Construction method and system of intelligent potential feature identification power marketing risk rule base

Abstract

The invention discloses a construction method and system of an intelligent potential feature identification power marketing risk rule base, and belongs to the field of unsupervised risk identification. The method comprises the following steps: collecting power marketing data, and converting the power marketing data into an acceptable format of an LSTM-VAE model; constructing an LSTM-VAE model to extract potential features in the power marketing data; training an LSTM-VAE model by using the power marketing data; constructing an auto-encoder model based on a neural network, and carrying out sparse coding and reconstruction on the potential features extracted by the LSTM-VAE model; training an auto-encoder model by using the potential features extracted by the LSTM-VAE model; using the trained auto-encoder model to calculate a reconstruction error; performing abnormal point detection on the reconstruction error by using a 3 sigma principle; and adding the data judged to be abnormal points into a risk rule base, then classifying the added data and samples in the base by using a K-mean clustering method, and updating the risk rule base in real time. According to the invention, power marketing data anomaly detection under unsupervised learning is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of unsupervised risk identification, and specifically relates to a method and system for constructing an intelligent potential feature identification power marketing risk rule base based on unsupervised learning. Background Art

[0002] In the field of power marketing events, the identification of abnormal risk events is crucial to ensure the stable operation of the power system and prevent potential economic losses. However, due to the low frequency of abnormal events and the lack of clear definitions, coupled with the lack of real label information, there is relatively little research on unsupervised anomaly detection methods. In addition, in power marketing events, there is also a lack of real labels, which further increases the difficulty of detecting abnormal risk events.

[0003] With the increasing number of sensors in smart grids and modern power marketing systems, anomaly detection of power marketing events has become increasingly difficult. How to effectively process measurement data from multiple sensors to learn the dynamics and context of the system and improve the operating efficiency of the system has become a new challenge. Traditional mathematical learning methods have limitations when dealing with large-scale data sources and have difficulty coping with the complexity and diversity of data. Therefore, the introduction of deep learning (DL) technology has become an attractive option. Deep learning can effectively capture the potential patterns of context perception and is widely regarded as one of the most advanced methods in various fields.

[0004] In deep learning-based representation learning, there have been some studies devoted to designing misbehavior detection methods to identify abnormal events in multi-sensor data. However, these methods usually fail to extract a strong and unified representation of the underlying normal behavior from multi-sensor data while preserving the dependencies between events. Therefore, further exploration and development of innovative methods are needed to overcome these challenges and improve the ability to accurately detect and effectively model risk events in multi-sensor data. Summary of the invention

[0005] In view of the problems existing in the above-mentioned prior art, the present invention proposes a method and system for constructing a power marketing risk rule base for intelligent potential feature identification, which utilizes a combined model (LSTM-VAE) of a long short-term memory network (LSTM) and a variational autoencoder (VAE) to extract potential features in power marketing data, and sparsely encodes and reconstructs these potential features through a neural network-based autoencoder (NNAE-AD) model, thereby realizing the identification of abnormal risk events.

[0006] To this end, the present invention adopts the following technical solution.

[0007] In a first aspect, the present invention provides a method for constructing a rule base for intelligent potential feature identification of power marketing risks, which comprises: Step 1), collect power marketing data and convert it into a format acceptable to the LSTM-VAE model; Step 2), build an LSTM-VAE model to extract latent features from power marketing data; Step 3), use power marketing data to train the LSTM-VAE model; Step 4), build a neural network-based autoencoder model to sparsely encode and reconstruct the latent features extracted by the LSTM-VAE model; Step 5), train the autoencoder model using the latent features extracted by the LSTM-VAE model; Step 6), use the trained autoencoder model to calculate the reconstruction error; Step 7), apply the 3σ principle to detect outliers in the reconstruction error; Step 8) Add the data judged as outliers to the risk rule library, and then use the K-mean clustering method to classify the added data and the samples in the library, and update the risk rule library in real time.

[0008] The present invention can effectively deal with the detection problem of abnormal events in power marketing data, provide an efficient and accurate risk identification and modeling method, and significantly improve the operating efficiency and safety of the power marketing system. The power marketing risk rule library refers to a library containing power marketing risk event rules.

[0009] The present invention is particularly suitable for scenarios where risk events are identified, and can locate abnormal events in complex monitoring scenarios. By combining LSTM-VAE and NNAE-AD models, efficient feature extraction and accurate anomaly detection can be achieved, and the risk rule base can be updated in real time. The present invention is suitable for power marketing, time series data analysis, and fields lacking labeled data, and can process data of unknown abnormal types. It has wide applicability and flexibility, and provides an efficient and reliable solution for power marketing risk control and other time series data analysis.

[0010] Furthermore, in step 2), the LSTM-VAE model: the encoder part is composed of an LSTM layer, which converts time series data into potential features; the decoder part is composed of an LSTM layer and a fully connected layer, which restores the potential features to original data.

[0011] Furthermore, in step 3), the training process of the LSTM-VAE model is as follows: The power marketing data is divided into a training set and a validation set, and then the data is input into the LSTM-VAE model. The output is calculated through forward propagation, and the reconstruction error between the output and the input and the potential feature regularization term are calculated as the loss function. The back propagation algorithm is used to update the model parameters according to the gradient of the loss function to optimize the model performance. During the training process, you need to set the number of training rounds and batch size, and monitor the loss values ​​of the training set and validation set to avoid overfitting.

[0012] Furthermore, the specific contents of step 4) are as follows: 41) Define the input layer of the autoencoder model, whose input feature dimension is the same as the latent feature dimension extracted by LSTM-VAE; 42) Then build the encoder part, compressing the input features into a low-dimensional sparse code through a fully connected layer; 43) Construct the decoder part to reconstruct the sparse code back to the original feature space; when compiling the model, select the appropriate optimizer and loss function to ensure that the model can effectively learn and reconstruct features during training.

[0013] Furthermore, in step 5), the latent features extracted by the LSTM-VAE model are used as input data and input into the autoencoder model for training. By training the autoencoder model, it is enabled to recognize and distinguish normal patterns from abnormal patterns.

[0014] Furthermore, when the input data is in a normal mode, the autoencoder model can accurately reconstruct the input data with a small reconstruction error; when the input data is in an abnormal mode, since its characteristics are significantly different from the normal mode, it is difficult for the autoencoder model to accurately reconstruct the input data and the reconstruction error is large; by monitoring the reconstruction error, the autoencoder model can effectively identify and distinguish between the normal mode and the abnormal mode of power marketing data.

[0015] Furthermore, in step 6), the calculation process of the reconstruction error is as follows: The reconstruction error is calculated by comparing the difference between the original latent features and the reconstructed latent features, using the mean square error as the metric.

[0016] Furthermore, in step 7), the specific steps of applying the 3σ principle to detect outliers in the reconstruction error are as follows: 71) Calculate the mean μ of all reconstruction errors and the standard deviation σ of all reconstruction errors; 72) Setting the threshold: According to the 3σ principle, the threshold is set to μ+3σ; data points whose reconstruction error exceeds this threshold are considered as outliers and have large reconstruction errors.

[0017] Furthermore, the specific process in step 8) is as follows: 81) Add the identified outlier data to the risk rule library to ensure that all abnormal events are recorded and tracked; 82) Perform K-means clustering on the data in the risk rule library, classify the newly added data and samples in the library, and thus identify and classify different types of risk events; 83) Periodically or when new data is available, use the K-means model to classify the new data based on the newly added data to ensure that the classification and clustering results in the risk rule base are always the latest and most accurate.

[0018] In a second aspect, the present invention provides a system for constructing a power marketing risk rule base for intelligent potential feature identification, which is used to implement a method for constructing the power marketing risk rule base for intelligent potential feature identification, including: Data preparation unit: used to collect power marketing data and convert it into a format acceptable to the LSTM-VAE model; LSTM-VAE model building unit: used to build LSTM-VAE models to extract latent features from power marketing data; LSTM-VAE model training unit: Use power marketing data to train the LSTM-VAE model; Autoencoder model building unit: used to build an autoencoder model based on a neural network, and to perform sparse encoding and reconstruction on the latent features extracted by the LSTM-VAE model; Autoencoder model training unit: trains the autoencoder model using the latent features extracted by the LSTM-VAE model; Reconstruction error calculation unit: uses the trained autoencoder model to calculate the reconstruction error; Outlier detection unit: uses the 3σ principle to detect outliers in the reconstruction error; Risk rule base forming unit: add the data judged as abnormal points to the risk rule base, and then use the K-mean clustering method to classify the added data and the samples in the library, and update the risk rule base in real time.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1. Unsupervised learning capability: Traditional anomaly detection methods usually rely on a large amount of labeled data, while this paper adopts an unsupervised learning model. By combining LSTM-VAE and NNAE-AD models, it realizes efficient anomaly detection without labeled data, greatly reducing the cost and time of data annotation.

[0020] 2. Efficient feature extraction: The present invention utilizes the LSTM-VAE model to effectively extract potential features from power marketing data, especially when processing time series data. It can capture the temporal dependencies in the data and provide more accurate feature representation.

[0021] 3. Accurate anomaly detection: Through the sparse coding and reconstruction mechanism of the NNAE-AD model, the present invention can accurately identify and distinguish normal modes from abnormal modes, significantly improving the accuracy and reliability of anomaly detection.

[0022] 4. Real-time update: The present invention uses K-means clustering method to dynamically update the risk rule base to ensure that the system can adapt to new data and risk types and always maintain high efficiency and accuracy.

[0023] 5. Wide applicability: The present invention is not only applicable to power marketing data, but can also be applied to other time series data analysis fields, especially in the absence of labeled data and the presence of unknown anomaly types, with good adaptability and flexibility.

[0024] 6. Reduce manual intervention: By automatically extracting, encoding and reconstructing data features, the present invention realizes a highly automated anomaly detection process, reduces manual intervention and improves work efficiency.

[0025] In general, the present invention has significant benefits in reducing data annotation costs, improving the accuracy of feature extraction and anomaly detection, real-time system updates, and wide applicability, providing an efficient and reliable solution for power marketing risk control and time series data anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a principle diagram of a method for constructing a power marketing risk rule base for intelligent potential feature recognition according to the present invention; Figure 2 It is a structural diagram of a system for constructing a power marketing risk rule base using intelligent potential features for identifying the risks of the power marketing according to the present invention. DETAILED DESCRIPTION

[0027] The present invention will be further described below in conjunction with the accompanying drawings and specific implementation methods.

[0028] Example 1 In view of the shortcoming that the existing technology cannot detect abnormal risk events in marketing event scenarios, this embodiment provides a method for constructing an intelligent potential feature recognition power marketing risk rule library, which can accurately identify abnormal risk events in marketing event data.

[0029] The method for constructing the intelligent potential feature recognition power marketing risk rule base is as follows: Figure 1As shown, the steps are as follows: Step 1), collect power marketing data and convert it into a format acceptable to the LSTM-VAE model; Step 2), build an LSTM-VAE model to extract latent features from power marketing data; Step 3), use power marketing data to train the LSTM-VAE model; Step 4) construct a neural network-based autoencoder model (i.e., NNAE-AD model) to sparsely encode and reconstruct the latent features extracted by the LSTM-VAE model; Step 5), train the autoencoder model using the latent features extracted by the LSTM-VAE model; Step 6), use the trained autoencoder model to calculate the reconstruction error; Step 7), apply the 3σ principle to detect outliers in the reconstruction error; Step 8) Add the data judged as outliers to the risk rule library, and then use the K-mean clustering method to classify the added data and the samples in the library, and update the risk rule library in real time.

[0030] The present invention is particularly suitable for scenarios where risk events are identified. It can locate abnormal events in complex monitoring scenarios. By combining LSTM-VAE and NNAE-AD models, it can achieve efficient feature extraction and accurate anomaly detection, and can update the risk rule base in real time. It is suitable for power marketing, time series data analysis, and areas where labeled data is lacking. It can process data of unknown abnormal types, has wide applicability and flexibility, and provides an efficient and reliable solution for power marketing risk control and other time series data analysis.

[0031] Specifically, in step 2), the construction of the LSTM-VAE model is designed and constructed to effectively extract latent features from power marketing data. First, LSTM is used to process and capture temporal dependencies in time series data, and can remember and utilize long-term contextual information. The variational autoencoder (VAE) learns the distribution of data by encoding the input data into features in a latent space and then decoding from the latent space to reconstruct the input data. LSTM-VAE combines the time series processing capability of LSTM and the feature extraction capability of VAE. In the specific implementation process, the encoder part consists of LSTM layers, which convert time series data into latent features; the decoder part consists of LSTM layers and fully connected layers, which restore the latent features to the original data.

[0032] Specifically, in step 3), the training LSTM-VAE model: divide the power marketing data into a training set and a validation set, then input the data into the LSTM-VAE model, calculate the output through forward propagation, and calculate the reconstruction error between the output and the input and the regularization term of the potential features as the loss function. Use the backpropagation algorithm to update the model parameters according to the gradient of the loss function to optimize the model performance. During the training process, it is necessary to set the number of training rounds and batch size, and monitor the loss values ​​of the training and validation sets to avoid overfitting. After multiple iterations of training, the LSTM-VAE model can accurately reconstruct the input data and extract representative potential features, laying the foundation for subsequent anomaly detection.

[0033] Specifically, in step 4), the construction of the autoencoder model (NNAE-AD model) is to design an autoencoder model (NNAE-AD model) based on a neural network for sparse encoding and reconstruction of the latent features extracted by the LSTM-VAE model. The specific contents are as follows: 41) Define the input layer of the autoencoder, whose input dimension is the same as the latent feature dimension generated by the LSTM-VAE model.

[0034] 42) Then construct the encoder part, which compresses the input features into a lower-dimensional sparse code through a fully connected layer.

[0035] 43) Build the decoder part to reconstruct the sparse coding back to the original feature space. When compiling the model, choose the appropriate optimizer and loss function to ensure that the model can effectively learn and reconstruct features during training.

[0036] The NNAE-AD model is a neural network-based autoencoder model designed to sparsely encode and reconstruct the latent features extracted by the LSTM-VAE model. During training, the autoencoder learns how to compress the input features into a lower-dimensional sparse encoding and reconstruct the original features from this representation. Through repeated training, the NNAE-AD model gradually grasps the normal pattern of the data, that is, the feature distribution and structure in most cases.

[0037] Specifically, in step 5), the latent features extracted by the LSTM-VAE model are used as input data and input into the autoencoder model for training. By training the autoencoder model, it is enabled to recognize and distinguish normal patterns from abnormal patterns, providing a solid foundation for subsequent anomaly detection.

[0038] When the input data is in a normal mode, the autoencoder can reconstruct the input data well with a small reconstruction error. However, when the input data is in an abnormal mode, since its characteristics are significantly different from the normal mode, it is difficult for the autoencoder to accurately reconstruct the input data, resulting in a large reconstruction error.

[0039] By monitoring the reconstruction error, the NNAE-AD model can effectively identify and distinguish normal and abnormal patterns in power marketing data. This capability provides a solid foundation for subsequent anomaly detection, because by calculating the reconstruction error, potential anomalies can be quickly detected and marked, enabling timely early warning and processing of risk events in power marketing data.

[0040] Specifically, in step 6), the calculation process of the reconstruction error is as follows: The calculation process of the reconstruction error is as follows: The reconstruction error is calculated by comparing the difference between the original latent features and the reconstructed latent features, using the mean square error (MSE) as a metric. Specifically, the square error between the original features and the reconstructed features of each data point is calculated and averaged. A large reconstruction error usually indicates that the data point is abnormal, because the reconstruction error of a normal data point should be small. By analyzing the reconstruction error, anomalies in power marketing data can be effectively identified, providing a basis for the detection and handling of risk events.

[0041] Specifically, in step 7), in order to determine the outliers, the 3σ principle can be used to detect outliers on the reconstruction error. The specific steps are as follows: 71) Calculate the mean (μ) and standard deviation (σ) of all reconstruction errors; 72) Setting the threshold: According to the 3σ principle, the threshold is set to μ+3σ. Data points whose reconstruction error exceeds the threshold are considered abnormal points, and the reconstruction error is large; data points whose reconstruction error is within the range of μ+3σ have a small reconstruction error. In this way, there is no need to manually set the threshold, but the threshold is given by a mathematical formula.

[0042] In this way, abnormal events in power marketing data can be effectively identified. These abnormal events usually represent potential risks or abnormal behaviors. Timely detection and processing of these events can help prevent potential losses and optimize power marketing strategies. With this approach, abnormal points can be determined more scientifically and the accuracy and reliability of risk event detection can be improved.

[0043] Specifically, in step 8), the risk rule base is written as follows: The data judged as abnormal risk events are written into the risk rule library, and then the K-means clustering method is used to classify the newly added data and the samples in the library, and the risk rule library is updated in real time. The specific steps are as follows: 81) Write to risk rule base: Write the identified abnormal point data into the risk rule base to ensure that all abnormal events are recorded and tracked; 82) Use K-means clustering for classification: Perform K-means clustering on the data in the risk rule library, classify the newly added data with the samples in the library, and thus identify and classify different types of risk events; 83) Periodically or when new data is available, use the K-means model to classify the new data based on the newly added data to ensure that the classification and clustering results in the risk rule base are always the latest and most accurate.

[0044] Example 2 This embodiment provides a system for constructing a power marketing risk rule base for intelligent potential feature identification, which is used to implement the above-mentioned method for constructing a power marketing risk rule base for intelligent potential feature identification, such as Figure 2 As shown, it consists of a data preparation unit, an LSTM-VAE model construction unit, an LSTM-VAE model training unit, an autoencoder model construction unit, a reconstruction error calculation unit, an anomaly detection unit and a risk rule base formation unit.

[0045] Data preparation unit: used to collect power marketing data and convert it into a format acceptable to the LSTM-VAE model; LSTM-VAE model building unit: used to build LSTM-VAE models to extract latent features in power marketing data.

[0046] LSTM-VAE model training unit: Use power marketing data to train the LSTM-VAE model.

[0047] Autoencoder model building unit: used to build an autoencoder model based on a neural network, and to perform sparse encoding and reconstruction on the latent features extracted by the LSTM-VAE model.

[0048] Autoencoder model training unit: Train the autoencoder model using the latent features extracted by the LSTM-VAE model.

[0049] Reconstruction error calculation unit: Use the trained autoencoder model to calculate the reconstruction error.

[0050] Outlier detection unit: Use the 3σ principle to detect outliers in the reconstruction error.

[0051] Risk rule base forming unit: add the data judged as abnormal points to the risk rule base, and then use the K-mean clustering method to classify the added data and the samples in the library, and update the risk rule base in real time.

[0052] It should be noted that each unit in the above-mentioned system for constructing a rule base for power marketing risks using intelligent potential features can be implemented in whole or in part by software, hardware, and a combination thereof. Each of the above-mentioned units can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each of the above-mentioned units. For specific limitations on a system for constructing a rule base for power marketing risks using intelligent potential features, please refer to the limitations on a method for constructing a rule base for power marketing risks using intelligent potential features above. The two have the same functions and effects, and will not be repeated here.

[0053] Example 3 A method for constructing a power marketing risk rule base based on intelligent potential features identification is provided, and the steps are as follows.

[0054] (1) Data preparation: In this step, data related to power marketing needs to be collected. This data may include various types of information, such as customer power consumption, payment records, power failure records, user behavior data, etc. The collected data needs to be preprocessed to convert it into a format suitable for LSTM-VAE model processing. Since the LSTM-VAE model is good at processing time series data, the data needs to be organized into a time series format, that is, a sequence of multiple time steps is input into the model.

[0055] Assume that the collected data is ,in Represents a single data point. Each data point It may contain information in multiple dimensions, such as electricity consumption, payment amount, etc. Convert this data into a time series format , each data point is a multidimensional vector. This conversion process may involve preprocessing steps such as data normalization, missing value filling, and noise filtering.

[0056] (2) Constructing LSTM-VAE model: The LSTM-VAE (Long Short-Term Memory-Variational Autoencoder) model combines the advantages of LSTM network and variational autoencoder to process time series data and extract its latent features. The LSTM network is used to capture the long-term dependencies of the time series, while the variational autoencoder (VAE) part is used to generate latent variables to represent the data.

[0057] LSTM encoder: takes the input data Pass it to the LSTM encoder to get the hidden state and cell status .

[0058] (1)

[0059] The LSTM encoder processes time series data through a multi-layer recurrent neural network (RNN) to capture its temporal dependencies.

[0060] Latent features: hidden state Mapping to latent features Through the fully connected layer.

[0061] (2)

[0062] Among the potential features, is through the mean and standard deviation This step allows for probabilistic modeling of the data.

[0063] LSTM decoder: From latent features Reconstructing input data .

[0064] (3)

[0065] LSTM decoder uses latent features Generated and original input data Similar reconstruction data .

[0066] (3) Training the LSTM-VAE model: The collected power marketing data is used to train the LSTM-VAE model. By optimizing the loss function of the LSTM-VAE model, the model can effectively capture the potential characteristics of the data.

[0067] The loss function of the LSTM-VAE model is trained by using reconstruction loss and KL divergence loss.

[0068] (4)

[0069] Among them, the reconstruction loss measures the difference between the original data and the reconstructed data, and the KL divergence (Kullback-Leibler Divergence) measures the difference between the potential distribution and the prior distribution. By minimizing the total loss function, the LSTM-VAE model can learn effective potential features.

[0070] (4) Constructing the NNAE-AD model: The NNAE-AD (neural network-based autoencoder-anomaly detection) model is used to further process and analyze the latent features extracted by LSTM-VAE. The model is designed to sparsely encode and reconstruct these features for subsequent anomaly detection. The autoencoder of NNAE-AD is designed by designing a neural network autoencoder (NNAE) for sparse encoding and reconstruction of latent features.

[0071] (5)

[0072] The autoencoder consists of an encoder and a decoder. The encoder transforms the latent features Mapping to a more compact representation , and the decoder attempts to Reconstruct the original input data .

[0073] (5) Extract latent features and train NNAE-AD model: Extract latent features from the LSTM-VAE model and use these latent features to train the NNAE-AD model. By minimizing the reconstruction loss, the autoencoder is trained to accurately reconstruct the input features.

[0074] Feature Extraction: Extracting latent features from the LSTM-VAE model .

[0075] Train the autoencoder: Use these latent features to train an autoencoder model, minimizing the reconstruction loss.

[0076] (6)

[0077] The training process optimizes the loss function so that the autoencoder can learn effective sparse coding and reconstruct features.

[0078] (6) Calculate reconstruction error: Use the trained NNAE-AD model to calculate the reconstruction error of each data point. The reconstruction error is the difference between the input features and the reconstructed features, which is used to evaluate the reconstruction ability of the model.

[0079] The trained NNAE-AD model is used to calculate the reconstruction error of each data point. The calculation formula is as follows.

[0080] (7)

[0081] The size of the reconstruction error reflects the accuracy of the model in reconstructing features. A larger reconstruction error may indicate that a data point is abnormal. The present invention detects this error through the following 3σ principle.

[0082] (7) Risk event identification: Apply the 3σ principle to detect outliers, that is, by calculating the mean and standard deviation of the reconstruction error, we can identify risk events corresponding to abnormally large reconstruction errors.

[0083] 3σ principle: Apply the 3σ principle to detect abnormal points. When the reconstruction error is larger than this value, it can be said that the power marketing data is abnormal, and then the data is written into the risk rule library. The formula is as follows.

[0084] (8)

[0085] in, and are the mean and standard deviation of the reconstruction error, respectively. This step uses statistical principles to identify reconstruction errors that exceed the normal range as outliers.

[0086] (8) Writing into the risk rule base: Risk events identified as abnormal are written into the risk rule base, and the new data is classified with samples in the base using the K-means clustering method to ensure real-time updating and accurate classification of the risk rule base.

[0087] Classify and update risk events, write risk events that are judged to be abnormal into the risk rule library, and use the K-means clustering method to classify the added data and samples in the library. The calculation formula is as follows: (9)

[0088] The K-means clustering method can group similar risk events into one category, allowing the rule base to better organize and classify risk events. In the future, when new data arrives, similar risk events can be quickly identified and handled based on the classification results in the rule base.

[0089] Through the above steps, an intelligent latent feature identification power marketing risk rule base can be constructed, and the LSTM-VAE and NNAE-AD models can be used to effectively extract and analyze the latent features in power marketing data, and identify and classify risk events.

[0090] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A method for constructing a risk rule base for power marketing based on intelligent potential features, It is characterized in that include: Step 1), collect power marketing data and convert it into a format acceptable to the LSTM-VAE model; Step 2), build an LSTM-VAE model to extract latent features from power marketing data; Step 3), use power marketing data to train the LSTM-VAE model; Step 4), build a neural network-based autoencoder model to sparsely encode and reconstruct the latent features extracted by the LSTM-VAE model; Step 5), train the autoencoder model using the latent features extracted by the LSTM-VAE model; Step 6), use the trained autoencoder model to calculate the reconstruction error; Step 7), apply the 3σ principle to detect outliers in the reconstruction error; Step 8) Add the data judged as outliers to the risk rule library, and then use the K-mean clustering method to classify the added data and the samples in the library, and update the risk rule library in real time.

2. According to the method for constructing a rule base for intelligent potential feature identification of power marketing risks according to claim 1, It is characterized in that In step 2), the LSTM-VAE model described above: the encoder part consists of LSTM layers, which converts time series data into latent features; The decoder part consists of LSTM layers and fully connected layers to restore the latent features to the original data.

3. According to the method for constructing a rule base for intelligent potential feature identification of power marketing risks according to claim 1, It is characterized in that In step 3), the training process of the LSTM-VAE model is as follows: The power marketing data is divided into a training set and a validation set, and then the data is input into the LSTM-VAE model. The output is calculated through forward propagation, and the reconstruction error between the output and the input and the regularization term of the potential features are calculated as the loss function. The back propagation algorithm is used to update the model parameters according to the gradient of the loss function to optimize the model performance. During the training process, you need to set the number of training rounds and batch size, and monitor the loss values ​​of the training set and validation set to avoid overfitting.

4. According to the method for constructing a rule base for intelligent potential feature identification of power marketing risks according to claim 1, It is characterized in that The specific contents of step 4) are as follows: 41) Define the input layer of the autoencoder model, whose input feature dimension is the same as the latent feature dimension extracted by LSTM-VAE; 42) Then build the encoder part, compressing the input features into a low-dimensional sparse code through a fully connected layer; 43) Construct the decoder part to reconstruct the sparse code back to the original feature space; when compiling the model, select the appropriate optimizer and loss function to ensure that the model can effectively learn and reconstruct features during training.

5. According to the method for constructing a rule base for intelligent potential feature identification of power marketing risks according to claim 1, It is characterized in that In step 5), the latent features extracted by the LSTM-VAE model are used as input data and input into the autoencoder model for training. By training the autoencoder model, it can recognize and distinguish normal patterns from abnormal patterns.

6. According to the method for constructing a rule base for intelligent potential feature identification of power marketing risks according to claim 5, It is characterized in that When the input data is in a normal pattern, the reconstruction error is small, and the autoencoder model can accurately reconstruct the input data; when the input data is in an abnormal pattern, since its characteristics are significantly different from the normal pattern, it is difficult for the autoencoder model to accurately reconstruct the input data and the reconstruction error is large; by monitoring the reconstruction error, the autoencoder model can effectively identify and distinguish between the normal and abnormal patterns of power marketing data.

7. According to claim 1, the method for constructing a rule base for intelligent potential feature identification of power marketing risks, It is characterized in that In step 6), the calculation process of the reconstruction error is as follows: The reconstruction error is calculated by comparing the difference between the original latent features and the reconstructed latent features, using the mean square error as the metric.

8. According to claim 1, the method for constructing a rule base for intelligent potential feature identification of power marketing risks, It is characterized in that In step 7), the specific steps of applying the 3σ principle to detect outliers in the reconstruction error are as follows: 71) Calculate the mean μ of all reconstruction errors and the standard deviation σ of all reconstruction errors; 72) Setting the threshold: According to the 3σ principle, the threshold is set to μ+3σ; data points whose reconstruction error exceeds this threshold are regarded as outliers and the reconstruction error is considered large.

9. According to claim 1, the method for constructing a rule base for intelligent potential feature identification of power marketing risks, It is characterized in that The specific process in step 8) is as follows: 81) Add the identified outlier data to the risk rule library to ensure that all abnormal events are recorded and tracked; 82) Perform K-means clustering on the data in the risk rule library, classify the newly added data and samples in the library, and thus identify and classify different types of risk events; 83) Periodically or when new data is available, use the K-means model to classify the new data based on the newly added data to ensure that the classification and clustering results in the risk rule base are always the latest and most accurate.

10. A system for constructing a rule base for intelligent latent feature identification of power marketing risks, used to implement the method for constructing a rule base for intelligent latent feature identification of power marketing risks as described in any one of claims 1 to 9, It is characterized in that include: Data preparation unit: used to collect power marketing data and convert it into a format acceptable to the LSTM-VAE model; LSTM-VAE model building unit: used to build LSTM-VAE models to extract latent features from power marketing data; LSTM-VAE model training unit: Use power marketing data to train the LSTM-VAE model; Autoencoder model building unit: used to build an autoencoder model based on a neural network, and to perform sparse encoding and reconstruction on the latent features extracted by the LSTM-VAE model; Autoencoder model training unit: trains the autoencoder model using the latent features extracted by the LSTM-VAE model; Reconstruction error calculation unit: uses the trained autoencoder model to calculate the reconstruction error; Outlier detection unit: uses the 3σ principle to detect outliers in the reconstruction error; Risk rule base forming unit: add the data judged as abnormal points to the risk rule base, and then use the K-mean clustering method to classify the added data and the samples in the library, and update the risk rule base in real time.

Images