Industry classification method based on multi-source data fusion and related equipment

Abstract

The application provides an industry classification method based on multi-source data fusion and related equipment, which comprises the following steps: obtaining enterprise characteristic information of a target power-using enterprise; preprocessing the enterprise characteristic information to obtain an enterprise text word sequence and a power consumption time sequence characteristic matrix; inputting the enterprise text word sequence into a pre-trained text encoding model for semantic encoding to obtain a text feature vector, and inputting the power consumption time sequence characteristic matrix into a pre-trained time sequence encoder for time sequence encoding to obtain a power consumption time sequence vector; weighting and fusing the text feature vector and the power consumption time sequence vector based on a pre-set attention fusion mechanism to determine a plurality of candidate industry categories and a probability value of each candidate industry category, and determining an industry category of the target power-using enterprise according to the probability values of the plurality of candidate industry categories. The application can fully mine the power-using enterprise text data and power consumption data, and improve the accuracy and reliability of the enterprise industry category.

Description

Technical Field

[0001] This application relates to the field of electricity enterprise classification technology, and in particular to an industry classification method and related equipment based on multi-source data fusion. Background Technology

[0002] In the power sector, the determination of user electricity prices is closely related to multiple factors, including their industry classification, electricity consumption category, and operating capacity. With continuous economic development and technological advancements, emerging and cross-sectoral business models such as data centers, vehicle charging stations, cold chain logistics, and high-tech laboratories are constantly emerging. Users are no longer limited to traditional industrial, agricultural, and residential categories, but exhibit diversified development characteristics, making user industry classification and electricity consumption characteristics increasingly complex. Consequently, relying on traditional industry judgment standards based on experience is insufficient to fully cover and adapt to the current complex classification scenarios, significantly increasing the difficulty of determining user electricity prices.

[0003] In related technologies, after obtaining enterprise text data that reflects the direction of enterprise production and operation and business activities, text classification algorithms can be used to identify and label the enterprise text data in order to determine the national economic industry category to which the corresponding enterprise belongs. However, the current application of classification methods has limited context length processing capabilities. When processing enterprise text data containing long texts, segmentation is required, which can easily destroy the semantic coherence of the text, making it difficult to accurately understand the overall description of the enterprise, thereby reducing the accuracy of industry classification. Summary of the Invention

[0004] In view of this, the purpose of this application is to propose an industry classification method and related equipment based on multi-source data fusion, so as to solve some or all of the technical problems mentioned above.

[0005] To achieve the above objectives, this application provides an industry classification method based on multi-source data fusion, including:

[0006] Obtain enterprise characteristic information of the target electricity-consuming enterprise; The enterprise feature information is preprocessed to obtain the enterprise text word sequence and electricity consumption time series feature matrix; The enterprise text word sequence is input into a pre-trained text encoding model for semantic encoding to obtain a text feature vector, and the electricity consumption time series feature matrix is ​​input into a pre-trained time series encoder for time series encoding to obtain an electricity consumption time series vector. The text feature vector and the electricity consumption time series vector are weighted and fused based on a preset attention fusion mechanism to determine multiple candidate industry categories and the probability value of each candidate industry category. The industry category of the target electricity-consuming enterprise is determined based on the probability values ​​of the multiple candidate industry categories.

[0007] Based on the same inventive concept, this application also provides an industry classification device based on multi-source data fusion, comprising: The data acquisition module is configured to acquire enterprise characteristic information of the target electricity-consuming enterprise; The data preprocessing module preprocesses the enterprise feature information to obtain the enterprise text word sequence and electricity consumption time series feature matrix; The data encoding module is configured to input the enterprise text word sequence into a pre-trained text encoding model for semantic encoding to obtain a text feature vector, and to input the electricity consumption time series feature matrix into a pre-trained time series encoder for time series encoding to obtain an electricity consumption time series vector. The category determination module is configured to perform weighted fusion of the text feature vector and the electricity consumption time series vector based on a preset attention fusion mechanism to determine multiple candidate industry categories and the probability value of each candidate industry category, and determine the industry category of the target electricity-consuming enterprise based on the probability values ​​of the multiple candidate industry categories.

[0008] Based on the same inventive concept, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.

[0009] As can be seen from the above, the industry classification method and related equipment based on multi-source data fusion provided in this application preprocess the acquired enterprise feature information to obtain enterprise text word sequences and electricity consumption time-series feature matrices. This method can remove data noise from enterprise feature information, standardize data format, and extract effective feature data to improve the overall quality and reliability of basic data. By using a pre-trained text encoding model and encoder to encode and represent the enterprise text word sequences and electricity consumption time-series feature matrices respectively, text feature vectors and electricity consumption time-series vectors can be obtained. This can fully explore the semantic information of the text and the changing patterns of electricity consumption time-series, generate highly recognizable and robust feature vectors, realize the conversion of multi-source data to a unified feature space, and balance the representation effect with the efficiency of batch processing. By introducing and utilizing an attention fusion mechanism to adaptively weight and fuse the two types of feature vectors, multiple candidate industry categories and the probability value of each candidate industry category can be determined. This fully leverages the complementary advantages of multimodal features, weakens the interference of invalid features, and combines probability discrimination to achieve accurate determination of enterprise industry categories, reduce the classification bias caused by single features, and improve the accuracy, stability, and generalization ability of classification results. This method can fully mine multi-source data such as text data and electricity consumption data of electricity-consuming enterprises, avoid the limitations of single information judgment, and effectively improve the classification accuracy and reliability of enterprise industry categories. At the same time, it can be adapted to the large-scale and engineering application scenarios of large-scale enterprise data in the industry, thereby improving the accuracy of enterprise industry category positioning and the efficiency of subsequent electricity price determination. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart of the industry classification method in the embodiments of this application; Figure 2 This is a structural block diagram of the industry classification device in the embodiments of this application; Figure 3 This is a structural block diagram of the electronic device in the embodiments of this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0013] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0014] In the power sector, the determination of user electricity prices is closely related to multiple factors, including their industry classification, electricity consumption category, and operating capacity. With economic development and technological progress, emerging and cross-sectoral business models such as data centers, electric vehicle charging stations, cold chain logistics, and high-tech laboratories are constantly appearing. User types are no longer limited to traditional industrial, agricultural, and residential categories, but exhibit diversified development characteristics, making the classification of user industries and electricity consumption characteristics increasingly complex. Based on this, industry judgment standards relying on traditional experience are insufficient to fully cover and adapt to the current complex classification scenarios, thus significantly increasing the difficulty of determining user electricity prices.

[0015] Enterprise text data (such as company name, business scope, main business, registration information, and operating qualifications) is crucial for determining the national economic industry category of an enterprise. It reflects the enterprise's production and operation direction and business activities, thus being a significant factor influencing electricity pricing. Since enterprise text data contains detailed information related to industry categories, it can be identified and labeled using text classification algorithms after acquisition to determine the enterprise's corresponding national economic industry category. In related technologies, the industry widely uses models such as BERT (Bidirectional Encoder Representations from Transformers) to process enterprise text data, determining the corresponding enterprise's industry category based on traditional machine learning and deep learning text classification methods. However, these text classification methods have limitations. In processing enterprise text data, their context processing capabilities are limited. For long text data, segmentation is required, which can easily disrupt the semantic coherence of the enterprise text data, leading to an inaccurate understanding of the overall enterprise description, thus reducing the accuracy of industry classification and affecting the accuracy of electricity pricing.

[0016] In some examples, traditional machine learning text classification methods can be used to process enterprise text data. Specifically, after obtaining the enterprise text data, expert experience can be used to manually extract features from the data, combining these features into a high-dimensional feature vector. After obtaining the high-dimensional feature vector, traditional classifiers such as Naive Bayes and Support Vector Machines can be used to complete industry classification. However, for traditional machine learning text classification methods, the classification effect is highly dependent on the quality of manual feature engineering, requiring high standards for feature extraction and selection. Furthermore, the process of feature analysis and processing is relatively complex, consuming significant manpower and time, resulting in insufficiently comprehensive features that are difficult to accurately adapt to emerging or complex industry types, easily leading to industry classification biases.

[0017] In some examples, deep learning-based text classification methods can also be used to process enterprise text data. Compared to traditional machine learning-based text classification methods, deep learning-based methods utilize more complex deep learning model structures, allowing for direct learning and automatic extraction of high-level semantic features from the original enterprise text data, reducing reliance on manual feature engineering. Specifically, after acquiring the enterprise text data, it undergoes text preprocessing and vectorization. Then, a pre-built and trained neural network model is used to process the preprocessed enterprise text data, achieving automatic classification of enterprise industry categories. Simultaneously, the model can be trained based on the enterprise text data, automatically learning key parameters to complete the industry classification task. However, while deep learning-based text classification methods can automatically learn features, current mainstream solutions are typically based on models like BERT. These models have a context length limit of 512 units, requiring segmentation into smaller segments before processing enterprise text data containing long texts. However, text segmentation can easily disrupt the semantic coherence of the original data, making it difficult for the model to comprehensively and accurately grasp the overall enterprise information.

[0018] In view of this, this application provides an industry classification method based on multi-source data fusion, which can be implemented using a classification controller or a computer as the execution subject, such as... Figure 1 As shown, the method may include: S100: Obtain enterprise characteristic information of the target electricity-consuming enterprise; Enterprise characteristic information of target electricity-consuming enterprises can serve as key data for determining their industry category. This enterprise characteristic information includes characteristic data that is related to their industry category. Specifically, it can include text data that reflects the enterprise's production and operation direction and business activities, as well as electricity consumption data that reflects the enterprise's electricity consumption behavior patterns and electricity load characteristics, so as to provide reliable data support for the industry category classification of target electricity-consuming enterprises.

[0019] Furthermore, compared to relying solely on text data of target electricity-consuming enterprises, using enterprise characteristic information that includes both text data and electricity consumption data can more comprehensively and accurately depict the industry attributes of target electricity-consuming enterprises. Specifically, text data can clearly define the business operations and industry attributes of target electricity-consuming enterprises, while electricity consumption data can accurately reflect the electricity load characteristics of target electricity-consuming enterprises. In this way, through the mutual complementarity between data, misjudgment of industry categories caused by similar single text data can be avoided, thereby improving the reliability and accuracy of subsequent industry classification.

[0020] In practice, a classification controller can be used as the execution subject. The classification controller can obtain the enterprise characteristic information of the target electricity-consuming enterprise, so as to provide a reliable data foundation for the classification method of the industry.

[0021] For example, when obtaining text data from enterprise characteristic information, it is possible to connect to the business platform system related to the target electricity-consuming enterprise, use a preset identification mark (such as the unified social credit code) as the retrieval basis, and download or collect the text data of the target electricity-consuming enterprise from the corresponding platform system, including but not limited to the enterprise name, business scope, main business, registration information and business qualifications.

[0022] For example, when obtaining electricity consumption data from enterprise characteristic information, it is possible to connect to the electricity consumption information collection system corresponding to the target electricity-consuming enterprise and continuously collect the load data and daily electricity consumption data of the target electricity-consuming enterprise within a preset time period.

[0023] S200: Preprocess the enterprise feature information to obtain the enterprise text word sequence and electricity consumption time series feature matrix; Preprocessing enterprise characteristic information can remove invalid noise, standardize the data format of enterprise characteristic information, and extract effective features related to the classification of target electricity-consuming enterprises, so as to provide a standardized and high-quality data foundation for subsequent processing.

[0024] For example, when preprocessing the enterprise characteristic information of the target electricity-consuming enterprise, the text data in the enterprise characteristic information can first be cleaned to remove irrelevant characters and obtain cleaned text data. At this time, the cleaned text data can be segmented based on a preset power industry professional dictionary and Chinese word segmentation tool, and common stop words in the text data can be removed during the segmentation process to reduce interference factors in the text data and obtain segmented text data. After that, the segmented text data can be converted into enterprise text word sequences to lay the data foundation for the subsequent vectorization processing of text data.

[0025] For example, when preprocessing the enterprise characteristic information of the target electricity-consuming enterprise, the electricity consumption data can be divided and organized according to the time series of the electricity consumption data. Taking a preset time granularity of 15 minutes as a sampling point, the 24-hour (i.e., the whole day) electricity consumption duration of the target electricity-consuming enterprise is divided and organized into multiple electricity consumption time series data (i.e., 96 sampling points), and a standard daily load curve of the target electricity-consuming enterprise for the whole day can be generated. At the same time, multiple electricity consumption features can be extracted from each electricity consumption time series data, and the extracted multiple electricity consumption features are used to construct an electricity consumption time series feature matrix so that it can reflect the electricity consumption scale and electricity consumption pattern of the target electricity-consuming enterprise.

[0026] Furthermore, the electricity consumption characteristics constituting the electricity consumption time series characteristic matrix may include total daily electricity consumption, maximum load and minimum load, and can respectively characterize the target electricity consumption enterprise's 24-hour electricity consumption scale, daily peak electricity consumption limit level and daily basic guaranteed load level; it also includes average load, load factor, peak-to-valley difference, coefficient of variation and daily load stability calculated based on at least one of the parameters of total daily electricity consumption, maximum load and minimum load.

[0027] Furthermore, the average load can be calculated based on the total daily electricity consumption and the duration of the electricity consumption data, reflecting the overall average consumption intensity of the target electricity-consuming enterprise throughout the day; the load factor can be calculated based on the average load and the maximum load, reflecting the overall fullness and utilization efficiency of the target electricity-consuming enterprise's electricity load; the peak-valley difference can be calculated based on the maximum load and the minimum load, reflecting the magnitude of the fluctuation in the target electricity-consuming enterprise's daily electricity consumption; the daily load stability can be the standard deviation of the daily load data, which can be determined based on the load data of multiple electricity consumption time series data, used to measure the degree of dispersion of the daily electricity load fluctuation; the coefficient of variation can be calculated based on the standard deviation of the daily load data and the average load, used to eliminate the interference caused by the mean, and used to reflect the relative fluctuation stability of the target electricity-consuming enterprise's electricity load.

[0028] In practical implementation, after determining that the classification controller has acquired enterprise characteristic information, the text data and electricity consumption data in the enterprise characteristic information can be preprocessed separately. For the text data, redundant and irrelevant characters can be removed first, and word segmentation can be completed based on a professional dictionary in the power field. At the same time, common stop words are removed, and a standardized and usable enterprise text word sequence is obtained after conversion. For the electricity consumption data, cleaning and correction can be carried out first to remove outliers, missing values, and interference data. Then, time alignment and standardization processing are performed to obtain multiple electricity consumption time series data and a multi-point daily load curve composed of multiple electricity consumption time series data. Key electricity consumption features are extracted from multiple electricity consumption time series data to construct an electricity consumption time series feature matrix that can completely represent the enterprise's electricity consumption behavior patterns and load characteristics. This step can remove noise from the enterprise text information, standardize the data format, and screen and extract effective features, providing standardized and high-quality data support for subsequent multi-source data processing, improving data processing efficiency and the accuracy of industry classification results.

[0029] S300: Input the enterprise text word sequence into a pre-trained text encoding model for semantic encoding to obtain text feature vectors, and input the electricity consumption time series feature matrix into a pre-trained time series encoder for time series encoding to obtain electricity consumption time series vectors; Among them, the pre-trained text encoding model can use labeled enterprise text data to build a training set to complete sample training in advance and form an encoder representation model with professional semantic understanding capabilities, so that the text encoding model can complete semantic encoding processing of the input enterprise text word sequence; the text feature vector can be a numerical vector obtained by semantic encoding of the enterprise text word sequence, which is used to condense the text semantic information that carries the enterprise's business operations and industry attributes.

[0030] Pre-trained text encoding models can employ the ModernBERT (Modern Bidirectional Encoder Representations from Transformers) model. The ModernBERT model is an encoder-type pre-trained language model built on the Transformer architecture, capable of handling sequences with a length of 8192 tokens. Compared to the BERT model, the ModernBERT model has made comprehensive improvements in network architecture, positional encoding methods, attention mechanisms, and pre-training task design. It can be applied to scenarios such as large-scale code search and long document retrieval, overcoming the inherent shortcomings of the BERT model, such as limited context length, insufficient semantic modeling capabilities, and low inference efficiency.

[0031] Furthermore, compared to the BERT model, the ModernBERT model uses RoPE positional encoding, which replaces traditional absolute positional encoding, to better capture the relative positional relationships and semantic dependencies between words in text data. It also introduces a lightweight attention optimization structure and a novel activation function to improve the model's semantic representation capabilities while reducing computational overhead and memory usage. Therefore, the ModernBERT model can extend the contextual input length of text data, fully supporting longer business texts with denser technical terms. Moreover, by relying on massive amounts of general and domain-specific texts for pre-training, it possesses stronger general semantic understanding capabilities and professional text adaptation capabilities, resulting in significantly better overall performance than the BERT model in tasks such as semantic representation and domain feature extraction in text data processing.

[0032] In practical implementation, the enterprise text word sequence is input into a pre-trained ModernBERT model for semantic encoding. This fully leverages the performance of the ModernBERT model in long-sequence text processing, specialized semantic mining, and fine-grained semantic representation. Since the enterprise text word sequence carries key textual data such as the target electricity-consuming enterprise's business scope, main business, industry attributes, and operating qualifications, and contains a large number of industry-specific terms and business descriptions, traditional BERT models are prone to information truncation and partial semantic understanding due to text length limitations. The ModernBERT model, however, can fully accept the entire enterprise text word sequence without losing contextual information. This model can automatically mine deep semantic relationships between words, distinguish attribute differences between similar business expressions, automatically filter invalid and redundant information and irrelevant interference in the text, and accurately extract semantic features that represent the enterprise's industry type and production and operation attributes. This results in outputting text feature vectors with high semantic recognizability and strong feature correlation.

[0033] Furthermore, semantic encoding using the ModernBERT model can map discretized enterprise text word sequences into text feature vectors with fixed dimensions and dense semantic information, achieving the transformation of unstructured text information into a vectorized feature space. The output text feature vectors can highly encompass the industry attributes and operational characteristics carried by the enterprise text, possessing high semantic discriminative power and strong feature robustness. This provides high-quality and uniform text representation support for subsequent feature vector fusion with electricity consumption time-series features and for classifying electricity-consuming enterprises by industry. Simultaneously, based on the excellent inference performance of the ModernBERT model itself, it ensures efficient processing of large-scale enterprise text data encoding and improves the adaptability of enterprise feature data to enterprise electricity consumption engineering application scenarios.

[0034] Among them, the pre-trained time encoder can be a dedicated encoding module pre-trained for the time-series data structure, adapted to the regular characteristics of the electricity consumption time-series data, so that the time encoder can perform time-series encoding on the input electricity consumption time-series feature matrix; the electricity consumption time-series vector can be a representation vector generated after the electricity consumption time-series feature matrix is ​​time-series encoded, used to condense and express the electricity load change pattern and electricity consumption behavior characteristics of the target electricity-consuming enterprise.

[0035] Pre-trained temporal encoders can employ LSTM (Long Short-Term Memory) encoders, suitable for sequence-to-sequence (Seq2Seq) encoding tasks. These encoders compress the input sequence into a fixed-dimensional context vector for subsequent decoder use. This LSTM encoder is a classic temporal deep learning model optimized iteratively based on traditional recurrent neural networks, addressing the vanishing and exploding gradient problems that recurrent neural networks often encounter when processing long temporal sequences. Through a three-gated structure (input gate, forget gate, and output gate), this model autonomously controls the memory, retention, and discarding of temporal information. It can effectively capture long-distance dependencies and patterns of change in time-series data, possessing strong temporal information memory and dynamic feature mining capabilities. It is suitable for modeling tasks involving continuous temporal correlations and complex fluctuations in sequence-type data, and applicable to scenarios such as temporal feature extraction, trend fitting, and state representation.

[0036] After pre-training with sufficient time series samples, the LSTM encoder can model general time series patterns and abstract features. It can also automatically learn the implicit change patterns, fluctuation amplitudes and periodic characteristics within time series data. In turn, it can compress and abstract the structured multidimensional time series matrix data, transforming high-dimensional and complex raw data such as time series feature matrices into dimensionally regular and information-concentrated representation vectors, providing standardized time series feature inputs for subsequent classification, fusion and discrimination tasks.

[0037] In practice, the electricity consumption time-series feature matrix is ​​input into a pre-trained LSTM encoder for time-series encoding, which fully leverages the advantages of the LSTM encoder in capturing long-term temporal dependencies and mining dynamic features. Since the electricity consumption time-series feature matrix carries load change data from multiple sampling points throughout the day, as well as various quantitative indicators such as total daily electricity consumption, maximum load, minimum load, and coefficient of variation, it contains the temporal fluctuation trend of the enterprise's daily electricity load, peak-valley variation patterns, and electricity consumption behavior patterns corresponding to production schedules. The data exhibits close temporal correlation and state continuity. Using the LSTM encoder, the dependencies between load data at different times can be mined layer by layer based on the time dimension, accurately capturing the periodic fluctuations, peak-valley switching rhythms, and stability changes of the electricity load. Simultaneously, it removes irregular minor disturbances and noise from the time-series data, retaining core temporal information that truly reflects the enterprise's electricity consumption habits and load operating status. This allows for the output of text feature vectors with relatively high semantic recognizability and temporal feature integration.

[0038] Furthermore, after the electricity consumption time-series feature matrix is ​​time-series encoded by the LSTM encoder, it can be mapped from a relatively complex feature matrix in terms of dimension and structure to a unified, highly concentrated time-series vector, realizing the transformation of multi-dimensional structured electricity consumption time-series data into a low-dimensional dense feature space. The generated electricity consumption time-series vector can fully reflect the overall electricity consumption pattern, load fluctuation characteristics, and electricity consumption patterns corresponding to production and operation of enterprises. It has excellent feature recognition and robustness, and can form good cross-modal adaptability with the text feature vector obtained by text encoding. This provides stable and reliable time-series representation support for subsequent multi-source feature fusion and accurate classification of electricity-consuming enterprise industry categories. At the same time, the lightweight structure of the LSTM encoder can ensure the operational efficiency of encoding and processing large batches of enterprise electricity consumption time-series data, thus adapting to the engineering needs of enterprise time-series features.

[0039] S400: Based on a preset attention fusion mechanism, the text feature vector and the electricity consumption time series vector are weighted and fused to determine multiple candidate industry categories and the probability value of each candidate industry category. Based on the probability values ​​of multiple candidate industry categories, the industry category of the target electricity-consuming enterprise is determined.

[0040] To achieve effective fusion of text feature vectors and electricity consumption time-series vectors for accurate classification in subsequent steps, a pre-defined attention fusion mechanism can be introduced in this step. As a multimodal feature fusion architecture with adaptive weight allocation, the attention fusion mechanism can autonomously assess the importance of information in the two types of vectors and dynamically complete the weight allocation so that the industry category of the target electricity-consuming enterprise and the probability value of the corresponding industry category can be calculated in the subsequent steps.

[0041] Furthermore, the attention fusion mechanism is used to generate a fusion feature of the textual semantic attributes and electricity consumption behavior characteristics of the target electricity-consuming enterprise. Based on this fusion feature, the probability values ​​of multiple candidate industry categories are calculated. In this way, the attention fusion mechanism can avoid the classification bias caused by single feature judgment, strengthen the complementary advantages of multi-source features, and improve the rationality and accuracy of industry category evaluation.

[0042] Furthermore, determining the industry category of the target electricity-consuming enterprise based on the probability values ​​of multiple candidate industry categories can align with the enterprise's actual operating attributes and electricity consumption behavior patterns, effectively improving the reliability and generalization performance of industry classification results. At the same time, utilizing the attention fusion mechanism, it possesses good adaptability and scalability, enabling it to adapt to batch fusion and classification scenarios of multi-source features of different electricity-consuming enterprises, thus meeting the requirements of practical engineering applications.

[0043] It should be noted that the attention weight allocation for the text feature vector and the electricity consumption time series vector is related to the effective state of the vectors. For example, when the text feature vector of the target electricity consumer is detected to be non-zero and the electricity consumption time series vector is zero, it indicates that the target electricity consumer lacks effective electricity consumption time series data, and therefore is likely a newly connected enterprise. In this case, the attention weights of the two types of features can be configured as [1, 0]. Similarly, when the text feature vector of the target electricity consumer is detected to be zero and the electricity consumption time series vector is non-zero, it indicates that the text data of the target electricity consumer has serious missing or data distortion problems. In this case, the attention weights can be configured as [0, 1].

[0044] In practical implementation, an attention fusion mechanism can be used to fuse and weight the obtained text feature vector and electricity consumption time series vector. The attention weights of the attention fusion mechanism are calculated to quantify the intrinsic correlation between the two features. Combined with the requirements of the industry classification task, the contribution ratio of the target electricity enterprise's text data and electricity consumption data to its classification results is adaptively evaluated. Features with high correlation and strong discriminative power are given higher weights to enhance the credibility of effective features, while weakening the interference from ineffective and redundant features. This completes the adaptive weighted fusion of two types of heterogeneous features and generates a fused feature vector. Simultaneously, the fused feature vector is input into a classifier associated with the attention fusion mechanism. Probability normalization is performed based on the Softmax activation function. Multiple candidate industry categories corresponding to the target electricity enterprise and the probability value of each candidate industry category are analyzed and calculated as the candidate industry categories and their matching probability values ​​for the target electricity enterprise. The predicted probability corresponding to each category is quantified and output. Then, based on the obtained probability values ​​of each industry category and each candidate industry category, the category with the highest fit with the target electricity enterprise is determined from multiple candidate industry categories as its final industry category. This step enables adaptive fusion of feature vectors from different modalities, fully leveraging the complementary advantages of text features and electricity consumption time-series features, avoiding classification bias caused by single feature judgment, and effectively improving the accuracy, stability, and generalization ability of industry classification results for electricity-consuming enterprises.

[0045] In summary, the method provided in this application preprocesses the acquired enterprise feature information to obtain enterprise text word sequences and electricity consumption time-series feature matrices. This process removes data noise from the enterprise feature information, standardizes the data format, and extracts effective feature data, thereby improving the overall quality and reliability of the basic data. By using a pre-trained text encoding model and encoder to encode and represent the enterprise text word sequences and electricity consumption time-series feature matrices respectively, text feature vectors and electricity consumption time-series vectors are obtained. This fully explores the semantic information of the text and the changing patterns of electricity consumption time-series data, generating highly recognizable and robust feature vectors. This achieves the transformation of multi-source data into a unified feature space, balancing representational effectiveness with engineering batch processing efficiency. By introducing and utilizing an attention fusion mechanism to adaptively weight and fuse the two types of feature vectors, multiple candidate industry categories and the probability value of each candidate industry category can be determined. This fully leverages the complementary advantages of multimodal features, weakens the interference of invalid features, and combines probabilistic discrimination to achieve accurate determination of enterprise industry categories, reducing classification bias caused by single features and improving the accuracy, stability, and generalization ability of the classification results. This method can fully tap the intrinsic value of multi-source data such as text data and electricity consumption data of electricity-consuming enterprises, avoid the limitations of single information judgment, and effectively improve the classification accuracy and reliability of enterprise industry categories. At the same time, it can be adapted to the large-scale and engineering application scenarios of large-scale enterprise data in the industry, thereby improving the accuracy of enterprise industry category positioning and the efficiency of subsequent electricity price determination.

[0046] In some embodiments, based on the content described in S400, the industry category of the target electricity-consuming enterprise is determined according to the probability values ​​of multiple candidate industry categories, including: S410: Compare the probability value of each candidate industry category with the preset first confidence threshold; The preset first confidence threshold can be a pre-defined probability judgment benchmark value, used to measure the credibility of the predicted probability of industry category, so as to select industry categories with the required confidence level as the basis for determining the industry affiliation of enterprises.

[0047] In practice, after obtaining multiple candidate industry categories and their corresponding probability values ​​through an attention fusion mechanism, a pre-configured and set first confidence threshold can be retrieved from a preset database. The probability values ​​of each candidate industry category are then compared with this first confidence threshold to quantitatively determine whether a suitable industry category exists for the target electricity-consuming enterprise. This step quantitatively assesses the reliability of each industry category result, avoiding biases caused by subjective judgment and providing a standardized and reliable basis for accurately determining the target electricity-consuming enterprise's industry category.

[0048] S420: In response to determining that there is a candidate industry category with a probability value greater than the first confidence threshold among multiple candidate industry categories, determine the candidate industry category with a probability value greater than the first confidence threshold as the target industry category, and determine the industry category of the target electricity-consuming enterprise based on the target industry category; The target industry category can be an industry category with a probability value greater than a preset first confidence threshold among multiple candidate industry categories. It can determine that the credibility of the corresponding candidate industry category meets the requirements and can be used as the basis for determining the industry category of the target electricity-consuming enterprise.

[0049] In practice, after comparing the probability values ​​of multiple candidate industry categories one by one with a preset first confidence threshold, if it is determined that there is a candidate industry category with a probability value greater than the first confidence threshold, it indicates that there is a candidate industry category that meets the credibility requirements and is suitable for the target electricity user. At this point, the candidate industry category with a probability value greater than the first confidence threshold can be identified as the target industry category, serving as the basis for determining the industry affiliation of the target electricity user. Subsequently, the identified target industry category can be matched with the target electricity user to determine the final industry category to which the target electricity user belongs, providing data support for subsequent electricity price assessment for the target electricity user. This step can accurately determine the valid industry category that meets the credibility requirements from multiple candidate industry categories, improving the efficiency and accuracy of determining the industry affiliation of the target electricity user.

[0050] S430: In response to determining that there is no candidate industry category with a probability value greater than the first confidence threshold among multiple candidate industry categories, sort the multiple candidate industry categories in descending order of probability value, and determine the industry category of the target electricity-consuming enterprise based on the sorted candidate industry category sequence.

[0051] In practice, after comparing the probability values ​​of multiple candidate industry categories one by one with a preset first confidence threshold, if the probability values ​​of all candidate industry categories are less than or equal to the first confidence threshold, it indicates that the target industry category that meets the requirements cannot be determined from the candidate industry categories based on the first confidence threshold. At this point, all candidate industry categories can be sorted in descending order according to the probability value ranking rule to form a regular industry category sequence. This sequence allows for further determination of the target electricity-consuming enterprise's industry category, ensuring the smooth progress of the target electricity-consuming enterprise's industry classification process and achieving refined identification of the target electricity-consuming enterprise's industry category. This step can reasonably infer the industry affiliation even when the overall confidence of the candidate industry categories is low, effectively ensuring the process integrity and adaptability of the overall industry classification method to real-world scenarios.

[0052] In some embodiments, based on the content described in S410, before comparing the probability value of each candidate industry category with a preset first confidence threshold, the method further includes: S401: Find multiple leaf nodes in the preset industry classification tree that correspond to multiple candidate industry categories, and obtain the number of samples for each leaf node. The pre-built industry classification tree can be a pre-constructed hierarchical system for electricity industry classification. The industry classification tree adopts a tree-like hierarchical architecture, dividing industry branches at all levels from top to bottom, and further subdividing the ends into leaf nodes that cannot be further divided. Each leaf node corresponds to a unique and specific subdivided industry category. Moreover, for each leaf node, the number of samples in the corresponding industry is statistically collected. In this way, the pre-built industry classification tree can be used for auxiliary verification and discrimination of subsequent industry affiliation, providing standardized hierarchical basis and data support for candidate industry category matching, node retrieval, and refined classification judgment.

[0053] In practice, after identifying multiple candidate industry categories through the attention fusion mechanism, each candidate industry category can be used as a retrieval basis to traverse the hierarchical structure of the preset industry classification tree. The bottom-level leaf nodes corresponding to each of the multiple candidate industry categories are located at the end of the tree. After locating the leaf nodes, the historical sample count corresponding to each matching leaf node can be retrieved and obtained one by one. This allows for the adjustment of the first confidence threshold based on the sample size of the leaf nodes, thereby determining whether there are sparse industry categories among the multiple candidate industry categories. Simultaneously, based on the sparsity of the industry categories within the preset industry classification tree, updating the first confidence threshold adjusts the verification strictness of the target industry category, further improving the accuracy of determining the target electricity-consuming enterprise's industry category.

[0054] S402: Adjust the first confidence threshold based on the number of samples in multiple leaf nodes.

[0055] In the pre-defined industry classification tree, the number of samples in a leaf node is positively correlated with the sparsity of its corresponding industry type. For leaf nodes with relatively few samples, the sparser the corresponding industry category samples, the greater the probability that it belongs to an emerging or cross-industry sector. To improve the representation effect of leaf nodes under sparse categories, EDA (Easy Data Augmentation) text augmentation technology is used to generate new training samples. Methods such as synonym replacement and random insertion are used to expand the sample data. Simultaneously, by using a thesaurus and randomly replacing non-stop words in the sample text, at least one new sample with semantic consistency but different expression can be generated for each category sample. This effectively expands the training data scale of sparse industry categories, continuously improves the industry classification tree, and avoids overfitting due to fixed word features.

[0056] In practice, after obtaining the sample size of leaf nodes corresponding to multiple candidate industry categories, based on a pre-defined industry classification tree, the industry category attribute can be determined according to the sample size of the leaf nodes, thereby identifying sparse industry categories with fewer samples from multiple candidate industry categories. Specifically, for leaf nodes with sufficient sample size, their corresponding candidate industry categories can be dense industry categories, so a stricter judgment standard can be adopted in the judgment process to improve the accuracy of identifying the target electricity-consuming enterprise's industry category. For leaf nodes with fewer sample size, their corresponding candidate industry categories are sparse industry categories, so the judgment and verification standards can be appropriately relaxed in the judgment process to prevent the wrong elimination of reasonable sparse industry categories due to scarce sample size. Among them, the pre-defined first confidence threshold serves as the benchmark for determining whether a candidate industry category can be used as the target industry category, so the first confidence threshold can be adjusted differently based on the actual sample size of each leaf node.

[0057] In some embodiments, based on the content described in S402, adjusting the first confidence threshold according to the number of samples in multiple leaf nodes includes: S4021: Compare the number of samples in multiple leaf nodes with a preset number threshold; The preset quantity threshold can be a pre-defined benchmark for judging the number of samples, used to determine the sample size of each leaf node in the industry classification tree, and to distinguish the corresponding industry category as a dense industry category or a sparse industry category based on the sample size of the leaf node.

[0058] In practice, after obtaining the sample count of the leaf nodes corresponding to each candidate industry category, the sample count of each leaf node can be compared and matched with a preset quantity threshold one by one. This allows for the quantitative differentiation of the sample size of each leaf node, determining the sparsity of the industry category to which each leaf node belongs. Consequently, it determines whether the corresponding candidate industry category is a dense or sparse industry category, providing a clear quantitative basis for subsequent adjustment of the first confidence threshold. In other words, this step can accurately distinguish between dense and sparse industry categories, providing a reliable discrimination standard for the adaptive adjustment of the first confidence threshold.

[0059] S4022: In response to determining that there are leaf nodes among multiple leaf nodes whose sample count is less than a preset threshold, lower the first confidence threshold. In practice, after comparing the sample size of each leaf node with a preset threshold, if a leaf node with a sample size less than the preset threshold is detected, it indicates that the leaf node belongs to a sparse industry category with few samples, and its corresponding candidate industry category is also determined to be a sparse industry category. To avoid erroneously eliminating sparse industry categories due to small sample size and low predicted probability values, a preset first baseline confidence level is used as the upper limit of the first confidence threshold. The current first confidence threshold is adaptively lowered to appropriately relax the admission criteria for sparse industry categories, ensuring that sparse industry categories that meet the actual attribution characteristics can be included in the corresponding judgment range. Therefore, this step can adaptively relax the judgment criteria for sparse industry categories with few samples, reduce misjudgments and omissions of sparse industry categories, and improve the rationality and accuracy of the target electricity-consuming enterprise industry classification results.

[0060] S4023: In response to determining that there are no leaf nodes among multiple leaf nodes with a sample count less than a preset threshold, the preset first baseline confidence level is used as the first confidence threshold.

[0061] In practice, after comparing the sample count of each leaf node with a preset threshold, if the sample count of all leaf nodes is not less than the threshold, it indicates that all candidate industry categories are dense industry categories with sufficient sample reserves, and there are no sparse industry categories with scarce samples. At this point, there is no need to relax the judgment criteria; the preset first baseline confidence level can be directly fixed as the currently used first confidence level threshold. All candidate industry categories are then screened and judged based on this first confidence level threshold, ensuring the rigor and reliability of the dense industry category classification results. Therefore, this step maintains the standard confidence level judgment rule for candidate industry categories belonging to dense industry categories, ensuring the rigor of the industry classification judgment criteria and guaranteeing the accuracy and stability of the industry classification results for electricity-consuming enterprises.

[0062] It should be noted that the preset first baseline confidence level can be a pre-set basic standard reference value, which can be used as a criterion for determining industry categories. When it is determined that there are no sparse industry categories among multiple candidate industry categories, the first baseline confidence level can be directly used to replace the formal first confidence level threshold. For example, the value range of the first baseline confidence level can be 0.6 to 0.9.

[0063] In some embodiments, based on the content described in S430, the industry category of the target electricity-consuming enterprise is determined according to the sorted candidate industry category sequence, including: S431: Based on the industry category sequence, compensate the probability values ​​of a preset number of candidate industry categories to obtain multiple compensated probability values ​​of candidate industry categories; For the candidate industry category sequence obtained after sorting in descending order, a fixed number of candidate industry categories with relatively high ranking can be selected according to a pre-set selection rule as the objects of subsequent probability value compensation. For example, the value range of this pre-set number can be set to 2 to 4.

[0064] In practice, all candidate industry categories are sorted in descending order of probability value, forming an ordered sequence of industry categories. Based on a pre-set number of categories, the top-ranked candidate industry categories are selected as the objects to be compensated. The probability values ​​of these selected categories are then compensated according to a pre-set compensation rule, resulting in the compensated probability values ​​for each candidate industry category. This step effectively corrects the probability values ​​of the top-ranked candidate industry categories, increases the probability differentiation between different industry categories, weakens the judgment bias caused by uneven sample distribution, and effectively improves the accuracy and rationality of the final industry category judgment results.

[0065] Furthermore, in this step, compensation methods such as adjacent difference amplification, temperature scaling, Bayesian uncertainty estimation, calibration based on sample local density / neighborhood, and Dirichlet calibration can be used to compensate for the probability values ​​of candidate industry categories.

[0066] The principle of the adjacent difference amplification compensation method is based on the arrangement order of the probability values ​​of candidate industry categories. It uses an amplification coefficient to uniformly stretch and adjust the probability differences between adjacent categories. While maintaining the original order of probability values, it widens the probability gap between different candidate industries, weakening the problem of convergence in probability values ​​for similar categories, and effectively improving the clarity and discernibility between different industry categories. For example, when using the adjacent difference amplification compensation method to compensate the probability values ​​of a preset number of candidate industry types, the original probability difference between every two adjacent candidate industry categories in the sequence can be calculated sequentially. A fixed difference amplification coefficient is set, and this coefficient is used to proportionally stretch all adjacent probability differences. Then, the probability values ​​corresponding to each candidate industry category are recalculated based on the stretched differences, maintaining the original order of the candidate industry categories' probability values, resulting in compensated probability values ​​with larger intervals and more obvious distinguishing effects.

[0067] Temperature scaling is a simple and efficient probability calibration method. By introducing a temperature parameter to smoothly scale the original output logic value of the model, it can correct the problem of the model's predicted probability being generally too high or too low, reduce the deviation between the predicted probability and the true confidence level, and adapt to the problem of probability distribution imbalance in industry classification scenarios. The temperature coefficient, as a learnable parameter, can assist the model's output logic value in scaling and achieve the result output by relying on softmax. For example, when using temperature scaling to compensate the probability value of a preset number of candidate industry types, the probability value of the candidate industry type can be determined by combining the probability value of the candidate industry type and the temperature parameter, so as to compress excessively high probability values, increase excessively low probability values, correct the confidence deviation in the model prediction process, and complete the overall probability compensation work.

[0068] Bayesian uncertainty estimation primarily relies on Bayesian statistical principles. It combines industry categories and dynamically adjusts the original predicted probabilities based on the degree of uncertainty, thereby improving the rationality of probability predictions for sparse industry categories. For example, when using Bayesian uncertainty to compensate for the probability values ​​of a preset number of candidate industry types, historical distribution data for each industry category can be retrieved. The uncertainty of the classification results can be calculated by combining this data with sample information from the candidate industry categories. Correction weights are then set according to the level of uncertainty, appropriately lowering the probability values ​​of candidate industry categories with excessively high uncertainty and appropriately increasing the probability values ​​of candidate industry categories with excessively low uncertainty. This numerical correction is then performed using Bayesian inference, generating the compensated probability values ​​for the candidate industry categories.

[0069] Calibration based on local sample density / neighborhood relies on the spatial distribution patterns of samples within the same industry category. It determines the confidence level of a sample classification by assessing the density of neighboring samples within the feature space, and then uses sample cluster density to perform probability calibration compensation. For example, when using this calibration method to compensate the probability values ​​of a preset number of candidate industry types, samples of the same category and neighborhood can be retrieved from the feature space. The local cluster density of these samples is then calculated. Higher sample density indicates higher confidence in the industry classification, leading to a slight upward adjustment of the corresponding probability; conversely, lower sample density indicates lower confidence in the industry classification, leading to an appropriate downward adjustment of the corresponding probability. Calibration is performed one by one according to the neighborhood density to obtain the compensated probability values ​​for each candidate industry category.

[0070] Dirichlet calibration utilizes the Dirichlet distribution to fit the overall probability distribution of multiple industry categories. It adapts to the interconnected probability changes of various candidate industry categories under different classification tasks, correcting probability crowding and uneven distribution issues that occur when multiple candidate industry categories are predicted simultaneously. This makes the overall probability distribution of candidate industry categories more closely reflect actual business distribution. For example, when using Dirichlet calibration to compensate for the probability values ​​of a preset number of candidate industry types, it fits the overall probability distribution trend of all candidate industries within the preset number based on the Dirichlet distribution. After fitting, the distribution parameters are adjusted according to the industry category sequence order, re-optimizing the probability distribution relationship between multiple candidate industry categories, correcting probability crowding bias between industry categories, and outputting the probability values ​​of candidate industry categories after overall distribution calibration.

[0071] S432: Compare the probability value of each compensated candidate industry category with the preset second confidence threshold; The preset second confidence threshold can be a pre-set probability judgment benchmark value, which is used to compare with the probability values ​​of each candidate industry category after compensation, so as to further determine the target industry category that can be used as the final judgment result from the preset number.

[0072] In practice, after obtaining multiple compensated probability values ​​for candidate industry categories, a preset second confidence threshold can be used as the judgment benchmark value for a preset number of candidate industry categories. By comparing the probability value of each compensated candidate industry category with the second confidence threshold one by one, it is verified whether the probability value of each compensated candidate industry category meets the preset judgment criteria. This determines whether there are candidate industry categories that meet the score conditions, while filtering out invalid candidate industry categories that do not meet the score conditions. This step enables quantitative screening of the compensated probability values, accurately eliminating invalid candidate industry categories that do not meet the score conditions, and ensuring the reliability and standardization of the industry category judgment results.

[0073] S433: In response to the determination that there is at least one candidate industry category with a probability value greater than the second confidence threshold among the candidate industry categories after compensation, the candidate industry category with a probability value greater than the second confidence threshold is determined as the target industry category, and the industry category of the target electricity-consuming enterprise is determined based on the target industry category.

[0074] In practice, after comparing the probability values ​​of each compensated candidate industry category with a preset second confidence threshold one by one, if at least one candidate industry category with a probability value greater than the preset second confidence threshold is detected among multiple compensated candidate industry categories, then the corresponding candidate industry category is determined to meet the industry classification requirements. At this point, the candidate industry category with a probability value greater than the second confidence threshold can be taken as the target industry category, and the determined target industry category is matched with the target electricity-consuming enterprise to determine the final industry category to which the target electricity-consuming enterprise belongs, thereby providing data support for subsequent electricity price assessment for the target electricity-consuming enterprise. Therefore, this step, based on the second confidence threshold, can accurately determine the target industry category from a preset number of candidate industry categories, so as to quickly lock in the true industry classification of the enterprise and improve the effectiveness and reliability of industry classification determination.

[0075] It should be noted that, among the preset number of candidate industry categories, if only one compensated candidate industry category has a probability value greater than the second confidence threshold, then that candidate industry category will be directly determined as the target industry category; if there are two or more compensated candidate industry categories with probability values ​​greater than the second confidence threshold, then one of the candidate industry categories that meets the conditions will be randomly selected as the final target industry category.

[0076] In some embodiments, based on the content described in S432, before comparing the probability value of each compensated candidate industry category with a preset second confidence threshold, the method further includes: S4301: Determine the probability difference between the probability values ​​of any two candidate industry categories in the preset quantity; Among a preset number of candidate industry categories, the probability value of each category is used to characterize the confidence level of the corresponding industry affiliation. The smaller the probability difference between two candidate industry categories, the more similar the confidence levels and the more consistent the judgment tendencies when using the preset attention mechanism to affix the two categories. This also reflects a high degree of similarity in business characteristics, electricity consumption characteristics, and attribute characteristics between the two candidate industry categories, making it difficult for the attention fusion mechanism to effectively distinguish their differences. Therefore, a relatively small probability difference between two candidate industry categories indicates a relatively high probability that they are a similar industry pair.

[0077] In practice, after obtaining a preset number of candidate industry categories, any two different candidate industry categories are selected through iteration to extract their probability values. The probability difference between the two probability values ​​is then calculated. The probability difference between two candidate industry categories can quantify the degree of industry difference between them, verifying whether there are similar industry pairs among the preset number of candidate industry categories, thus providing reliable data support for the accurate determination of subsequent industry categories.

[0078] S4302: In response to determining that there is at least one probability difference less than a preset probability difference threshold, increase the second confidence threshold; The preset probability difference threshold can be a pre-defined fixed numerical benchmark, which can serve as a critical boundary for evaluating the probability difference between two candidate industry categories. When the probability difference between any two sets of candidate industry categories is less than this value, it can be determined that their classification confidence is highly similar, their industry characteristics are highly similar, and they constitute a similar industry pair.

[0079] In practice, after calculating the probability difference between any two candidate industry categories within a preset range, each probability difference is compared sequentially with a preset probability difference threshold. If any probability difference is found to be less than the preset threshold, it indicates the existence of similar industry pairs among the candidate industry categories within the preset range. The confidence levels of these similar industry pairs are too close, resulting in low feature differentiation and making identification difficult. In this case, according to a preset adjustment rule, a second baseline confidence level can be used as the lower limit of the second confidence threshold to appropriately increase the threshold, raising the entry standard for industry category identification and rigorously screening eligible candidate industry categories to prevent misjudgments due to the presence of similar industry pairs. This step accurately identifies similar industry pairs among candidate industry categories, raises the judgment standard for similar industry pairs, avoids confusion caused by similar industry pairs, and further improves the accuracy of industry category identification for target electricity-consuming enterprises.

[0080] S4303: In response to determining that all probability differences are greater than or equal to a preset probability difference threshold, a preset second baseline confidence level is used as the second confidence level threshold.

[0081] In practice, after calculating the probability difference between any two candidate industry categories within a preset range, each probability difference can be compared sequentially with a preset probability difference threshold. If all probability differences are greater than or equal to the preset probability difference threshold, it indicates that there are no similar industry pairs among the candidate industry categories within the preset range, the probability scores between candidate industry categories are significantly different, and the industry feature differentiation is good. Therefore, there is no need to adjust the judgment criteria. At this point, the preset second baseline confidence level can be directly used as the second confidence level threshold to facilitate the subsequent comparison and screening process of probability values ​​for subsequent candidate industry categories. This step ensures that the industry category screening rules remain stable and consistent when the differentiation between candidate industry categories is significant, balancing judgment efficiency and industry identification reliability.

[0082] In some embodiments, enterprise characteristic information includes text data, wherein the text data is electricity consumption data that reflects the enterprise's electricity consumption behavior patterns and electricity load characteristics; for example, the text data may include, but is not limited to, the enterprise name, business scope, main business, registration information and business qualifications.

[0083] Based on the content described in S100, the enterprise feature information is preprocessed to obtain the enterprise text word sequence, including: S110: Clean the text data, segment it, and remove stop words to obtain the processed text data; In practice, the obtained raw text data can be cleaned to remove garbled characters, special symbols, invalid punctuation, redundant spaces and meaningless redundant characters, and to standardize the text format and content. Then, the cleaned text data is segmented into words according to industry-specific semantics and text sentence logic, dividing the continuous text into independent words and professional terminology units.

[0084] Furthermore, a pre-set stop word list can be loaded, and the segmentation results can be traversed to filter out commonly used function words, auxiliary words, and general high-frequency meaningless words that do not contribute to actual semantics, thus removing redundant and interfering words. In this way, after the processing of cleaning, segmentation, and stop word removal, invalid noise information can be eliminated, and effective words with semantic representation capabilities can be retained, ultimately resulting in well-organized, low-noise text data that can be used for feature extraction in subsequent models.

[0085] S120: Transform and sort the processed text data to obtain the enterprise text word sequence.

[0086] In practice, after obtaining the processed text data, the words in the processed text data can be formatted to unify the vocabulary expression. Then, the words are sorted according to the degree of semantic relevance, frequency of occurrence, or contextual order to organize the messy and scattered words into a coherent and orderly word arrangement structure, so as to form a relatively standardized corporate text word sequence, so as to facilitate the subsequent model to read, understand the semantics of the text, and extract effective features.

[0087] In some embodiments, enterprise characteristic information includes electricity consumption data; the electricity consumption data can reflect the patterns of enterprise electricity consumption behavior and the characteristics of electricity load, so as to provide reliable data support for the industry classification of target electricity-consuming enterprises.

[0088] Based on the content described in S100, the enterprise characteristic information is preprocessed to obtain the electricity consumption time-series vector matrix, including: S130: Divide and organize the electricity consumption data according to the preset time granularity to obtain multiple electricity consumption time series data; In practice, a preset fixed time interval can be used as the time granularity (e.g., 15 minutes). Based on this time granularity, the collected continuous electricity consumption data can be divided into multiple electricity consumption data segments with the same time period and independent of each other. At this time, missing data and abnormal fluctuation data in the electricity consumption data segments can be supplemented and corrected, and abnormal values ​​with obvious deviations can be removed to obtain multiple sets of standardized electricity consumption time series data, which is convenient for subsequent analysis and processing of electricity consumption characteristics through preset models.

[0089] S140: Extract electricity consumption features from each electricity consumption time series data, and determine the electricity consumption time series feature matrix based on the obtained multiple electricity consumption features.

[0090] In practice, after acquiring multiple electricity consumption time-series data, relatively key electricity consumption features can be extracted from each segmented and regularized electricity consumption time-series data. After acquiring the electricity consumption features of each electricity consumption time-series data, the electricity consumption features corresponding to each set of time-series data can be arranged sequentially and integrated according to a fixed row and column structure. All extracted electricity consumption features are then organized and summarized to construct a regular electricity consumption time-series feature matrix, so as to transform the scattered electricity consumption patterns into a structured feature matrix, which facilitates the model to quickly learn and identify the electricity consumption behavior characteristics of enterprises.

[0091] It should be noted that the method in this embodiment can be executed by a single device, such as a computer or server. The method can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method in this embodiment, and the multiple devices will interact with each other to complete the method described.

[0092] It should be noted that the above description describes some embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0093] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides an industry classification device based on multi-source data fusion.

[0094] like Figure 2 As shown, the industry classification device based on multi-source data fusion includes a data acquisition module 11, a data preprocessing module 12, a data encoding module 13, and a type determination module 14. The data acquisition module is configured to acquire the enterprise characteristic information of the target electricity-consuming enterprise. The data preprocessing module is configured to preprocess enterprise feature information to obtain enterprise text word sequences and electricity consumption time series feature matrices; The data encoding module is configured to input the enterprise text word sequence into a pre-trained text encoding model for semantic encoding to obtain a text feature vector, and input the electricity consumption time series feature matrix into a pre-trained time series encoder for time series encoding to obtain an electricity consumption time series vector. The type determination module is configured to perform weighted fusion of text feature vectors and electricity consumption time series vectors based on a preset attention fusion mechanism, determine multiple candidate industry categories and the probability value of each candidate industry category, and determine the industry category of the target electricity-consuming enterprise based on the probability values ​​of multiple candidate industry categories.

[0095] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware.

[0096] The apparatus described above is used to implement the corresponding industry classification method based on multi-source data fusion in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0097] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the industry classification method based on multi-source data fusion as described in any of the above embodiments.

[0098] Figure 3 This embodiment illustrates a more specific hardware structure of an electronic device. The device may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0099] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0100] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0101] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0102] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0103] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0104] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0105] The electronic devices described above are used to implement the corresponding industry classification methods based on multi-source data fusion in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0106] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the industry classification method based on multi-source data fusion as described in any of the above embodiments.

[0107] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0108] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the industry classification method based on multi-source data fusion as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0109] Based on the same concept, corresponding to any of the above embodiments, this application also provides a computer program product, including computer program instructions, which, when run on a computer, cause the computer to perform the method described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0110] It is understood that before using the technical solutions of the various embodiments in this application, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.

[0111] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose, based on the prompt message, whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media performing the operations described in this application.

[0112] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0113] It is understood that the above notification and user authorization process is merely illustrative and does not limit the implementation of this application. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this application.

[0114] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0115] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0116] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0117] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the claims of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. An industry classification method based on multi-source data fusion, characterized in that, include: Obtain enterprise characteristic information of the target electricity-consuming enterprise; The enterprise feature information is preprocessed to obtain the enterprise text word sequence and electricity consumption time series feature matrix; The enterprise text word sequence is input into a pre-trained text encoding model for semantic encoding to obtain a text feature vector, and the electricity consumption time series feature matrix is ​​input into a pre-trained time series encoder for time series encoding to obtain an electricity consumption time series vector. The text feature vector and the electricity consumption time series vector are weighted and fused based on a preset attention fusion mechanism to determine multiple candidate industry categories and the probability value of each candidate industry category. The industry category of the target electricity-consuming enterprise is determined based on the probability values ​​of the multiple candidate industry categories.

2. The method according to claim 1, characterized in that, Determining the industry category of the target electricity-consuming enterprise based on the probability values ​​of the multiple candidate industry categories includes: The probability value of each candidate industry category is compared with the preset first confidence threshold. In response to determining that there is a candidate industry category among the plurality of candidate industry categories with a probability value greater than the first confidence threshold, the candidate industry category with a probability value greater than the first confidence threshold is determined as the target industry category, and the industry category of the target electricity-consuming enterprise is determined based on the target industry category; In response to determining that there is no candidate industry category among the plurality of candidate industry categories with a probability value greater than the first confidence threshold, the plurality of candidate industry categories are sorted in descending order of probability value, and the industry category of the target electricity-consuming enterprise is determined based on the sorted candidate industry category sequence.

3. The method according to claim 2, characterized in that, Before comparing the probability value of each candidate industry category with a preset first confidence threshold, the process further includes: Find multiple leaf nodes in the preset industry classification tree that correspond to the multiple candidate industry categories, and obtain the number of samples for each leaf node; The first confidence threshold is adjusted based on the number of samples in the plurality of leaf nodes.

4. The method according to claim 3, characterized in that, The step of adjusting the first confidence threshold based on the number of samples in the plurality of leaf nodes includes: The sample count of the multiple leaf nodes is compared with a preset quantity threshold; In response to determining that there are leaf nodes among the plurality of leaf nodes whose sample count is less than a preset threshold, the first confidence threshold is lowered. In response to determining that there are no leaf nodes among the plurality of leaf nodes with a sample count less than a preset threshold, a preset first baseline confidence level is used as the first confidence threshold.

5. The method according to claim 2, characterized in that, Determining the industry category of the target electricity-consuming enterprise based on the sorted candidate industry category sequence includes: Based on the industry category sequence, the probability values ​​of a preset number of candidate industry categories are compensated to obtain multiple compensated probability values ​​of candidate industry categories. The probability value of each compensated candidate industry category is compared with a preset second confidence threshold. In response to determining that there is at least one candidate industry category with a probability value greater than the second confidence threshold among the compensated candidate industry categories, the candidate industry category with a probability value greater than the second confidence threshold is determined as the target industry category, and the industry category of the target electricity-consuming enterprise is determined based on the target industry category.

6. The method according to claim 5, characterized in that, Before comparing the probability value of each compensated candidate industry category with a preset second confidence threshold, the method further includes: Determine the probability difference between the probability values ​​of any two candidate industry categories from a preset number; In response to determining that there is at least one probability difference less than a preset probability difference threshold, the second confidence threshold is increased; In response to determining that all probability differences are greater than or equal to a preset probability difference threshold, a preset second baseline confidence level is used as the second confidence threshold.

7. The method according to claim 1, characterized in that, The enterprise characteristic information includes text data. The step of preprocessing the enterprise feature information to obtain the enterprise text word sequence includes: The text data is cleaned, segmented, and stop words are removed to obtain the processed text data; The processed text data is then transformed and sorted to obtain a sequence of enterprise text words.

8. The method according to claim 1, characterized in that, The enterprise characteristic information includes electricity consumption data; The step of preprocessing the enterprise feature information to obtain the electricity consumption time-series vector matrix includes: The electricity consumption data is divided and organized according to a preset time granularity to obtain multiple electricity consumption time series data. Electricity consumption features are extracted from each electricity consumption time series data, and an electricity consumption time series feature matrix is ​​determined based on the obtained multiple electricity consumption features.

9. An industry classification device based on multi-source data fusion, characterized in that: include: The data acquisition module is configured to acquire enterprise characteristic information of the target electricity-consuming enterprise; The data preprocessing module is configured to preprocess the enterprise feature information to obtain the enterprise text word sequence and electricity consumption time series feature matrix; The data encoding module is configured to input the enterprise text word sequence into a pre-trained text encoding model for semantic encoding to obtain a text feature vector, and to input the electricity consumption time series feature matrix into a pre-trained time series encoder for time series encoding to obtain an electricity consumption time series vector. The type determination module is configured to perform weighted fusion of the text feature vector and the electricity consumption time series vector based on a preset attention fusion mechanism to determine multiple candidate industry categories and the probability value of each candidate industry category, and determine the industry category of the target electricity-consuming enterprise based on the probability values ​​of the multiple candidate industry categories.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.

Images