Intelligent customer service interaction content recommendation method and system based on artificial intelligence

Abstract

The invention provides an intelligent customer service interaction content recommendation method and system based on artificial intelligence, and the method comprises the steps: obtaining power grid equipment state data and user behavior data, and carrying out the weighted fusion through an attention mechanism, and outputting a fusion feature vector; predicting a prediction vector set of a future time period based on the fused feature vector; screening a demand event set higher than a threshold in the prediction vector set, and generating a recommended content set; performing emotion recognition according to a current input text of the user, and calculating a recommended content score and a pushing priority score in combination with the recommended content set; a cross-modal consistency regular term is introduced to correct the push strategy vector, and an adjusted recommendation push score and a corresponding recommendation content set are obtained; and converting the adjusted recommendation push score and the corresponding recommendation content set into a scheduling task, and performing linkage execution with a scheduling system. According to the invention, efficient, active and intelligent upgrading of the intelligent customer service system in a power grid scene is realized.

Description

Technical Field

[0001] This invention belongs to the field of intelligent customer service, and in particular relates to a method and system for recommending intelligent customer service interactive content based on artificial intelligence. Background Technology

[0002] With the rapid development of artificial intelligence and data analytics, intelligent customer service systems have become an important means of improving service quality and customer satisfaction in the power industry. Power grid companies handle a large volume of customer inquiries and repair requests through customer service platforms. Traditional human customer service relies on fixed scripts and experience-based judgment, which cannot provide accurate and efficient responses in the complex and dynamic power grid environment. Existing intelligent customer service systems are mostly based on keyword matching or preset templates for question and answer, which, while improving routine business processing, still has significant limitations.

[0003] First, power grid operation has strong time-series characteristics and regional differences. User problems are often closely related to factors such as the status of power grid equipment, load fluctuations, and weather conditions. Traditional systems cannot integrate these multi-source data in real time for dynamic response.

[0004] Secondly, existing customer service systems lack long-term modeling and personalized analysis of user behavior, making it impossible to identify potential user needs or abnormal electricity consumption. For example, in the event of a partial power outage or equipment overload, the system can only passively receive user reports of faults, unable to predict user inquiry hotspots or provide preventative advice. Furthermore, customer service recommendations remain at the template-based push stage, lacking a personalized recommendation mechanism tailored to user scenarios, emotional states, and the actual operating status of the power grid. Especially in emergency scenarios such as peak load periods, extreme weather, or regional faults, the system fails to establish effective coordination with power grid operations and maintenance, resulting in delayed user response, degraded service experience, and limited operational efficiency.

[0005] Current intelligent customer service models are insufficient to meet the multi-level, real-time, and interpretability requirements of the power grid industry. There is an urgent need for an intelligent recommendation mechanism that can combine power grid operation status, user behavior patterns, and emotion recognition to achieve closed-loop optimization from data perception, demand prediction, content generation to service execution. Summary of the Invention

[0006] The purpose of this invention is to propose an intelligent customer service interactive content recommendation method and system based on artificial intelligence to address the above-mentioned shortcomings.

[0007] To achieve the above objectives, a first aspect of the present invention provides an intelligent customer service interaction content recommendation method based on artificial intelligence, the method comprising:

[0008] S1. Acquire power grid equipment status data and user behavior data, and output a fused feature vector by weighted fusion through an attention mechanism;

[0009] S2. Based on the fused feature vector, a set of prediction vectors for future time periods is predicted using an improved gated cyclic unit network; wherein each vector in the set of prediction vectors represents the prediction result of user demand or power grid status for the corresponding time slice;

[0010] S3. Select the set of demand events that are higher than the threshold from the set of predicted vectors, and generate a set of recommended content based on the set of demand events and the fused feature vector; perform emotion recognition based on the user's current input text to obtain a joint emotion vector, and calculate the recommended content score and push priority score in combination with the set of recommended content.

[0011] S4. Introduce a cross-modal consistency regularization term based on joint sentiment vector to modify the push strategy vector, and obtain the adjusted recommendation push score and the corresponding set of recommended content.

[0012] S5. The adjusted recommendation push score and the corresponding set of recommended content are converted into a scheduling task and executed in conjunction with the scheduling system.

[0013] Furthermore, the power grid equipment status data includes the operating parameters of equipment in the transmission, distribution, and terminal metering stages; the user behavior data includes users' electricity consumption curves, inquiry content, and historical electricity consumption events.

[0014] Furthermore, the weighted fusion output of the fused feature vector through the attention mechanism specifically includes:

[0015] The power grid equipment status data is input into the power grid equipment channel to extract features from the time series load curve, thereby obtaining the power grid equipment status features. The lengths of each convolutional kernel are 3, 5, and 7, respectively, to capture short-term fluctuations and medium-term trends. The power grid equipment channel is a three-layer convolutional network.

[0016] The user behavior data is input into the user behavior channel to obtain the state vector of user behavior in each time segment;

[0017] The outputs of the power grid equipment channel and the user behavior channel are weighted and fused using an over-attention mechanism to output a fused feature vector.

[0018] Furthermore, the improved gated cyclic unit network introduces a grid stability constraint term in the gating mechanism to express the system's inertia at the physical level and generate a grid state representation; wherein, the grid stability constraint term represents a system stability index calculated based on actual grid data, obtained through historical load standard deviation or equipment health function.

[0019] Furthermore, the improved gated recurrent unit network also adopts a structure combining two stacked temporal convolutional networks with an attention layer to generate user behavior representations, which are used to identify user behavior clustering phenomena within a certain time window.

[0020] The set of prediction vectors for predicting future time periods is specifically as follows:

[0021] After concatenating the power grid state representation and the user behavior representation, a set of prediction vectors is output through a linear mapping layer. Each prediction vector in the set of prediction vectors represents the user demand or power grid state prediction result for the corresponding time slice.

[0022] The linear loss function includes a power grid security constraint and a behavior prediction objective; the power grid security constraint is used to prevent frequent fluctuations or oversensitivity in the model output; the behavior prediction objective is cross-entropy loss.

[0023] Furthermore, the set of recommended content includes recommended text, explanatory statements, and operational suggestions;

[0024] Specifically, the step of performing emotion recognition based on the user's current input text to obtain a joint emotion vector is as follows:

[0025] The user's current input text is input into a dual-channel fusion model, which includes a text sentiment channel and a behavioral sentiment channel. The text sentiment channel uses a pre-trained language model BERT to output a sentiment vector, while the behavioral sentiment channel inputs a user behavior representation and outputs a sentiment vector through two layers of gated recurrent units. The sentiment vectors are then concatenated to obtain a joint sentiment vector.

[0026] Furthermore, the recommended content score is generated by calculating the predicted vector set and the event topic vector, and is used to evaluate the degree of fit between the recommended content and the current scenario;

[0027] The push priority score is calculated and generated based on the recommended content score, joint sentiment vector, user's historical acceptance rate on this type of recommendation, and the current load level of the customer service system.

[0028] Then, based on the push priority score, several recommended content items are selected, sorted by priority, and put into the push queue; if the push priority score exceeds a certain threshold and the joint sentiment vector is below the critical value, the recommended content will be marked as "push immediately", otherwise it will be put into the delayed push list or handed over to customer service for review.

[0029] Furthermore, the cross-modal consistency regularization term is calculated and generated based on the emotion vector and the emotion vector.

[0030] Furthermore, S5 specifically includes:

[0031] Based on the geographic region code field carried in the recommended content set, it is aligned with the operating status data in the real-time power grid equipment status data. If the alignment matches, the task matching process begins.

[0032] The task matching includes:

[0033] Based on the adjusted recommendation push score, the system calculates the task activation score by combining the power grid operation indicators of the corresponding region and the current load status of the network-wide dispatch resources. When the task activation score exceeds a set threshold, the system will construct a dispatch task request, which includes the task type, associated device ID, recommendation source, expected response window, and suggested execution method.

[0034] The scheduling task request is pushed to the scheduling management system or operation and maintenance system via API interface.

[0035] In another aspect of the invention, an intelligent customer service interaction content recommendation system based on artificial intelligence is provided, the system comprising:

[0036] The data fusion module is used to acquire power grid equipment status data and user behavior data, and outputs a fused feature vector through weighted fusion using an attention mechanism;

[0037] The predictive analysis module is used to predict a set of prediction vectors for future time periods based on the fused feature vectors through an improved gated cyclic unit network; wherein each vector in the set of prediction vectors represents the prediction result of user demand or power grid status for the corresponding time slice.

[0038] The recommendation engine module is used to filter out the set of demand events that are higher than a threshold from the set of predicted vectors, generate a set of recommended content based on the set of demand events and the fused feature vector; perform emotion recognition based on the user's current input text to obtain a joint emotion vector, and calculate the recommended content score and push priority score in combination with the recommended content set;

[0039] The sentiment analysis module is used to introduce a cross-modal consistency regularization term based on the joint sentiment vector to modify the push strategy vector, so as to obtain the adjusted recommendation push score and the corresponding set of recommended content.

[0040] The operation and maintenance collaboration module is used to transform the adjusted recommendation push score and the corresponding set of recommended content into scheduling tasks, and execute them in conjunction with the scheduling system.

[0041] The beneficial technical effects of the present invention are at least as follows:

[0042] This invention proposes an AI-based intelligent customer service interactive content recommendation method and system for power grids, enabling intelligent management of the entire process from power grid operation status perception to service delivery and operation and maintenance collaboration. The system acquires power grid equipment operation data and user behavior data through a multi-source fusion model, extracts power grid time-series features and user behavior patterns using a dual-channel neural structure, and generates a unified fusion feature vector to provide basic data support for subsequent prediction and recommendation. Based on this, a hybrid time-series prediction model with physical constraints is constructed by introducing power grid stability constraints and a behavior sparsity reinforcement mechanism, enabling forward-looking prediction of user needs and power grid status. The recommendation module uses the prediction results as trigger conditions, combines contextual features to generate structured recommendation content, and quantifies the credibility and timeliness of the recommendations through a confidence score function. To address different user emotional states and power grid service pressures, the system further introduces emotion recognition and dynamic push control mechanisms. A dual-channel emotion recognition model analyzes user language and behavioral signals, and an emotion regulation function adjusts the push rhythm and tone template, making the recommended content more humanized and acceptable. Finally, based on the recommendation results and real-time power grid operation data, the system establishes a task activation and scheduling feedback mechanism to achieve closed-loop collaboration between the intelligent customer service system and power grid operation and maintenance. In scenarios such as faults and load anomalies, it can automatically trigger task issuance, personnel scheduling, and execution feedback, forming a linkage system between intelligent customer service and power grid management. This invention, through a systematic design of data fusion, predictive analysis, emotion adjustment, and operation and maintenance linkage, solves the problems of delayed response, generalized recommended content, lack of emotion adaptation, and lack of execution closed loop in existing power grid customer service systems, achieving a highly efficient, proactive, and intelligent upgrade of the intelligent customer service system in the power grid scenario. Attached Figure Description

[0043] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0044] Figure 1 This is a flowchart of an intelligent customer service interactive content recommendation method based on artificial intelligence, according to an embodiment of the present invention.

[0045] Figure 2 This is a framework diagram of an intelligent customer service interactive content recommendation system based on artificial intelligence, according to an embodiment of the present invention. Detailed Implementation

[0046] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0047] like Figure 1 As shown in the embodiment of the present invention, an intelligent customer service interaction content recommendation method based on artificial intelligence is provided. The method includes:

[0048] S1. Obtain power grid equipment status data and user behavior data, and output a fused feature vector by weighted fusion through an attention mechanism.

[0049] Specifically, the goal of this step is to construct a structured, reusable dataset by bidirectionally collecting and fusing power grid equipment operation data and user behavior data, providing a unified data input foundation for subsequent demand forecasting and recommendation generation. This data fusion is not a simple data splicing, but rather involves intelligent matching and normalization of data in both time and semantic dimensions, enabling the dynamic operating characteristics of the power grid and the behavioral characteristics of users to be understood and used by the same model, thereby forming a computable and learnable input representation.

[0050] Furthermore, the input consists of two types of data: power grid equipment status data and user behavior data. The former includes equipment operating parameters from transmission, distribution, and terminal metering stages, such as load, voltage, current, power factor, and alarm and fault information for each substation. This data is obtained through the remote acquisition interface of the SCADA system (Supervisory Control and Data Acquisition) and smart meters. The acquisition frequency is dynamically adjusted between seconds and minutes based on the importance of the equipment. The latter mainly comes from the customer service system and electricity data management platform, including users' electricity consumption curves (e.g., 24-hour periodic load changes), inquiry content (e.g., power outage inquiries, payment inquiries, fault reporting, etc.), and historical electricity events (e.g., abnormal fluctuations, power outage records due to arrears, etc.). This data is automatically extracted through the Customer Information Management System (CIS) and online customer service log system, and mapped to power grid equipment data using user IDs and power supply area IDs.

[0051] Furthermore, after data collection, the first step is to standardize and time-align the data from different sources. Power grid equipment data, due to its high sampling frequency and massive volume, requires a distributed data cleaning module to remove invalid records such as communication packet loss and outliers (e.g., current values ​​suddenly turning negative). The cleaned data is then grouped by timestamp, forming a time-series sequence with minute-level granularity. User behavior data, due to its diverse modalities (including text, time series, and discrete events), requires multi-dimensional feature extraction. User inquiry texts are processed using a word segmentation model to extract keywords, with the keyword importance calculated using the TF-IDF algorithm. Historical electricity consumption curves are processed using a sliding window to calculate daily average load and peak load times, thereby extracting periodic features. The key processing element is ensuring the synchronization of different modalities of data across time. For example, if a region experiences a sudden load surge at 2:00 PM, and multiple users submit inquiries about "voltage instability" between 2:05 PM and 2:10 PM, this data, after fusion, will form a feature sample with a clear temporal causal relationship.

[0052] The key to data fusion lies in mapping power grid equipment data and user behavior data to a unified feature space through a dual-channel embedding model. The power grid equipment channel uses a three-layer convolutional network to extract features from the time-series load curve, with kernel lengths of 3, 5, and 7 for each layer, respectively, to capture short-term fluctuations and medium-term trends. The user behavior channel uses a two-layer Long Short-Term Memory (LSTM) network to model user time-series behaviors (such as electricity consumption cycles and fault reporting frequency). The two channels are finally weighted and fused using an attention mechanism, outputting a unified fused feature vector. This vector contains the spatiotemporal correspondence between the current power grid operating status and user behavior characteristics.

[0053]

[0054] in, This represents the state features of power grid equipment extracted through a three-layer one-dimensional convolutional neural network. The input to this network is load data organized in a time series, and the output of each convolutional layer is passed to the next layer after ReLU activation. It is a representation of user behavior data extracted by two layers of LSTM. The input is the user behavior encoding sequence, and the output is the state vector of user behavior in each time segment. It is a weighted matrix generated by the attention mechanism, through... and The importance of each dimension is normalized to obtain the result; It is a fusion function that combines the features of power grid status and user behavior, and outputs a fused feature vector. The attention mechanism employs additive attention, which calculates and normalizes a weighted score for each user behavior vector and the power grid state vector to control the degree of influence of each channel on the final output.

[0055] Furthermore, after the fusion is completed, the system outputs It is a set of high-dimensional feature vectors with region-time granularity, each This represents the combined characteristics of the power grid status and user behavior in a certain area during that time period. All It is organized as a structured data table with fields including time index, region number, load characteristics, behavioral characteristics, and fused representation. This structure can be directly input into subsequent demand forecasting models.

[0056] S2. Based on the fused feature vector, a set of prediction vectors for future time periods is predicted through an improved gated cyclic unit network; wherein each vector in the set of prediction vectors represents the prediction result of user demand or power grid status for the corresponding time slice.

[0057] Specifically, the goal of this step is to use the fused feature vector output from the previous step. This paper establishes a user demand prediction and power grid state analysis model suitable for power grid scenarios. This model enables the proactive identification of potential user electricity demand, equipment failure risks, and high load conditions, providing spatiotemporal predictive input for the subsequent recommendation engine. This step plays a crucial role in the entire patent system, transforming the static fusion characteristics of "power grid state-user behavior" into predictive dynamic information. It is a key link in enabling proactive recommendations and precise interactions in the intelligent customer service system.

[0058] Furthermore, the input is the fused feature vector output from step S1. ,in The status of power grid equipment has been integrated. With user behavior characteristics The time series representation is described. To enable the model to accurately predict future power grid operating conditions and user behavior trends, this step designs a hybrid time series prediction model that combines physical constraints with behavioral pattern learning. In power grid scenarios, changes in equipment load often have a nonlinear coupling relationship with user behavior. Simply relying on data-driven predictions can easily lead to overfitting, while physical modeling alone struggles to capture changes in user behavior characteristics. Therefore, this model adopts a dual-structure strategy of "power grid physical constraints + user behavior memory" in its design.

[0059] Furthermore, the first part of the model is used to model grid state changes, employing an improved gated cyclic unit (GRU) network to capture the evolution of power load over multiple time periods. Unlike conventional GRUs, this model introduces grid stability constraints into the gating mechanism to express the system's physical inertia. For example, when the load fluctuates drastically in a short period, the constraint term limits the model's overfitting to outliers. The hidden state update of the model can be represented as:

[0060]

[0061] in, The power grid state is represented by the time indicated. The hidden state; To update the gate vector; To fuse features at time Input; This is the hidden layer weight matrix; The grid stability constraint coefficient controls the intensity of the influence of physical constraints on the model. This represents a system stability index calculated based on actual power grid data, which can be obtained through historical load standard deviation or equipment health functions. When system load fluctuations are significant... Increase the size of the constraint to enhance smoothness and prevent the model from making non-physical predictions under sudden disturbances.

[0062] Furthermore, the second part of the model models user behavior trends, employing a two-layer stacked Temporal Convolutional Network (TCN) combined with an attention layer to identify clusters of user behavior within a certain time window. For example, if a certain type of user frequently queries "power outage notification" or "voltage anomaly" within a short period, it indicates a potential increase in service risk in that area. The attention mechanism automatically adjusts weights based on the importance of user behavior, focusing the model on representative behavioral patterns. The output is a representation of user behavior. Defined by the following formula:

[0063]

[0064] in, The user behavior embedding sequence at time 1 The query, key, and value matrix is ​​generated by the pre-convolutional layer; The dimension of the key vector; A sparse matrix encoding representing a user's historical behavior records the distribution of user behavior types at several recent moments; This is a behavior sparsity adjustment coefficient, used to enhance the model's focus on users with frequent abnormal behaviors. By adding... The model can adaptively enhance its ability to identify short-term, high-frequency behavioral patterns, thereby identifying potential high-demand users in advance within the customer service system.

[0065] Representing the power grid state User behavior representation After concatenation, a joint prediction vector is output through a single linear mapping layer. This vector represents the future. The model predicts the intensity of various types of electricity demand or grid events that users may experience within a given time slice. The loss function design incorporates a dual optimization mechanism combining grid security constraints and behavioral prediction objectives. In addition to the traditional cross-entropy loss, a grid state smoothing regularization term is added to ensure the physical consistency of the model's predictions over continuous time. The overall optimization objective is as follows:

[0066]

[0067] in, This is the total loss function; For real labels, indicating the time. Did the first one occur? Class-specific events; Predict probabilities for the model; Weighting coefficients are applied to event categories to reflect the importance of different events to the business. The coefficients for the smoothing regularization term are used to limit abrupt changes in the predicted state between adjacent time slices, making the model more stable and reliable in power grid scenarios. The second term... This is the power grid state smoothing term, which has a clear physical meaning and prevents the model output from fluctuating frequently or becoming overly sensitive.

[0068] Furthermore, during model training, a sliding time window method is used to generate training samples, with each sample covering the previous one. Historical Inputs and Future of a Time Slice The prediction target is set for each time slice, and the window sliding step size can be set according to the sampling frequency of the power grid data (e.g., sliding once every 5 minutes). In this way, the model can continuously learn the latest power grid state and user behavior evolution patterns, achieving online updates and dynamic predictions.

[0069] The output is a set of prediction vectors. each This represents the predicted user demand or power grid status for the corresponding time slice. Specifically, it includes user-level demand types (such as equipment repair requests, power outage inquiries, and abnormal power consumption) and system-level risk warnings (such as overload and declining equipment health). The prediction results will be used as input to the intelligent recommendation engine in the next step to generate targeted recommended content and service strategies.

[0070] The innovation of this step is reflected in three aspects: First, it introduces a power grid stability constraint term. With state smoothing regularization term This approach combines data-driven models with the physical characteristics of the power grid, ensuring the rationality and interpretability of prediction results. Secondly, it adds sparse behavioral reinforcement terms to the behavioral channels. The model can identify high-frequency abnormal consultation groups, providing early warning capabilities for the customer service system; thirdly, the overall model structure takes into account the temporal differences between the power grid operation rules and user behavior characteristics, achieving a qualitative improvement from passive response to proactive prediction in the intelligent customer service recommendation scenario.

[0071] S3. Select the set of demand events that are higher than the threshold from the set of predicted vectors, and generate a set of recommended content based on the set of demand events and the fused feature vector; perform emotion recognition based on the user's current input text to obtain a joint emotion vector, and calculate the recommended content score and push priority score in combination with the set of recommended content.

[0072] Specifically, this step aims to utilize the set of predicted vectors output from the previous step. fused input features with corresponding time slices This patent constructs a recommendation content generation and push mechanism for intelligent customer service scenarios in the power grid. Unlike traditional recommendation systems, the power industry, in which this patent operates, is characterized by high-voltage service, significant fault impact, and complex user emotions. Therefore, the recommendation system must not only generate personalized content but also consider business rationality, user acceptance, and system controllability, truly serving the patent objective of "interactive content recommendation."

[0073] Input in progress, This refers to the probability distribution of future user demand events predicted in step two through power grid state analysis and user behavior trend modeling. The time-series input features fused in step one include the grid load status, equipment operating status, user query records, and historical behavior vectors for the current time slice. In customer service operations, recommended content should simultaneously satisfy three dimensions of adaptability: semantic relevance (the recommended content matches the current user scenario), behavioral matching degree (the recommended content conforms to user preferences or historical feedback), and contextual timeliness (the recommended content is relevant at the current time). Therefore, this step proposes a two-stage recommendation engine: the first stage generates the recommended content, and the second stage determines the push strategy.

[0074] Furthermore, the content generation module, based on event-driven modeling principles, will use the prediction results... Consider it as a "behavioral motivation" triggered by the recommendation, and combine it with the context. The system generates structured recommended content. First, it filters out... Medium above the threshold demand event set and for each event type The recommended content structure is constructed, including recommended text, explanatory statements, and action suggestions. To enhance the explanatory power of the content and its relevance to service actions, a confidence-based content scoring function is introduced. This is used to assess how well the recommended content fits the current context.

[0075]

[0076] in, For event types The credibility rating of the recommended content; It is an event The predicted probability; For the current input features Event topic vector cosine similarity, It is a content center vector extracted from historical samples of this type of event; This represents the estimated user acceptance rate after the recommendation is triggered in the current scenario, and is obtained by modeling historical push-feedback data using a lightweight feedforward neural network. This represents a content redundancy penalty factor, which measures whether the currently recommended content duplicates content that the user has recently received. It is calculated by... Recommended topic vectors and The average similarity was obtained; To control the coefficients of different influencing factors, this scoring function integrates prediction probability, semantic matching degree, acceptance tendency, and content novelty, ensuring that the generated recommended content has both task relevance and user acceptability.

[0077] For each selected recommended content structure, the system matches a corresponding structured template based on the event type. The template consists of three parts: "Recommendation Main Sentence + Recommendation Explanation + Operational Suggestion." For example, the main sentence might be "There is a risk of overload in your area," the explanation would be "According to equipment operation data, voltage fluctuations have exceeded the normal range within 15 minutes," and the operational suggestion would be "Please avoid using high-power appliances during this period, or contact on-duty customer service." The content slots in the template (such as "equipment operation data," "voltage fluctuations," etc.) are selected from... The system dynamically extracts key information and automatically fills in the blanks, ensuring the content's adaptability to the current context. This structured generation method ensures semantic accuracy and professional expression, while also making subsequent review and manual intervention more efficient and controllable.

[0078] After the recommended content is built, the push strategy module determines whether to push, when to push, and how to push. This module introduces a recommendation push priority function. It is used to filter the optimal push content from multiple candidate content and dynamically control the push strategy:

[0079]

[0080] in, It is an event Push priority score; Rate the recommended content; This is the joint emotion vector (output by the subsequent sentiment analysis module; a larger value indicates a more intense emotion). The user's historical acceptance rate for this type of recommendation is provided by the behavioral profile module; This indicates the current load level of the customer service system (such as the number of users currently in the queue or the density of fault events in the current area), which is obtained in real time through the power grid dispatching system. The weight parameters, adjusted for experience (e.g., 0.4, 0.3, 0.2, 0.1), are used to balance the weight relationship between content quality, user emotional state, behavioral preferences, and system resource pressure.

[0081] The joint sentiment vector is generated through a dual-channel fusion model: a text sentiment channel and a behavioral sentiment channel. The text channel uses a pre-trained language model BERT (base version, 12 layers, 768 hidden dimensions), with the input being the text sequence within the user's current dialogue window. The output is a context-aware sentence vector. To address the sparsity of Chinese semantics, a custom emotion word attention layer is added after the BERT output. This layer assigns higher weights to emotional keywords (such as "angry," "out of power," and "complaint") to enhance the model's sensitivity to emotion-related vocabulary. This attention layer is built upon a keyword vocabulary derived from frequently used user complaint terms in customer service logs over the past year. The behavior channel input is the user's short-term behavior vector. Features include the number of queries within the past 15 minutes, keyword category distribution (e.g., whether they are concentrated in the "fault" category), page dwell time, and whether users attempted to switch to human customer service. This vector is processed through a two-layer GRU (64 units per layer) to generate a behavioral sentiment representation. The GRU model uses labeled data alignment during training. The labels are derived from historical emotional human annotation records (such as service evaluations of "emotionally agitated") or existing human-generated event identifiers within the system.

[0082] The outputs of the two channels mentioned above will be merged during the fusion stage by splicing. and Obtain the joint sentiment vector .

[0083] The final system will be based on Select several recommended items, sort them by priority, and then put them into the push queue; if Exceeding a certain threshold and If the value is below a threshold (e.g., 0.3), the recommended content will be marked as "immediate push"; otherwise, it will be placed in a delayed push list or submitted to customer service for review. Push methods include text notifications, voice synthesis broadcasts, WeChat pushes, etc., determined by the push preference field in the user profile.

[0084] The output consists of two parts: one is a set of recommended content. Each recommendation includes fields such as recommendation text, explanation, operation suggestions, and credibility score; secondly, the push strategy vector. This information guides the delivery method and priority of each recommended piece of content. This content will then be passed to the intelligent customer service front-end system for the final service interaction presentation.

[0085] S4. Introduce a cross-modal consistency regularization term based on the joint sentiment vector to modify the push strategy vector, and obtain the adjusted recommendation push score and the corresponding set of recommended content.

[0086] Specifically, this step aims to implement a sentiment analysis and interaction optimization mechanism for a smart customer service system for the power grid. The goal is to utilize users' natural language text and behavioral signals during interactions to identify their emotional state in real time, and then adjust the tone and push strategy of the system's recommended content accordingly. This will provide a more personalized and humanized interactive experience in high-voltage service scenarios such as power outages and load anomalies. This mechanism uses the set of recommended content already output in step three. With push rating Based on, combined with the user's current input text By integrating historical behavior records, and enabling intelligent filtering and dynamic adjustment of recommended content, a key step in the transformation of intelligent customer service systems from "static response" to "emotion-aware proactive interaction" is achieved. The input for this step includes three types of data sources. The first type is... A collection of recommended content structures generated by the recommendation engine, with each recommended item... This includes fields for recommended topics, contextual explanations, and actionable suggestions; the second category is... The first category represents the priority score for each recommendation, derived from the multi-factor scoring mechanism in the recommendation engine; the second category is the user's current natural language input. This includes the needs and opinions expressed by users through text conversations or voice transcription. In addition, the system also needs to access users' recent interaction history, including short-term behavioral characteristics (such as query frequency within 5 minutes and keyword category distribution) and medium-term behavioral characteristics (such as the number of complaints within this week and whether they have been upgraded to human customer service). This data is extracted from the intelligent customer service platform logs and the user profiling system.

[0087] In the joint sentiment vector Based on this, a cross-modal consistency regularization term is further introduced. This constrains the representational distance between the two channels, ensuring logical consistency between the emotional expression in language and behavior, and enhancing discrimination ability, especially in scenarios where emotional expression is unclear but behavior is drastically abnormal. Its regularization term... The definition is as follows:

[0088]

[0089] in, This is the sentiment vector output from the text channel. The emotion vector output by the behavioral channel. The weight parameter for controlling semantic-behavior consistency constraints is typically set to 0.2. This parameter, along with cross-entropy, participates in the optimization training of the total loss function of the emotion recognition model, enabling the model to have cross-modal self-correction capabilities. This is particularly suitable for users in power grid customer service scenarios who often exhibit "indistinct emotional expression but strong behavior."

[0090] Furthermore, after completing emotion recognition, the system will The output is used to dynamically adjust the push strategy and tone of recommended content. The adjustment of the push strategy is based on the push score. Based on this, an emotion adjustment coefficient is introduced into the original rating to avoid immediately pushing content to users when they are in an overly emotional state, which could cause offense. The adjustment function is designed as follows:

[0091]

[0092] in, The adjusted recommendation score. This is the original recommendation rating. This is a normalized index of emotional intensity (between 0 and 1). This is a push notification delay adjustment factor, typically set around 0.3. It means that when a user is emotionally agitated, the pace of push notifications will be appropriately slowed down to guide them towards calm. The system sets a threshold, for example, when... At this time, the currently recommended content will not be pushed out, but will be slowly loaded in the user interface in the form of "preparing", and the push will be triggered after the emotions have subsided.

[0093] Furthermore, regarding tone and expression, the recommended templates will adjust content slots based on the intensity of emotion. The system maintains three types of tone style templates: formal, neutral, and reassuring. For example, for events like "overload," the formal template is "We suggest you turn off some high-power devices," the neutral template is "The current electricity consumption in the area is high; appropriate energy saving will be safer," and the reassuring template is "We have noticed that the current load in the area is high; to ensure your electricity safety, we suggest making some adjustments." Template switching logic is implemented. Value segmentation control, such as Use the official template at times. Use a neutral template. A soothing template is used. The output consists of two parts: a structured emotional state vector. The first aspect involves adjusting recommended content during the current round of conversation and tracking user status in subsequent rounds of conversation; the second aspect involves the set of emotion-adjusted recommended content and its push strategies. This includes fields such as adjusted text content, tone style category, and suggested push method, which are directly displayed and executed by the intelligent customer service front-end.

[0094] S5. The adjusted recommendation push score and the corresponding set of recommended content are converted into a scheduling task and executed in conjunction with the scheduling system.

[0095] Specifically, this step aims to integrate the sentiment-modified recommended content from the previous stage with the push strategy set. This translates into practically executable power grid operation and maintenance response actions, achieving a service loop from user interaction to power grid resource coordination. Traditional intelligent customer service systems typically only handle information display and user communication. However, the system designed in this patent not only identifies user needs and emotional states but also proactively coordinates resources from multiple departments, including power grid dispatching, maintenance, and emergency repair, enabling intelligent response and optimized scheduling for emergencies and high-load areas. This mechanism has significant practical implications for power grid companies: when facing regional faults, sudden load surges, or concentrated user inquiries, it can shorten the processing and response chain, improve user satisfaction, and reduce system load pressure.

[0096] The input for this step is ,in This indicates that the recommended content has been optimized for emotional impact, and includes a clear event type, related area, and problem manifestation (e.g., "Users in this area have reported a concentrated number of unstable voltage issues"). This is a comprehensive score for the content, taking into account multiple factors such as user sentiment, content credibility, and historical acceptance preferences. Additionally, this step requires access to real-time power grid operation status data. This data is provided by the energy management system (EMS) or distribution network automation system (D-SCADA) of the provincial or municipal dispatch center and is updated periodically through boundary interface services (such as CIM / XML interface or WebSocket API). It contains structured data such as the load factor of each distribution area, load growth rate, current fault alarm list, emergency repair work order execution status, and dispatch load. Each field has a timestamp, area number, and multi-dimensional indicator value. For example, A data entry might be "Area ID: 0401, Current load rate: 92%, Growth rate in 24 minutes: +6%, Status: Non-alarm".

[0097] Furthermore, the system first bases its recommendations on content. The geographic area code field carried in it, and its relationship with Align the runtime status data. For example, The event topic is "Concentrated complaints about low voltage in this area", and the area number is 0401. The system will search for... For record number 0401, its current load rate, voltage fluctuation value, and fault status fields are read, and the system is evaluated to determine if there is objective data to support the recommendations. If a strong consistency match exists (e.g., voltage fluctuation amplitude is below 95% of the rated value for 15 consecutive minutes), the system enters the "task matching" process.

[0098] To determine whether the recommendation needs to be converted into a scheduled task, the system designed a task activation scoring function. This function integrates push ratings Regional operational pressure indicators and current scheduled resource load Three factors:

[0099]

[0100] in, Source The system uses a multi-index aggregation function to normalize the power grid operation indicators of the corresponding region, such as load factor, voltage fluctuation frequency, and average fault recovery time, to a risk score between 0 and 1. It indicates the current load status of the entire network's scheduling resources, including the number of scheduling / work order tasks currently in execution, the maximum concurrent task capacity set by the system, etc., reflecting the availability of current response resources; and This is an adjustable parameter used to balance the weighting of risk-driven factors and resource scheduling. For example, when... , , , , hour, This indicates that the task has high activation value and should be included in the scheduling process.

[0101] when If the value exceeds a set threshold (e.g., 0.85), the system will construct a service task request. Its structure includes fields such as task type (e.g., load adjustment, fault inspection, work order dispatch), associated device ID, recommended source, expected response window (e.g., response within 15 minutes), and suggested execution method (automatic / manual). Scheduling tasks are pushed to the scheduling management system or operations and maintenance system (e.g., OMS) via API. Task execution is assigned by the system based on existing personnel schedules, equipment online status, and other scheduling rules. For tasks involving manual operation, such as "maintenance notification" or "temporary supply change," the system will integrate GIS system data with operations and maintenance personnel location data, prioritizing the nearest repair personnel to the target area, and automatically generating a work order with navigation guidance, which is then pushed to their mobile terminal.

[0102] To ensure that the results of task execution can have a positive impact on system decision-making, a feedback channel mechanism is designed. After the task is completed, the system... The system retrieves records of changes in the operational status of the area and, in conjunction with user feedback or complaint data, calculates a task response effectiveness score. The scoring structure is as follows:

[0103]

[0104] in, This indicates the magnitude of change in the power grid pressure index in the region after the task is executed; This indicates the degree of improvement in user feedback, such as a decrease in user complaints and the adoption of recommended content. and To adjust the weights, adjust them according to the actual business needs (e.g., setting them when more emphasis is placed on physical metrics). The value is 0.7, but it is increased when user feedback is involved. The system will be high. The recommendation-task link records are used as positive samples and enter the "reward mechanism" of the model reinforcement learning to optimize future recommendation content selection strategies and task activation mechanisms.

[0105] like Figure 2 As shown, another embodiment of the present invention provides an intelligent customer service interaction content recommendation system based on artificial intelligence, the system comprising:

[0106] Data fusion module 301 is used to acquire power grid equipment status data and user behavior data, and output a fused feature vector through weighted fusion using an attention mechanism;

[0107] The predictive analysis module 302 is used to predict a set of prediction vectors for future time periods based on the fused feature vectors through an improved gated cyclic unit network; wherein each vector in the set of prediction vectors represents the prediction result of user demand or power grid status for the corresponding time slice.

[0108] The recommendation engine module 303 is used to filter out the set of demand events that are higher than a threshold from the set of predicted vectors, generate a set of recommended content based on the set of demand events and the fused feature vector; perform emotion recognition based on the user's current input text to obtain a joint emotion vector, and calculate the recommended content score and push priority score in combination with the recommended content set;

[0109] The sentiment analysis module 304 is used to introduce a cross-modal consistency regularization term based on the joint sentiment vector to modify the push strategy vector, so as to obtain the adjusted recommendation push score and the corresponding set of recommended content.

[0110] The operation and maintenance collaboration module 305 is used to convert the adjusted recommendation push score and the corresponding set of recommended content into a scheduling task, and execute it in conjunction with the scheduling system.

[0111] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0112] In addition, for technical details not described in detail in this embodiment, please refer to the parameter operation method provided in any embodiment of the present invention, which will not be repeated here.

[0113] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0114] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0115] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0116] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. An AI-based intelligent customer service interaction content recommendation method, characterized in that, The method includes: S1. Acquire power grid equipment status data and user behavior data, and output a fused feature vector by weighted fusion through an attention mechanism; S2. Based on the fused feature vector, a set of prediction vectors for future time periods is predicted using an improved gated cyclic unit network; wherein each vector in the set of prediction vectors represents the prediction result of user demand or power grid status for the corresponding time slice; S3. Select the set of demand events that are higher than the threshold from the set of predicted vectors, and generate a set of recommended content based on the set of demand events and the fused feature vector; perform emotion recognition based on the user's current input text to obtain a joint emotion vector, and calculate the recommended content score and push priority score in combination with the set of recommended content. S4. Introduce a cross-modal consistency regularization term based on joint sentiment vector to modify the push strategy vector, and obtain the adjusted recommendation push score and the corresponding set of recommended content. S5. The adjusted recommendation push score and the corresponding set of recommended content are converted into a scheduling task and executed in conjunction with the scheduling system.

2. The method for recommending intelligent customer service interactive content based on artificial intelligence according to claim 1, characterized in that, The power grid equipment status data includes the operating parameters of equipment in the transmission, distribution, and terminal metering stages; the user behavior data includes users' electricity consumption curves, inquiry content, and historical electricity consumption events.

3. The method for recommending intelligent customer service interactive content based on artificial intelligence according to claim 1, characterized in that, The weighted fusion output of the fused feature vector through the attention mechanism specifically includes: The power grid equipment status data is input into the power grid equipment channel to extract features from the time series load curve, thereby obtaining the power grid equipment status features. The lengths of each convolutional kernel are 3, 5, and 7, respectively, to capture short-term fluctuations and medium-term trends. The power grid equipment channel is a three-layer convolutional network. The user behavior data is input into the user behavior channel to obtain the state vector of user behavior in each time segment; The outputs of the power grid equipment channel and the user behavior channel are weighted and fused using an over-attention mechanism to output a fused feature vector.

4. The method for recommending intelligent customer service interactive content based on artificial intelligence according to claim 1, characterized in that, The improved gated cyclic unit network introduces a grid stability constraint term into the gating mechanism to express the system's inertia at the physical level and generate a grid state representation. The grid stability constraint term represents a system stability index calculated based on actual grid data, obtained through historical load standard deviation or equipment health function.

5. The method for recommending intelligent customer service interactive content based on artificial intelligence according to claim 4, characterized in that, The improved gated recurrent unit network also adopts a structure combining two stacked temporal convolutional networks with an attention layer to generate user behavior representations, which are used to identify user behavior clustering phenomena within a certain time window. The set of prediction vectors for predicting future time periods is specifically as follows: After concatenating the power grid state representation and the user behavior representation, a set of prediction vectors is output through a linear mapping layer. Each prediction vector in the set of prediction vectors represents the user demand or power grid state prediction result for the corresponding time slice. The linear loss function includes a power grid security constraint and a behavior prediction objective; the power grid security constraint is used to prevent frequent fluctuations or oversensitivity in the model output; the behavior prediction objective is cross-entropy loss.

6. The method for recommending intelligent customer service interactive content based on artificial intelligence according to claim 1, characterized in that, The recommended content set includes recommended text, explanatory statements, and operation suggestions; Specifically, the step of performing emotion recognition based on the user's current input text to obtain a joint emotion vector is as follows: The user's current input text is input into a dual-channel fusion model, which includes a text sentiment channel and a behavioral sentiment channel. The text sentiment channel uses a pre-trained language model BERT to output a sentiment vector, while the behavioral sentiment channel inputs a user behavior representation and outputs a sentiment vector through two layers of gated recurrent units. The sentiment vectors are then concatenated to obtain a joint sentiment vector.

7. The method for recommending intelligent customer service interactive content based on artificial intelligence according to claim 6, characterized in that, The recommended content score is generated by calculating the predicted vector set and the event topic vector, and is used to evaluate the degree of fit between the recommended content and the current scenario; The push priority score is calculated and generated based on the recommended content score, the joint sentiment vector, the user's historical acceptance rate on the current type of recommendation, and the current load level of the customer service system. Then, based on the push priority score, several recommended content items are selected, sorted by priority, and put into the push queue; if the push priority score exceeds a certain threshold and the joint sentiment vector is below the critical value, the recommended content will be marked as "push immediately", otherwise it will be put into the delayed push list or handed over to customer service for review.

8. The method for recommending intelligent customer service interactive content based on artificial intelligence according to claim 6, characterized in that, The cross-modal consistency regularization term is calculated and generated based on the emotion vector and the emotion vector.

9. The method for recommending intelligent customer service interactive content based on artificial intelligence according to claim 1, characterized in that, Specifically, S5 is: Based on the geographic region code field carried in the recommended content set, it is aligned with the operating status data in the real-time power grid equipment status data. If the alignment matches, the task matching process begins. The task matching includes: Based on the adjusted recommendation push score, and combined with the power grid operation indicators of the corresponding region and the current load status of the network-wide scheduling resources, the task activation score is calculated. When the task activation score exceeds a set threshold, the system will construct a scheduling task request, which includes the task type, associated device ID, recommendation source, expected response window, and suggested execution method. The scheduling task request is pushed to the scheduling management system or operation and maintenance system via API interface.

10. An intelligent customer service interaction content recommendation system based on artificial intelligence, characterized in that: The system includes: The data fusion module is used to acquire power grid equipment status data and user behavior data, and outputs a fused feature vector through weighted fusion using an attention mechanism; The predictive analysis module is used to predict a set of prediction vectors for future time periods based on the fused feature vectors through an improved gated cyclic unit network; wherein each vector in the set of prediction vectors represents the prediction result of user demand or power grid status for the corresponding time slice. The recommendation engine module is used to filter out the set of demand events that are higher than a threshold from the set of predicted vectors, generate a set of recommended content based on the set of demand events and the fused feature vector; perform emotion recognition based on the user's current input text to obtain a joint emotion vector, and calculate the recommended content score and push priority score in combination with the recommended content set; The sentiment analysis module is used to introduce a cross-modal consistency regularization term based on the joint sentiment vector to modify the push strategy vector, so as to obtain the adjusted recommendation push score and the corresponding set of recommended content. The operation and maintenance collaboration module is used to transform the adjusted recommendation push score and the corresponding set of recommended content into scheduling tasks, and execute them in conjunction with the scheduling system.

Images