Intelligent association recommendation method and system based on power marketing data, and storage medium

Abstract

A kind of intelligent association recommendation method based on power marketing data, comprising: data acquisition is carried out to power system marketing data;Similar association calculation is carried out to the data collected, and the association data set of the data of different departments in marketing data is obtained;Association data set is collected, and the text of association data set is converted into two-dimensional matrix, for the data input of gated recurrent-attention mechanism model model;Two-dimensional matrix data input gated recurrent-attention mechanism model is trained, and the output two-dimensional matrix is converted back to text format, find the text keyword of association data set, and write into dictionary and store;Staff input the data that wants to query according to business requirement, and automatically recommend matching dictionary by SQL statement matching.The advantage is: the method is automatically identified by similar association calculation and gated recurrent-attention mechanism model, and the association between the data of different departments is established, to realize the intelligent recommendation of data demand for staff.

Description

Technical Field

[0001] This invention relates to an intelligent association recommendation method, system, and storage medium based on electricity marketing data. Background Technology

[0002] Currently, cross-business related data tables stored in the power system data center, such as the business expansion application form under customer service, require manual input. This task is extremely large, and staff still use field matching technology when querying data, which sometimes results in staff not finding the data they want.

[0003] Especially in the field of electricity marketing, power companies frequently need to process large amounts of data, including sales records, customer information, and market trends. However, this data is stored in a highly fragmented manner within the data platform, leading to difficulties in retrieving and utilizing the data. To address this issue, there is an urgent need to develop an intelligent association and recommendation algorithm for electricity marketing data. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an intelligent association recommendation method, system and storage medium based on electricity marketing data. The method automatically identifies and establishes the association relationship between data through algorithms and models to realize intelligent recommendation for staff to meet their data needs.

[0005] The technical solution of this invention is: A smart association recommendation method based on electricity marketing data includes: Collect marketing data from the power system; Similarity correlation calculations are performed on the collected data to obtain the associated datasets of data from different departments within the marketing data. Collect associated datasets and convert the text in the associated datasets into two-dimensional matrices for use as data input to the gated recurrent attention machine model; The two-dimensional matrix data is input into the gated recurrent attention mechanism model for training. The output two-dimensional matrix is ​​converted back into text format, the text keywords of the associated dataset are found, and the found text keywords are written into a dictionary and stored. Input the data you want to query based on your business needs, and the system will automatically recommend matching dictionaries through SQL statement matching.

[0006] Furthermore, the power system marketing data includes structured data, semi-structured data, and unstructured data.

[0007] Furthermore, similarity association calculations are performed on the collected data, including: The cosine similarity algorithm is used to calculate the similarity within a single business category of the collected data, and to obtain text with high similarity for a single business category. Using the Apriori algorithm, association analysis is performed on highly similar texts of single business data in different business categories to obtain an association dataset of data from different departments in the marketing data.

[0008] Further optimization involves using a cosine similarity algorithm to calculate the similarity within a single business category of the collected data, including: Redundant information in electricity marketing data was removed. Based on the term frequency or term frequency-inverse document frequency algorithm, single types of business data are unified into a vector space of the same length; By using the features of individual business data and vector weighting, a weighted calculation is performed to obtain text with high similarity to individual business data.

[0009] Further optimization of the Apriori algorithm includes: Frequent itemset generation based on a single character; Frequent itemset generation based on candidate itemsets; Repeat the generation of frequent itemsets based on candidate itemsets until no new frequent itemsets are generated; Generate association rules, synthesize text datasets with values ​​above the minimum confidence threshold into data packets, and store them to obtain association datasets of data from different departments within the marketing data. A system for an intelligent association recommendation method based on electricity marketing data includes: The data acquisition module collects marketing data from the power system. The data association module performs similarity association calculations on the collected data to obtain the associated datasets of data from different departments in the marketing data. The associated data collection module collects associated datasets and converts the text of the associated datasets into two-dimensional matrices for use as data input to the gated recurrent attention mechanism model. The associated data text keyword generation module is trained using a gated recurrent attention mechanism model to find text keywords in the associated dataset, and writes the found text keywords into a dictionary and stores them. The data query module allows users to input the data they want to query based on business needs. It then automatically recommends matching dictionaries based on SQL statements.

[0010] A readable computer storage medium storing a computer program that, when executed by a processor, implements the aforementioned intelligent association recommendation method based on electricity marketing data.

[0011] The beneficial effects of this invention are: This method automatically identifies and establishes relationships between data from different departments through similarity association calculation and a gated loop-attention mechanism model, enabling intelligent recommendations for staff based on their data needs. The similarity association calculation employs the Apriori algorithm, which uses a step-by-step approach to increase the number of items and leverages the properties of frequent itemsets for pruning, thereby reducing the generation and computational load of candidate itemsets and improving the algorithm's efficiency. Attached Figure Description

[0012] Figure 1 This is a flowchart of the intelligent association recommendation method based on electricity marketing data according to the present invention; Figure 2 This is a system structure diagram of the present invention. Detailed Implementation

[0013] Example 1 The intelligent association recommendation method based on electricity marketing data includes the following steps: S1: Data collection for power system marketing data Electricity marketing data was collected from the Fuxin Power Bureau, including structured data, semi-structured data, and unstructured data. (1) Structured data: Structured data is data organized according to a fixed pattern and format. It usually exists in the form of rows and columns in tables or databases. In electricity marketing data, the following data belongs to structured data: Customer Information Form: Includes fields for customer's name, address, contact information, and electricity requirements; Sales record table: includes fields for sales date, sales amount, product type, and sales channel; Market Trends Report: Includes sales data, market share, and growth rate indicators for the specified time period; (2) Semi-structured data: Semi-structured data is data that has a certain structure but does not conform to the format requirements of traditional relational databases; in electricity marketing data, the following data belongs to semi-structured data: Emails and documents: These include email communications related to sales activities, quotations, and contract documents. This data may contain structured fields (such as sender, recipient, and date) as well as unstructured text content. Log files: contain system logs and sales personnel's operation records. This data is usually in text format, but may contain some structured fields. (3) Unstructured data: Unstructured data is data without a fixed structure and format, and usually exists in the form of free text, images, audio or video; in electricity marketing data, the following data belongs to unstructured data: Customer feedback and comments: This includes opinions and feedback provided by customers on social media, online forums, or surveys. This data can be in free text form and is difficult to extract and analyze directly.

[0014] S2 performs similarity correlation calculations on the collected data. S2.1 uses the cosine similarity algorithm to calculate the similarity within a single business category of the collected data. The calculation process is as follows: S2.1.1 Electricity marketing data is a type of data with strong standardization, usually composed of a combination of letters and numbers. Of course, many equipment or customer information may include additional information, such as user identifier, user number, user name, equipment name, meter box information, and capacitor meter information. Therefore, the data needs to be preprocessed to remove redundant information that does not need to participate in matching fields. The following steps are taken when removing redundant information from electricity marketing data: (1) Data backup: Before deleting redundant information, back up the original data first; this way, the data can be restored when needed, avoiding data loss due to accidental deletion; (2) Determine redundant information based on the meaning of the fields: Based on the meaning of each field, determine which fields contain redundant information; redundant information may be repetitive, irrelevant or unnecessary content; for example, the equipment name field contains redundant information such as equipment model and manufacturer. (3) Use string processing methods: Because redundant information in the device name or user information field always starts or ends with specific characters or keywords, use string truncation methods to delete redundant parts; S2.2.2. Based on the term frequency (TF-IDF) algorithm, term frequency statistics are performed on single-type business data of electricity marketing data (using metering box data as an example in this embodiment). Term frequency is a relative concept; for data, there is no strong semantic connection between characters, so each character can be treated as a word for term frequency statistics. In this embodiment, using metering box data as an example, the data length for each metering box model is inconsistent, resulting in an inconsistent vector space. The term frequency statistics results cannot be directly used as input for calculation. To unify all metering box model data into a vector space of the same length, the length of the vector space is determined by the longest field with the highest term frequency. Simultaneously, for ease of calculation, a fixed mapping relationship is designed to map the model data to the vector space. When the string lengths of two metering box specifications differ significantly, they are highly likely not the same metering box. Therefore, the length of the metering box model is used as the first dimension of the text vector, and the vector space is as follows: in The vector space has one dimension: character frequency. The entire space is 30-dimensional, with the first dimension being the length of the meter box model data. The following dimension represents the word frequency statistics of the bin data field. If a field has only 20 dimensions, then the information from the 21st to the 30th dimensions is recorded as 0. This completes the vectorization operation of the bin data, ensuring the length of the vector and the word frequency statistics results. S2.2.3 Weight Calculation Metering box model data features and vector weighting: Frequency analysis of metering box model data is sensitive not only to the positional information of the model data but also to the relative order of characters. The original component model data corresponding to the same frequency vector may be completely different. Specifically, two model numbers, "GHYHY27395" and "GHYYH27395", have the same frequency vector, but represent two completely different metering boxes. Therefore, to address this, a weighting method is proposed for the metering box model data vector based on frequency statistics. The weight is related to the position of the characters in the metering box model data; the earlier the character, the higher its weight. After a certain position, subsequent characters do not need further weighting, which is due to the attributes of the metering box model data. The specific weighting steps are as follows: Taking "GHYYH27395" as an example, the weighting method described above is used for calculation, and the steps are as follows: (1) Determine the weighting range: Select the first 5 characters of the weighted meter model data; (2) Weighting: Use a linear decay function to assign weights to characters within the weighted range, so that the characters at the beginning have higher weights; For the meter model data "GHYYH27395", the weighted range is the first 5 characters, and the weighting is assigned as follows: The weight of the first character 'G' is 1.0. The weight of the second character 'H' is 0.8. The weight of the third character 'Y' is 0.6. The weight of the fourth character 'Y' is 0.4. The weight of the fifth character 'H' is 0.2. (3) Standardized weights: The assigned weights are standardized to ensure that the sum of the total weights is 1; in the above process, the total weights are 1.0 + 0.8 + 0.6 + 0.4 + 0.2 = 3.0, so the standardized weights are: The normalized weight of the first character 'G' is 1.0 / 3.0 = 0.333. The normalized weight of the second character 'H' is 0.8 / 3.0 = 0.267. The normalized weight of the third character 'Y' is 0.6 / 3.0 = 0.200. The normalized weight of the fourth character 'Y' is 0.4 / 3.0 = 0.133. The normalized weight of the fifth character 'H' is 0.2 / 3.0 = 0.067. The final weights are [0.333, 0.267, 0.200, 0.133, 0.067]; (4) To distinguish the relative order of character positions, the character sequence is converted into one-hot encoding; in this way, each character will have a unique code, and the order of the codes is consistent with the order of the characters in the original meter box model data; For "GHYYH27395", the one-hot encoding is as follows: The one-hot encoding of the first character 'G' is [1, 0, 0, 0, 0]. The one-hot encoding of the second character 'H' is [0, 1, 0, 0, 0]. The one-hot encoding of the third character 'Y' is [0, 0, 1, 0, 0]. The one-hot encoding of the fourth character 'Y' is [0, 0, 1, 0, 0]. The one-hot encoding of the fifth character 'H' is [0, 1, 0, 0, 0]. (5) Calculate the weighted vector: Multiply each character of the one-hot encoding by its corresponding weight, and sum the results to obtain the weighted vector; For "GHYYH27395", the weighted vector is calculated as follows: Weighted vector = (First character in one-hot encoding × First character weight) + (Second character in one-hot encoding × Second character weight) + ... Weighted vector = ([1, 0, 0, 0, 0] × 0.333) + ([0, 1, 0, 0, 0] × 0.267) + ([0, 0, 1, 0, 0] × 0.200) + ([0, 0, 1, 0, 0] × 0.133) + ([0, 1, 0, 0, 0] × 0.067) Then the dimensional information in the vector space Replace it with the weighted value obtained in this step ((first character of one-hot encoding × first character weight) + (second character of one-hot encoding × second character weight) + ...), thus obtaining the weighted vector; S2.2.4 Calculate similarity: The two data points are respectively , After word frequency statistics and weighting, word frequency-weighted vectors are obtained respectively. and The cosine similarity between two data points is denoted as... The calculation formula is as follows: in, Representing vectors with vector dot product, and Representing vectors respectively with vector The similarity value is the length of the data. The closer the similarity value is to 1, the more similar the text features of the two data points are; the closer the similarity value is to 0, the lower the similarity between the two data points.

[0015] S2.2.5 Interpretation of similarity results: A similarity value close to 1 indicates that the two vectors are very similar. A similarity value close to -1 indicates that the two vectors are very dissimilar. A similarity value close to 0 indicates that the two vectors have low similarity. S2.2 uses the Apriori algorithm to perform association analysis on highly similar texts of single business data in different business categories, and obtains the association dataset of data from different departments in the marketing data.

[0016] In S2.1, the similarity values ​​of texts in the same type of data were obtained. S2.2, based on the high similarity texts obtained in S2.1, performs correlation analysis (i.e., the Apriori algorithm) on high similarity data in different categories of data (such as meter box information and user electricity consumption information). The Apriori algorithm has the following steps: (1) Generation of frequent itemsets based on a single item (character) Scan the dataset and count the support (frequency of occurrence) of each individual item. Based on the set minimum support threshold, items with support higher than the threshold are selected to form candidate frequent 1-itemsets; (2) Frequent itemset generation based on candidate itemsets For candidate frequent k-itemsets, generate candidate frequent (k+1)-itemsets through a join operation; For each candidate frequent (k+1)-itemset, check if all its k-item subsets are frequent; if not, discard the candidate itemset. Scan the dataset and calculate the support for each candidate frequent (k+1) itemset; Based on the set minimum support threshold, items with support higher than the threshold are selected to form a candidate frequent (k+1)-itemset; (3) Repeat step (2) until no new frequent itemsets are generated; (4) Generate association rules: For each frequent itemset, generate all its non-empty subsets as the antecedents of the rule; For each rule's antecedent, calculate its confidence score (the rule's support divided by the antecedent's support). Based on the set minimum confidence threshold, association rules with confidence levels higher than the threshold are selected. Text datasets with confidence levels higher than the threshold are then synthesized into data packets and stored to obtain association datasets of data from different departments within the marketing data.

[0017] S3 model training The associated datasets of marketing data from different departments are input into the gated recurrent-attention mechanism model for training. The specific process is as follows: S3.1 converts the associated datasets of different departments in the marketing data into a two-dimensional matrix form that can be used as input data for the gated recurrent attention mechanism model by converting the text data into a two-dimensional matrix using Python; S3.2 Using the GRU-Attention model to find text keywords. Since the input word vectors are the result of text segmentation preprocessing, there will be some relationships between the word vectors. The deep learning model based on the attention mechanism can extract important features in the text by training the word vectors to identify the importance of words in the text. In the GRU-Attention model, through the role of the attention mechanism, keywords in the text can be found. These words are often important information for discovering implicit relationships between texts. S3.2.1 Initialize the weights and biases in the GRU-Attention model using a uniformly distributed random initialization method; S3.2.2 Forward Propagation and Loss Calculation For each sample in the training set, the output and hidden state at each time step are calculated through forward propagation. The output of the model is compared with the target value of the training set, and the loss function is calculated. S3.2.3 Backpropagation and Gradient Calculation The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm; the gradient represents the direction and magnitude of the change of the loss function relative to the model parameters. S3.2.4: Parameter Update Stochastic gradient descent is used to update the model's parameters based on the gradient. By adjusting the parameters, the loss function is gradually reduced, and the model gradually fits the training data. S3.2.5 Repeat steps 3.2.2-3.2.4 to train the model until the number of iterations or the loss function converges; S4 data conversion and storage The two-dimensional matrix output by the GRU-Attention model is converted back into text data packets using Python. The text data packets obtained after training with the GRU-Attention model have more accurate correlations. Then, the data packets are written into a dictionary using Python and stored. This dictionary contains a correlation dictionary for data such as [Customer Service -> Business Expansion Application].

[0018] The system employing the above-mentioned intelligent association recommendation method based on electricity marketing data includes: The data acquisition module collects marketing data from the power system. The data association module performs similarity association calculations on the collected data to obtain the associated datasets of data from different departments in the marketing data. The associated data collection module collects associated datasets and converts the text of the associated datasets into two-dimensional matrices for use as data input to the gated recurrent attention mechanism model. The associated data text keyword generation module is trained using a gated recurrent attention mechanism model to find text keywords in the associated dataset, and writes the found text keywords into a dictionary and stores them. The data query module allows users to input the data they want to query based on business needs. It then automatically recommends matching dictionaries based on SQL statements.

[0019] A readable computer storage medium storing a computer program that, when executed by a processor, implements the aforementioned intelligent association recommendation method based on electricity marketing data.

[0020] The above are merely specific embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent association and recommendation based on electricity marketing data, characterized in that, include: Collect marketing data from the power system; Similarity correlation calculations are performed on the collected data to obtain the associated datasets of data from different departments within the marketing data. Collect associated datasets and convert the text in the associated datasets into two-dimensional matrices for use as data input to the gated recurrent attention mechanism model; The two-dimensional matrix data is input into the gated recurrent attention mechanism model for training. The output two-dimensional matrix is ​​converted back into text format, the text keywords of the associated dataset are found, and the found text keywords are written into a dictionary and stored. Input the data you want to query based on your business needs, and the system will automatically recommend matching dictionaries through SQL statement matching.

2. The intelligent association recommendation method based on electricity marketing data according to claim 1, characterized in that, The power system marketing data includes structured data, semi-structured data, and unstructured data.

3. The intelligent association recommendation method based on electricity marketing data according to claim 1, characterized in that, The similarity association calculation of the collected data includes: The cosine similarity algorithm is used to calculate the similarity within a single business category of the collected data, and to obtain text with high similarity for a single business category. Using the Apriori algorithm, association analysis is performed on highly similar texts of single business data in different business categories to obtain an association dataset of data from different departments in the marketing data.

4. The intelligent association recommendation method based on electricity marketing data according to claim 3, characterized in that, The cosine similarity algorithm is used to calculate the similarity within a single business category of the collected data, including: Redundant information in electricity marketing data was removed. Based on the term frequency or term frequency-inverse document frequency algorithm, single types of business data are unified into a vector space of the same length; By using the features of individual business data and vector weighting, a weighted calculation is performed to obtain text with high similarity to individual business data.

5. The intelligent association recommendation method based on electricity marketing data according to claim 3, characterized in that, The Apriori algorithm includes: Frequent itemset generation based on a single character; Frequent itemset generation based on candidate itemsets; Repeat the generation of frequent itemsets based on candidate itemsets until no new frequent itemsets are generated; Generate association rules, synthesize text datasets with values ​​higher than the minimum confidence threshold into data packets and store them to obtain association datasets of data from different departments in the marketing data.

6. A system employing the intelligent association recommendation method based on electricity marketing data as described in claim 1, characterized in that, include: The data acquisition module collects marketing data from the power system. The data association module performs similarity association calculations on the collected data to obtain the associated datasets of data from different departments in the marketing data. The associated data collection module collects associated datasets and converts the text of the associated datasets into two-dimensional matrices for use as data input to the gated recurrent attention mechanism model. The associated data text keyword generation module is trained using a gated recurrent attention mechanism model to find text keywords in the associated dataset, and writes the found text keywords into a dictionary and stores them. The data query module allows users to input the data they want to query based on business needs. It then automatically recommends matching dictionaries based on SQL statements.

7. A readable computer storage medium storing a computer program, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-5.

Images