S200: performing a first anomaly detection on the acquired historical power data by using a variational autoencoder to obtain normal power data that incorporates the encoder feature space;

[0056] S300: Build a reinforcement learning model and input the normal power data integrated with the encoder feature space and the real-time power data into the model;

[0057] S400: Active learning iterations are performed on the reinforcement learning model in combination with some labeled power data to obtain an optimized reinforcement learning model after training.

[0058] S500: Input the new real-time power data into the reinforcement learning model to obtain the power data audit classification result.

[0059] It should be noted that deep learning methods can automatically extract complex features and patterns from the massive amounts of data in GIM models, effectively identifying potential irregular data structures and logical errors. Furthermore, compared to simple rule-based scripts, they can adapt to the ever-changing design specifications and industry standards of substation engineering projects, improving the accuracy and consistency of GIM model data audits.

[0060] Therefore, steps S100-S500 aim to comprehensively analyze and conduct compliance audits on the digitally transferred GIM model. A deep learning model is constructed, combining a variational autoencoder (VAE), deep reinforcement learning (DQN), and active learning modules to detect irregular data structures and logical errors. After the model is constructed, it is trained and optimized using the dataset to improve its anomaly detection performance. Audit standards and thresholds are then set to determine data compliance.

[0061] Example 2, reference Figure 1-Figure 2 , which is an embodiment of the present invention, provides a GIM model power data standardization review method based on the above embodiment.

[0062] In the embodiment of the present application, in step S100, the real-time power data and historical power data of the GIM model are obtained, wherein the real-time data may include: current / voltage waveform data of the transmission line, transformer winding temperature data, and power load data; the historical data may include: equipment failure event records and equipment maintenance logs

[0063] In an optional embodiment, the acquisition method in step S100 may be to collect real-time data of power grid nodes through PMU (synchronized phasor measurement unit), and the real-time data may also include a distributed power output curve.

[0064] In another optional embodiment, the acquisition method in step S100 can be to connect the equipment asset management system (EAM) to the Oracle database to obtain structured data including equipment ledgers, maintenance records, etc. The historical data can also include: annual load characteristic curves and meteorological related data.

[0065] In the embodiment of the present application, preprocessing is performed after the data is acquired in step S100. The preprocessing includes:

[0066] Use Kalman filter to denoise and obtain denoised data;

[0067] The denoised data is combined with mean filling and LSTM model to perform data filling on different types of data;

[0068] After padding, standardization and normalization are performed to obtain preprocessed data for input.

[0069] Specifically, during denoising, in the power system, noise often presents a non-Gaussian distribution, and the basic Kalman filter method is used for denoising; for missing values ​​in the data, different filling methods are selected according to the data characteristics. For data with stable statistical characteristics, such as fixed parameters of certain substation equipment, mean filling is used. By calculating the average value of this type of data in the complete record, the missing values ​​are filled with this average value; for data with time series characteristics, such as equipment temperature, a prediction model based on the long short-term memory network (LSTM) is used. The LSTM model is first trained using complete historical data to learn the time series characteristics and change trends of the data, and then the trained model is used to predict and fill in the missing data. In order to meet the data input requirements of subsequent deep models, the data is standardized and normalized. During standardization, the mean and standard deviation of each feature in the data set are calculated, and then the mean is subtracted from each data point and divided by the standard deviation, thereby converting the data into standard normal distribution data with a mean of 0 and a standard deviation of 1; during normalization, the minimum-maximum normalization method is used, and its formula is as follows:

[0070]

[0071] Among them, X is the original data, X min and X max are the minimum and maximum values ​​of the feature data.

[0072] In an optional implementation, after the data is acquired in step S100, pre-processing may be performed through operations such as spatiotemporal alignment or pre-enhancement.

[0073] In the embodiment of the present application, in step S200, a first anomaly detection is performed on the acquired historical power data by using a variational autoencoder to obtain normal power data that integrates the encoder feature space, including the following steps A1-A3:

[0074] A1: Input normal historical power data into the variational autoencoder;

[0075] A2: Use a variational autoencoder to learn normal historical power data, perform sample reconstruction, and obtain the reconstruction loss;

[0076] A3: Obtain abnormal supervision signals based on reconstruction loss and generate normal power data that integrates the encoder feature space.

[0077] It should be noted that when evaluating unlabeled samples, a variational autoencoder (VAE) attempts to reconstruct each example. Normal examples can typically be reconstructed with minimal loss, indicating that they conform to the learned normal patterns. In contrast, anomalous examples, because they deviate from these patterns, tend to incur higher reconstruction losses. By setting a specific reconstruction loss threshold, this metric can be used as an anomaly score. Examples with reconstruction losses exceeding this threshold are classified as anomalous. This approach is similar to traditional outlier detection techniques, which flag examples that deviate from established norms as potential anomalies.

[0078] In an optional embodiment, the normal power data obtained in step S200 that is fused with the encoder feature space can also be obtained through an adversarial generative network + reconstruction error correction form, for example: a generator network (Generator) and a discriminator network (Discriminator) are constructed; normal historical power data (such as transformer temperature time series) is input, and the generator learns to reconstruct the normal data distribution while introducing Wasserstein distance constraints to ensure that the generated data satisfies physical laws; a 3σ boundary is calculated based on the score distribution of historical normal data, and the normal data and the generated data are aligned in the discriminator feature space by maximum mean difference (MMD), and the fused feature vector is output as reinforcement learning input;

[0079] In another optional embodiment, the normal power data obtained in step S200 that is fused with the encoder feature space can also be used as reinforcement learning input by combining time contrastive learning with prototype clustering to obtain fused feature data.

[0080] It should also be noted that the present invention prefers variational autoencoder (VAE) because it uses explicit probability distribution modeling through latent variables (such as Gaussian distribution), can accurately describe the physical constraints of power data, and directly reconstruct data that conforms to the differential equations of the power system; compared with other forms such as GAN, the implicit distribution is difficult to guarantee physical laws such as power conservation, and additional corresponding constraints need to be added. Contrastive learning lacks a generative model and cannot directly reconstruct data that conforms to the differential equations of the power system; at the same time, the reconstruction loss of VAE obeys a distribution that can be analytically derived, while the discriminator output of GAN has the risk of adversarial oscillation, and contrastive learning cannot directly generate verifiable samples.

[0081] In the embodiment of the present application, a reinforcement learning model is constructed in step S300, and the normal power data integrated with the encoder feature space and the real-time power data are input into the model, including the following steps B1-B5:

[0082] B1: Combining a variational autoencoder with a reinforcement learning model consisting of an agent, an environment, and a reward function;

[0083] B2: Based on the abnormal supervision signal of the variational autoencoder, a second anomaly detection is performed on the unlabeled real-time data. The anomaly score is generated by comparing the normal power data and the abnormal sequence reconstruction that integrates the encoder feature space.

[0084] B3: Based on the anomaly score, obtain the final abnormal data and normal data and input them into the reinforcement learning model;

[0085] B4: The reinforcement learning model's agent uses a Q-value function to evaluate the cumulative discounted reward of the current power dispatch action over the future time span. The extrinsic reward is associated with the cost-effectiveness of the operation, while the intrinsic reward is driven by the reconstruction error of the variational encoder to detect potential anomalies.

[0086] B5: By balancing the immediate benefits and long-term risks of power dispatch through discount factors, the intelligent agent can achieve the dual goals of cost minimization and active anomaly detection under safety constraints.

[0087] In the embodiment of the present application, step S300 is evaluated by the Q value function, which is specifically expressed as:

[0088]

[0089] Among them, E π represents the expectation of the power dispatch strategy; γ represents the discount factor considering the cost-effectiveness of power system operation, which is used to adjust the weight of future rewards in the cumulative rewards; T represents the future time; t represents the current time; s t represents the parameter status of the power grid at time t, including power consumption, transmission capacity, and power generation capacity, a tRepresents the decision action of the power grid dispatching center at time t.

[0090] It should be noted that if Figure 2 As shown in the figure, by integrating VAE into the weakly supervised learning environment and optimizing DQN as the deep learning method, combined with normal power data with encoder feature space for anomaly detection, the detection accuracy can be improved while exploring new anomaly categories in unlabeled data. Figure 2 The state represents the complete information representation of the current situation of the environment. It is the basis for the agent to make decisions. It is multi-source data, which can be real-time measurement data, equipment status and other information. The environment balances the use of known anomalies and the exploration of new data through a hybrid reward function. The state-action value function (Q-value function) is used to evaluate the effectiveness of actions and update strategies to maximize the cumulative discounted reward of the agent. In order to balance exploration and utilization, a reward function containing extrinsic rewards and intrinsic rewards is set. Figure 2 Here, Xr is the state reconstructed by VAE, and the reconstruction error serves as an intrinsic reward to help the agent explore potential irregular data.

[0091] In an optional embodiment, the reinforcement learning model of step S300 can also guide hierarchical reinforcement learning through a physical model. For example, regional load demands are generated based on VAE encoding features to form high-level demands, and low-level demands are generated in response to real-time power fluctuations at the minute level; soft constraints of the power flow equation are embedded in the reward function; action masks are dynamically generated using VAE reconstruction errors, and transfer training is performed to obtain the final strategy.

[0092] In another optional embodiment, the reinforcement learning model of step S300 can also be reinforced through multi-agent game learning, for example: constructing different types of agents: on the power generation side, such as thermal power / new energy units, the goal is to minimize power generation costs, and on the transmission and distribution side, substations / line controllers, the goal is to minimize network losses; constructing a non-zero-sum game framework based on the above goals; conducting federal curriculum learning based on the power grid topology, and introducing conditional risk value to obtain the final strategy.

[0093] It should also be noted that the present invention prefers to integrate VAE into a weakly supervised learning environment and prefers DQN as a deep learning method because VAE can generate pseudo labels through unsupervised reconstruction errors, allowing DQN to effectively learn with only 5%-10% labeled data, which can reduce manual labeling compared to other models that require 100% calibration data; at the same time, the joint model of VAE reconstruction error and DQN can be compressed to <8MB and deployed on edge devices such as distribution terminals (DTUs), which is more suitable for large-scale centralized scheduling systems, while multi-agent or multi-level models are basically larger than 200MB in size or have large video memory requirements.

[0094] In the embodiment of the present application, in step S400, active learning iteration is performed on the reinforcement learning model in combination with part of the labeled power data, including the following steps C1-C3:

[0095] C1: Introducing an active learning mechanism into the reinforcement learning model and using experience replay to store the state transition information of the power grid;

[0096] C2: Processing unlabeled real-time power data through marginal sampling;

[0097] C3: The intelligent agent selects quantitatively by calculating the Q value based on the real-time status of the power grid, selects some real-time power data based on the minimum margin value for labeling, and the labeled data enters subsequent training.

[0098] It should be noted that the introduction of active learning mechanism can improve data utilization efficiency and update stability.

[0099] In the embodiment of the present application, the training in step S400 includes the following steps D1-D2:

[0100] D1: Divide the preprocessed GIM model data into training set, validation set and test set;

[0101] D2: Adjust the model parameters through the cross entropy loss function and Adam optimizer, evaluate the model through the validation set, and obtain the optimized reinforcement learning model after training.

[0102] For example, we take the pre-processed GIM model data from the substation, use 70% of the data as the training set, 20% as the validation set, and the remaining 10% as the test set. We use the cross entropy loss function to measure the difference between the model prediction results and the true labels. The formula is:

[0103]

[0104] Among them, N is the number of samples, C is the number of categories (standard and non-standard are two categories), y ij Indicates that the i-th sample belongs to the j-th class label, p ij It represents the probability that the model predicts that the i-th sample belongs to the j-th class.

[0105] In an optional implementation, in step D2 of S400 , the model may be evaluated by calculating performance indicators such as accuracy, recall, and F1 value of the model on a validation set.

[0106] S500: Inputting new real-time power data into the reinforcement learning model to obtain power data audit classification results, such as whether the data is standardized or non-standard; compliant with standards or not.

[0107] In summary, this method combines substation engineering design specifications and industry standards, using a deep learning model to extract features from GIM model data and identify potential irregular data structures and logical errors. Based on established audit criteria, it performs multi-level logical analysis on the parsed file data, ensuring data integrity and consistency and generating detailed audit reports. This method not only improves the efficiency and accuracy of data audits but also reduces the need for daily manual intervention through intelligent means, providing a solid foundation for subsequent data applications and promoting the effective application and management of digital twin technology in power systems.

[0108] Example 3, the above is a schematic scheme of a GIM model power data standardization audit method. It should be noted that the technical solution of the GIM model power data standardization audit system and the technical solution of the GIM model power data standardization audit method above are of the same concept. For details not described in detail in the technical solution of the GIM model power data standardization audit system in this embodiment, please refer to the description of the technical solution of the GIM model power data standardization audit method above.

[0109] This embodiment also provides a GIM model power data standardization review system, including:

[0110] Acquisition module, used to obtain real-time power data and historical power data of GIM model;

[0111] An anomaly detection module, configured to perform a first anomaly detection on the acquired historical power data using a variational autoencoder to obtain normal power data that incorporates the encoder feature space;

[0112] A model building module, configured to input normal power data fused with encoder feature space and real-time power data into the model;

[0113] The iterative optimization module is used to combine some labeled power data to perform active learning iterations on the reinforcement learning model, and obtain an optimized reinforcement learning model after training;

[0114] The output module is used to input new real-time power data into the reinforcement learning model to obtain the power data audit classification results.

[0115] This embodiment also provides a computer device suitable for the standardization review of GIM model power data, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the standardization review method of GIM model power data proposed in the above embodiment.

[0116] This embodiment also provides a storage medium having a computer program stored thereon. When the program is executed by a processor, the method for implementing the GIM model power data standardization review proposed in the above embodiment is implemented.

[0117] The storage medium proposed in this embodiment and the method for implementing the GIM model power data standardization review proposed in the above embodiment belong to the same inventive concept. The technical details not described in detail in this embodiment can be referred to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0118] Through the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented with the help of software and necessary general hardware, and of course can also be implemented by hardware. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of the present invention.

[0119] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A GIM model power data standardization audit method, characterized by: include: Obtain real-time and historical power data of the GIM model; Performing a first anomaly detection on the acquired historical power data through a variational autoencoder to obtain normal power data that incorporates the encoder feature space; Building a reinforcement learning model, and inputting both normal power data integrated with the encoder feature space and real-time power data into the model; Combined with some labeled power data, the reinforcement learning model is actively iterated to obtain an optimized reinforcement learning model after training. The new real-time power data is input into the reinforcement learning model to obtain the power data audit and classification results.

2. A GIM model power data standardization audit method according to claim 1, characterized in that: The first anomaly detection is performed on the acquired historical power data using a variational autoencoder, including: Input normal historical power data into the variational autoencoder; The normal historical power data is learned through the variational autoencoder, and the sample reconstruction is performed to obtain the reconstruction loss; The abnormal supervision signal is obtained based on the reconstruction loss, and normal power data is generated by integrating the encoder feature space.

3. A GIM model power data standardization audit method according to claim 2, characterized in that: Construct a reinforcement learning model and input the normal power data integrated with the encoder feature space and the real-time power data into the model, including: Combine variational autoencoders with reinforcement learning models consisting of an agent, an environment, and a reward function; Based on the abnormal supervision signal of the variational autoencoder, a second anomaly detection is performed on unlabeled real-time data. The anomaly score is generated by comparing the normal power data and the abnormal sequence reconstruction that integrates the encoder feature space. Based on the anomaly score, the final abnormal data and normal data are obtained and input into the reinforcement learning model; The reinforcement learning model's intelligent agent uses a Q-value function to evaluate the cumulative discounted rewards of the current power dispatch action over the future time span. The extrinsic reward is associated with the cost-effectiveness of the operation, while the intrinsic reward is driven by the reconstruction error of the variational encoder to obtain potential anomalies. By balancing the immediate benefits and long-term risks of power dispatching through discount factors, the intelligent agent can achieve the dual goals of cost minimization and active anomaly detection under safety constraints.

4. A GIM model power data standardization audit method according to claim 3, characterized in that: Combined with some labeled power data, the reinforcement learning model is actively iterated, including: An active learning mechanism is introduced into the reinforcement learning model, and experience replay is used to store the state transition information of the power grid; Processing unlabeled real-time power data through marginal sampling; The intelligent agent calculates the Q value according to the real-time status of the power grid and selects quantitatively. It selects part of the real-time power data according to the minimum marginal value for labeling, and the labeled data enters the subsequent training.

5. A GIM model power data standardization audit method according to claim 4, characterized in that: The training includes: The preprocessed GIM model data is divided into training set, validation set and test set; The model parameters are adjusted through the cross entropy loss function and Adam optimizer, and the model is evaluated through the validation set to obtain the optimized reinforcement learning model after training.

6. A GIM model power data standardization audit method according to claim 5, characterized in that: The evaluation is performed through the Q-value function, which is specifically expressed as: Among them, E π represents the expectation of the power dispatch strategy; γ represents the discount factor considering the cost-effectiveness of power system operation, which is used to adjust the weight of future rewards in the cumulative rewards; T represents the future time; t represents the current time; s t represents the parameter status of the power grid at time t, including power consumption, transmission capacity, and power generation capacity, a t Represents the decision action of the power grid dispatching center at time t.

7. A GIM model power data standardization review method according to claim 1 or 6, characterized in that: After acquiring the data, preprocessing is performed, and the preprocessing includes: Use Kalman filter to denoise and obtain denoised data; The denoised data is combined with mean filling and LSTM model to perform data filling on different types of data; After padding, standardization and normalization are performed to obtain preprocessed data for input.

8. A GIM model power data standardization review system, applying the method according to any one of claims 1 to 7, characterized in that: include: Acquisition module, used to obtain real-time power data and historical power data of GIM model; An anomaly detection module, configured to perform a first anomaly detection on the acquired historical power data using a variational autoencoder to obtain normal power data that incorporates the encoder feature space; A model building module, configured to input normal power data integrated with the encoder feature space and real-time power data into the model; The iterative optimization module is used to combine some labeled power data to perform active learning iterations on the reinforcement learning model, and obtain an optimized reinforcement learning model after training; The output module is used to input new real-time power data into the reinforcement learning model to obtain the power data audit classification results.

9. A computer device comprising: memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the GIM model power data standardization review method described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions, when executed by a processor, implement the steps of the GIM model power data standardization review method according to any one of claims 1 to 7.

Images