CN118152428A - A method and device for predicting and enhancing query instructions of power customer service system - Google Patents

Abstract

A prediction and enhancement method and a device for query instructions of an electric power customer service system relate to the field of data query of the electric power customer service system. The method embeds the trained data of the power industry into a large language model for loading, and converts the input data text into a semantic vector; matching the query instruction of the user with the structured data, and returning the structured data matched with the query instruction; if the query instructions are not matched, adopting an error correction algorithm to correct the error query instructions; carrying out knowledge enhancement and text classification on the error correction string by using the structured data; acquiring a query instruction input by a user, performing vector similarity search on the query instruction and the structured data, obtaining a text queue with the most relevant semantics, and outputting and displaying; constructing a retrieval enhancement generation model, and calling a large language model to generate a response result matched with a query instruction of a user. The method is mainly used for predicting and enhancing the intention of the query instruction of the electric power customer service system.

Description

A prediction and enhancement method and device for query instructions of power customer service system

Technical Field

The present invention relates to the field of data query of an electric power customer service system, and in particular to a method and a device for predicting and enhancing a query instruction of an electric power customer service system.

Background technique

At present, the answering technology for user questions in the power customer service system is mainly based on the FAQ (Frequently Asked Questions) knowledge base, and the business personnel mainly set the relevant rules. When the question hits the rule, the relevant answer is given. This knowledge search mode has many problems:

(1) Limited understanding of questions: FAQ knowledge bases can only match existing questions and answer bases, and cannot provide a deeper understanding and analysis of language, making it difficult to make intelligent predictions and recommendations based on user needs.

(2) Unable to handle complex questions: FAQ knowledge bases can only handle simple questions, but cannot handle complex, frequent questions and multi-round conversations. They lack the ability to deeply understand and predict user needs and cannot accurately grasp user needs, resulting in poor user experience. Many question-and-answer rules require manual labeling, which increases the workload of customer service staff.

(3) High quality requirements for answers: The answers returned by the FAQ knowledge base must be accurate, otherwise it will affect user experience and trust.

In the power customer service system, accurate understanding of user queries (i.e. query instructions) is the key to providing customers with quality services. Due to its own professionalism and particularity, the power industry also has high requirements for the search of power knowledge. Generally speaking, there are multiple different business systems within the customer service of power companies, such as professional customer service assistants for customers, business systems for employees, and procurement systems for suppliers. The data and documents generated by the diverse businesses are massive, and the search terms of the power customer system are generally professional terms, short paragraphs or long documents. When customers ask questions, they often express unclearly, use voice input conversion inaccurately, omit expressions, etc., resulting in unsatisfactory search results and ineffective customer service.

Although the existing power customer service system has also developed some simple search models, the response to user queries and intent recognition mainly rely on keyword matching or are mainly based on natural language processing technology, but this may lead to inaccurate intent recognition in some cases. For example, users may use different words or phrases to express the same intent, or keyword matching fails due to errors in voice input.

Therefore, there is a need for a method and device for predicting and enhancing query instructions of an electric power customer service system that can accurately and deeply understand the user's intentions, predict the user's query instructions, and enhance the intentions of the user.

Summary of the invention

In order to solve the defects of the existing electric power customer service system inaccurately recognizing the intention of user query instructions and failing to deeply understand the user's intention, the present invention provides a method and device for predicting and enhancing the query instructions of the electric power customer service system, which can accurately and deeply understand the user's intention, predict the user's query instructions and enhance the intention.

The method for predicting and enhancing a query instruction of a power customer service system according to the present invention comprises the following steps:

S1: Collect and process data from the power industry and train it;

S2: embed the trained data into the large language model for loading, and use the large language model to convert the input data text into a semantic vector;

S3: Match the user's query instruction with the structured data. If the match is successful, directly return the structured data matching the query instruction; if the match is not successful, execute S4 for the query instruction with error;

S4: using an error correction algorithm to correct the erroneous query instruction, thereby obtaining an error correction string with the highest probability of being correct;

S5: performing knowledge enhancement and text classification on the error correction string using the structured data;

S6: Obtain the query instruction input by the user, perform vector similarity search on the query instruction and the structured data, obtain the text queue with the most semantic relevance, and output it for display;

S7: Construct a retrieval enhancement generation model, and call the large language model to generate a response result that matches the user's query instruction.

Furthermore: in S2, it also includes: constructing a vector database and storing the semantic vector in the vector database.

Further: in S7, it also includes: constructing a prompt library for guiding the large language model to answer questions according to the context;

Further: in S1, the collected data includes user query instructions and power field knowledge; the processing includes preprocessing and cleaning for removing irrelevant information and noise data; the training is first using a document loader to load different types of documents into plain text, and then using a text segmenter to segment the plain text.

Further: In S4, the error correction algorithm is modeled and calculated by the Bayesian formula, that is:

argmax p(r|q) = argmax p(q|r)p(r)/p(q) = argmax p(q|r)p(r);

Among them, argmax means finding the maximum value of the probability of all error-correcting strings occurring, q represents the original string, r represents the error-correcting string, p(r|q) represents the probability of the error-correcting string occurring when the original string q occurs, p(q|r) represents the transition probability from the original string to the error-correcting string, p(r) represents the probability of the error-correcting string being a normal query instruction, and p(q) is a constant that is independent of r.

Further: In S5, the structured data is stored in a knowledge graph or a knowledge base, and the knowledge enhancement refers to entity linking and semantic reasoning between the user's query instructions and the structured data, and training the large language model to understand and represent the structured data in the knowledge graph or knowledge base, thereby improving the large language model's ability to recognize the intent of the query instructions.

Further: In S5, the text classification is implemented by using a convolutional neural network, and the specific process is as follows:

S51: Construct a convolutional neural network model; the convolutional neural network model includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer; the input layer is used to receive text to be classified, the convolutional layer is used to extract local features in the text, the pooling layer is used to reduce the dimension of the data, the fully connected layer is used to map the extracted features to the output space, and the output layer is used to classify the text;

S52: Using the labeled structured data as a training set, training the convolutional neural network model so that it learns and recognizes different text categories from the structured data; the training of the convolutional neural network model is implemented using a cross entropy loss function and a gradient descent optimizer;

S53: Use the test set to test and evaluate the trained convolutional neural network model, and calculate the classification accuracy and/or recall rate indicators to evaluate the performance of the convolutional neural network model.

Further: in S7, the response result is displayed in a manner selected according to the needs of the user, and the display manner includes text form, data chart form, audio form and/or video form.

The device for predicting and enhancing query instructions of a power customer service system described in the present invention comprises a large language model, a data training module, a text conversion module, an intention recognition enhancement module, an error correction module and a response result generation and display module;

The data training module is used to train data of the power industry;

The text conversion module is used to call the large language model to convert the trained data text into a semantic vector;

The error correction module is used to correct the error query instruction, so as to obtain the error correction string with the highest correct probability;

The intention recognition enhancement module is used to perform knowledge enhancement and text classification on the error correction string;

The response result generation and display module is used to call the large language model to generate a response result that matches the user's query instruction.

Further: also including a vector database and a prompt library;

The vector database is used to store the semantic vector;

The prompt library is used to guide the large language model to generate a response result according to the context.

The beneficial effects of the present invention are:

The present invention is a method and device for accurately predicting and enhancing data query intentions in an electric power customer service system based on large model generative technology. It can improve the accuracy of user intention recognition in the electric power customer service system, improve the efficiency of unified search queries by electric power users, and improve service efficiency. It can understand and predict user intentions more conveniently, uniformly, quickly, and accurately, and improve customer service efficiency and customer service satisfaction.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a method for predicting and enhancing a query instruction;

FIG2 is a schematic diagram of the structure of knowledge enhancement.

Detailed ways

The following are only preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a technician familiar with the technical field within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. The embodiments described below are only used to explain the present invention and cannot be interpreted as limitations on the present invention. The protection scope of the present invention should be based on the protection scope of the claims. The embodiments of the present invention are described in detail below. In order to facilitate the description of the present invention and simplify the description, the technical terms used in the specification of the present invention should be interpreted in a broad sense, including but not limited to conventional replacement schemes not mentioned in this application, and also including direct implementation methods and indirect implementation methods.

Example 1

The present embodiment is described in conjunction with FIG. 1 and FIG. 2 . The present embodiment discloses a method for predicting and enhancing a query instruction of a power customer service system, which specifically includes the following steps:

S1: Collect and process data from the power industry and train it;

First, the data of the power industry is collected and processed to enhance the query data of the query instructions in the power customer service field. In order for the large language model to learn and understand the knowledge in the power customer service field, it is necessary to collect relevant question and answer data in the power industry field from systems such as the power customer service knowledge base, including user queries (query instructions) and power field knowledge. The collected data needs to be preprocessed and cleaned to remove irrelevant information and noise data. The collected data includes user query instructions and power field knowledge; the processing includes preprocessing and cleaning to remove irrelevant information and noise data; the training is to first use a document loader to load different types of documents into plain text, and then use a text segmenter to segment the plain text.

Large Language Models (LLMs) are an application of large models in the field of natural language processing. Large model generative technology refers to the technology and methods of using large models for generative tasks.

Document splitting can be performed for long document information. The context window of a large language model is limited. The context window refers to the range or length of text that a large language model can consider and remember when processing text input. The size of this window is critical to the performance of the model because it determines the amount of text that the model can refer to and understand when generating responses or performing tasks. A document is often large, so the file needs to be split and relevant content fragments need to be found based on user semantics. Use the document loader of langchain (LLM programming framework) to load different types of documents into plain text, and then split the document into segments.

Recursive Character Text Splitter is a text splitter that splits text into smaller chunks using a set of characters as separators, such as "\n\n", "\n", "space", and "-". The granularity and continuity of the segmentation can be adjusted by controlling the parameters of the splitter, such as the maximum size of the text chunks and the amount of overlap.

S2: Embed the trained power field data into the large language model for loading, so as to prepare for the conversion of text into semantic vectors. Use the large language model to convert the input data text into semantic vectors; for subsequent semantic search using vectors.

Construct a vector database. The big language model will convert the input text and other power customer information into semantic vectors and store them in the vector database. By using the big model generative technology, the input text and other power customer information are converted into semantic vectors. These semantic vectors can capture the deep semantic information in the text, so that we can more accurately understand the real needs and accurate intentions of the query instructions entered by the power customers. And storing the generated semantic vectors in the vector database can facilitate subsequent queries and use. When necessary, relevant vectors can be retrieved from the vector database to perform similarity calculations, classification and other tasks, thereby realizing fast and accurate processing of power customer query instructions.

As shown in Figure 1, Prompt-RAG is loaded in advance to guide the model to answer questions based on the context. Prompt-RAG is a method similar to RAG (Retrieval-Augmented Generation) but without vector database and embedding. It is used to optimize large language models (LLM) for specific domains. The large language model can be ERNIE Bot, which has the technical advantages of knowledge enhancement, retrieval enhancement and dialogue enhancement. The number of model parameters N (excluding embedding), the size of the data set D, the amount of computation used for training C, and the training process includes pre-training and fine-tuning.

S3: Match the user's query instruction with the structured data. If the match is successful, directly return the structured data matching the query instruction; if not, execute S4 for the erroneous query instruction; when the user enters an electricity-related query instruction query, since the query instruction query entered by the user may be special due to dialects or professional names, the query instruction query entered by the user will generally be judged. If it can directly hit the corresponding structured data in the vector database, directly return the structured data corresponding to the query instruction query; if it does not hit, the query will be corrected.

S4: Use an error correction algorithm to rewrite and correct the erroneous query instruction query, so as to obtain the error correction string with the highest probability of being correct; query error correction is the process of detecting and correcting errors in the query instruction query input by the power enterprise user, generally including word correction. Due to the particularity of the power industry, when users are consulting questions and inputting searches, they may not understand professional terms, make mistakes when inputting, make errors in voice input recognition, and input in dialects, etc., which will lead to certain errors in the search queries input by power users. Therefore, it is necessary to correct these erroneous queries so as not to affect the accuracy of the large language model matching, and ultimately affect the user's search experience.

There are generally two types of errors generated by queries: Non-word spelling errors and Real-word spelling errors. Non-word spelling errors indicate that the word itself does not exist in the dictionary. Real-word spelling errors refer to errors in the word itself, but do not conform to the context. They often involve errors at the grammatical and semantic levels, such as writing "I want to check the electricity bill" as "I want to check the electricity share". This type of error is more difficult to correct than a Non-word spelling error.

The following is an example of error correction for power customer service related queries:

Non-word spelling errors:

Pinyin error: chadianfen → check electricity bill;

Pinyin abbreviation: Cdf → check electricity bill;

Mixed pinyin and numbers: woyao 2 check electricity bill → I want to check the electricity bill;

Real-word spelling errors:

Missing words: I want to check the bill → I want to check the electricity bill;

Multi-word: I want to check the electricity bill → I want to check the electricity bill;

Reversal: I check the electricity bill → I want to check the electricity bill;

Homophone: I want to check the electricity waste → I want to check the electricity bill;

Confusing sound: rub electricity bill → check electricity bill;

Similar characters: 查由费→查电费;

Wrong phrase collocation: Business expansion registration → Business expansion explosion.

The error correction of user query commands includes error detection and error correction. Error detection is used to identify the location of incorrect words. It can be simply done by segmenting the input query and checking whether each word is in the maintained custom word list or the accumulated common error correction pairs. If not, the candidate is replaced based on the shape, pronunciation or similar characters of the input code and combined with the probability of the N-Gram language model to determine whether it is wrong. N-Gram is a language model commonly used in large vocabulary continuous speech recognition. For Chinese, it is called the Chinese Language Model (CLM). The Chinese language model uses the collocation information between adjacent words in the context to realize the automatic conversion of Chinese characters. This method does not fully consider the context information and can only be applied to the detection of common Chinese phrase collocation, English word errors, etc. A further approach is to identify the start and end positions of the error by training a sequence annotation model.

The error correction of user queries is mainly based on modeling the relationship between the original query and the corrected query. This process can be decomposed into two main probability calculations: p(q|r) and p(r). p(q|r) represents the transition probability from the original string to the corrected string, which can be calculated by calculating the edit distance between the original string and the corrected string, the co-click probability, and extracting corresponding features.

Commonly used edit distance algorithms include Levenshtein distance (edit distance that allows insertion, deletion, and replacement of a character) and Damerau-Levenshtein distance (edit distance that allows the positions of two adjacent characters to be swapped and allows insertion, deletion, and replacement of a character). It is an edit distance algorithm used to measure the degree of difference between two strings. It is an extension of the Levenshtein distance algorithm and can measure the similarity between two strings. The co-click probability refers to the probability that two strings are clicked at the same time in the same context, which can be calculated by analyzing user behavior data; the feature extraction method is to extract useful information from the query, such as keywords, grammatical structures, etc., to facilitate subsequent model prediction.

As shown in Figure 2, p(r) represents the probability that the error correction string is a normal query, which can be measured by language models, high-frequency queries, entity knowledge bases, etc. Language model is a statistical language understanding technology, and its quality can be evaluated by calculating the language probability of the error correction string. High-frequency queries refer to queries that are frequently used by a large number of users. These user queries tend to have high quality and credibility. Entity knowledge base is a database containing entity information, such as Entitiy1. The relevance and accuracy of the error correction string can be evaluated by querying the entity knowledge base.

The goal of the error correction algorithm is to find an error correction string r that maximizes the probability of transition from the original string q to the error correction string r, and at the same time, the probability of the error correction string r being a normal query is also high. This goal can be modeled and calculated using the Bayesian formula, i.e., argmaxp(r|q)=argmaxp(q|r)p(r)/p(q)=argmaxp(q|r)p(r). During the calculation process, various optimization algorithms and machine learning methods can be used to improve the accuracy and efficiency of the error correction algorithm. Among them, q represents the original string, r represents the error correction string, p(r|q) represents the probability of the error correction string occurring when the original string q occurs, p(q|r) represents the probability of transition from the original string to the error correction string, p(r) represents the probability of the error correction string being a normal query, and p(q) is a constant that is independent of r.

The first equation argmaxp(r|q)=argmaxp(q|r)p(r)/p(q) is based on Bayes' theorem. Since the argmax operation is looking for the value of r that maximizes the probability, and p(q) is a constant that is independent of r (for a given q), it will not affect the result of argmax. Therefore, p(q) in the denominator can be ignored, resulting in the second equation. The second equation argmaxp(q|r)p(r)/p(q)=argmaxp(q|r)p(r) is based on the above reason, that is, p(q) is a constant and will not affect the result of argmax. The above formula shows that when looking for the value of r that maximizes the conditional probability p(r|q), p(q) in the denominator can be ignored, and only the numerator p(q|r)p(r) needs to be considered. This simplifies the calculation process and quickly obtains the correct error correction string r with the highest probability.

S5: performing knowledge enhancement and text classification on the error correction string using the structured data;

The structured data is stored in a knowledge graph or knowledge base, and the knowledge enhancement refers to entity linking and semantic reasoning between the user's query instructions and the structured data, and training the model to understand and represent the structured data in the knowledge graph or knowledge base, thereby improving the model's ability to recognize the intent of the query instructions. The intent recognition module is usually a classification task, and its purpose is to identify the category that the user wants to query, and then output it to the recall and sorting module to ensure the category relevance of the final result. Intent recognition enhancement is mainly divided into two steps, mainly including customer service knowledge enhancement and text classification. Adding text classification can make user query searches more accurate.

Knowledge Enhancement:

Knowledge enhancement is the use of structured data in knowledge graphs or knowledge bases to perform entity linking and semantic reasoning on user queries, and to improve the model's ability to recognize complex query intentions by training models to understand and represent the knowledge in the knowledge graph or knowledge base; matching and associating user queries with the knowledge in the knowledge graph or knowledge base to provide users with more accurate and relevant results.

Text Categorization:

When doing text classification, you can use a neural network model, such as a convolutional neural network (CNN). CNN is widely used in the field of computer vision, but it can also be used for text classification. CNN captures local features in the text through convolution operations and reduces the dimensionality of the features through pooling operations. Finally, a fully connected layer maps the features to the output space.

S51: Build a convolutional neural network model, including input layer, convolution layer, pooling layer, fully connected layer and output layer. The convolution layer is used to extract local features in the text, the pooling layer is used to reduce the dimension of the data, the fully connected layer is used to map the extracted features to the output space, and the output layer is used to classify the text.

S52: Training model: Use a large amount of labeled text data as a training set to train the convolutional neural network model so that it can learn from the data and recognize different text categories.

S53: Testing and evaluation: Use the test set to test and evaluate the trained model, and calculate indicators such as classification accuracy and recall rate to evaluate the performance of the model.

The specific process is as follows:

① Input layer:

(X = [x_1; x_2; ...; x_n]);

Where (x_i) represents the embedding vector of the (i)th word and X represents the input matrix.

②Convolutional layer:

(H = [h_1; h_2; ...; h_m]);

Where H represents the height of the convolutional layer output matrix, (h_j) represents the output of the (j)th feature map, and the calculation formula is as follows:

(h_j = f(W_j * X + b_j));

Among them, (h_j) is the height of the (j)th convolution kernel, (W_j) is the weight matrix of the (j)th convolution kernel, (b_j) is the bias term, and (f) is the activation function (such as ReLU).

③Pooling layer:

Assuming the maximum pooling operation is used, the calculation formula is as follows:

((pooling_output) = max(h_1, h_2, ..., h_m));

Among them, (pooling_output) represents the output vector of the pooling layer.

④Fully connected layer:

Assuming the ReLU activation function is used, the calculation formula is as follows:

(FC = [relu(w_1h + b_1); relu(w_2h + b_2); ...; relu(w_k*h + b_k)]);

Among them, (wi) and (bi) are the weight and bias of the i-th neuron respectively, relu represents the activation function of each neuron, and the mathematical expression of the relu function is f(x)=max(0,x); FC represents the output vector of the fully connected layer.

⑤Output layer:

Use the softmax function to convert the output of the fully connected layer into a probability distribution. The calculation formula is as follows:

(P = [softmax(w'_1FC + b'_1); softmax(w'_2FC + b'_2); ...; softmax(w'_n*FC + b'_n)]),

Where (w'_i) is the weight of the (i)th neuron, (b'_i) is the bias term of the (i)th neuron, and P represents the output probability of the fully connected layer;

Among them, the mathematical expression of the softmax function is: ;

Loss function and optimizer:

The cross entropy loss function is used for model training. The loss function is calculated as follows:

;

Among them, Loss is the loss function, N is the number of samples, is the true label (i.e. 0 or 1), /> is the predicted probability, where the log function does not limit the base and calculates all possible values in a traversal form.

S6: Get the query command query input by the user, perform a vector similarity search on it and the structured data, as shown in Figure 2, that is, the entity queue and its introduction, obtain the most semantically relevant text queue, that is, obtain the top k most semantically relevant texts, that is, the text queue formed by similar questions, and output it for display.

S7: Construct a retrieval-augmented generation model, and call the large language model to generate a response result that matches the user's query instruction. Constructing rag_chain, that is, constructing rag_chain (Retriever-Augmented Generation, retrieval-augmented generation model) is a method that combines information retrieval and generative models, aiming to improve the accuracy and relevance of generative models when answering user queries. Calling a large language model to answer user questions, the output results can output basic text results according to different user needs, or generate data charts for related questions to achieve multimodal responses to user questions.

Example 2

This embodiment is described in combination with Example 1. This embodiment discloses a prediction and enhancement device for query instructions of an electric customer service system, including a large language model, a data training module, a text conversion module, an intention recognition enhancement module, an error correction module, and a response result generation and display module;

The data training module is used to train data of the power industry;

The text conversion module is used to call the large language model to convert the trained data text into a semantic vector;

The error correction module is used to correct the error query instruction, so as to obtain the error correction string with the highest correct probability;

The intention recognition enhancement module is used to perform knowledge enhancement and text classification on the error correction string;

The response result generation and display module is used to call the large language model to generate a response result that matches the user's query instruction.

Also includes vector database and hint library;

The vector database is used to store the semantic vector;

The prompt library is used to guide the large language model to generate a response result according to the context.

Claims (10)

1. A method for predicting and enhancing query instructions of an electric power customer service system, characterized in that it comprises the following steps: S1: Collect and process data from the power industry and train it; S2: embed the trained data into the large language model for loading, and use the large language model to convert the input data text into a semantic vector; S3: Match the user's query instruction with the structured data. If the match is successful, directly return the structured data matching the query instruction; if the match is not successful, execute S4 for the query instruction with error; S4: using an error correction algorithm to correct the erroneous query instruction, thereby obtaining an error correction string with the highest probability of being correct; S5: performing knowledge enhancement and text classification on the error correction string using the structured data; S6: Obtain the query instruction input by the user, perform vector similarity search on the query instruction and the structured data, obtain the text queue with the most semantic relevance, and output it for display; S7: Construct a retrieval enhancement generation model, and call the large language model to generate a response result that matches the user's query instruction. 2. A method for predicting and enhancing query instructions of an electric power customer service system according to claim 1, characterized in that, in S2, it also includes: constructing a vector database and storing the semantic vector in the vector database. 3. The method for predicting and enhancing query instructions of an electric power customer service system according to claim 1, characterized in that in S7, it also includes: constructing a prompt library to guide the large language model to answer questions according to the context. 4. According to claim 1, a method for predicting and enhancing query instructions of an electric power customer service system is characterized in that, in S1, the collected data includes user query instructions and electric power field knowledge; the processing includes preprocessing and cleaning to remove irrelevant information and noise data; the training is first using a document loader to load different types of documents into plain text, and then using a text segmenter to segment the plain text. 5. The method for predicting and enhancing query instructions of a power customer service system according to claim 1, characterized in that, in S4, the error correction algorithm is modeled and calculated by the Bayesian formula, that is: argmax p(r|q) = argmax p(q|r)p(r)/p(q) = argmax p(q|r)p(r); Among them, argmax means finding the maximum value of the probability of all error-correcting strings occurring, q represents the original string, r represents the error-correcting string, p(r|q) represents the probability of the error-correcting string occurring when the original string q occurs, p(q|r) represents the transition probability from the original string to the error-correcting string, p(r) represents the probability of the error-correcting string being a normal query instruction, and p(q) is a constant that is independent of r. 6. According to claim 1, a method for predicting and enhancing query instructions of an electric power customer service system is characterized in that, in S5, the structured data is stored in a knowledge graph or a knowledge base, and the knowledge enhancement refers to entity linking and semantic reasoning between the user's query instructions and the structured data, and training the large language model to understand and represent the structured data in the knowledge graph or knowledge base. 7. The method for predicting and enhancing query instructions of a power customer service system according to claim 1, characterized in that, in S5, the text classification is implemented by using a convolutional neural network, and the specific process is as follows: S51: Construct a convolutional neural network model; the convolutional neural network model includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer; the input layer is used to receive text to be classified, the convolutional layer is used to extract local features in the text, the pooling layer is used to reduce the dimension of the data, the fully connected layer is used to map the extracted features to the output space, and the output layer is used to classify the text; S52: Using the labeled structured data as a training set, training the convolutional neural network model so that it learns and recognizes different text categories from the structured data; the training of the convolutional neural network model is implemented using a cross entropy loss function and a gradient descent optimizer; S53: Use the test set to test and evaluate the trained convolutional neural network model, and calculate the classification accuracy and/or recall rate indicators to evaluate the performance of the convolutional neural network model. 8. A method for predicting and enhancing query instructions of an electric power customer service system according to claim 1, characterized in that, in S7, the response result is displayed in a manner selected according to user needs, and the display manner includes text form, data chart form, audio form and/or video form. 9. A prediction and enhancement device for query instructions of an electric power customer service system, characterized by comprising a large language model, a data training module, a text conversion module, an intention recognition enhancement module, an error correction module and a response result generation and display module; The data training module is used to train data of the power industry; The text conversion module is used to call the large language model to convert the trained data text into a semantic vector; The error correction module is used to correct the error query instruction, so as to obtain the error correction string with the highest correct probability; The intention recognition enhancement module is used to perform knowledge enhancement and text classification on the error correction string; The response result generation and display module is used to call the large language model to generate a response result that matches the user's query instruction. 10. A prediction and enhancement device for query instructions of a power customer service system according to claim 9, characterized in that it also includes a vector database and a prompt library; The vector database is used to store the semantic vector; The prompt library is used to guide the large language model to generate a response result according to the context.