Electric power marketing data analysis method based on AI large model

Abstract

The invention relates to the technical field of power marketing, in particular to an AI large model-based power marketing data analysis method, which comprises the following steps of: acquiring structured data and unstructured data in a power marketing system, and generating unified coded data after space-time alignment and pre-training word embedding model processing; inputting the unified coding data into a pre-trained power field large model, and extracting static, dynamic and semantic features through multi-modal fusion, feature decoupling and semantic anchoring; constructing entity link feature pairs in combination with a power knowledge graph, and enhancing fusion feature expression through a graph attention mechanism; and finally, inputting a dynamic weight gating network, outputting an abnormal user identification tag, a demand response strategy and a customer loss early warning probability, and executing strategy optimization under specific conditions. The method can be widely applied to risk identification, strategy making and user behavior prediction tasks in power marketing.

Description

Technical Field

[0001] The present invention relates to the field of power marketing technology, and in particular to a power marketing data analysis method based on an AI big model. Background Art

[0002] With the rapid development of new power systems and the gradual improvement of power market mechanisms, power marketing systems are shifting from a traditional, static management model centered around meter reading and billing to an intelligent operation model based on user behavior analysis, load forecasting, and precise service delivery. Power companies have accumulated a vast amount of structured data (such as user profiles, load curves, and electricity price information) and unstructured data (such as customer service texts, inspection images, and equipment logs). Efficiently integrating this heterogeneous data, extracting key features, and performing intelligent identification and prediction has become a core challenge for improving marketing efficiency, achieving refined regulation, and providing proactive services.

[0003] However, existing power marketing data analysis methods still have the following key technical bottlenecks: First, multi-source data lack a unified encoding mechanism in terms of time scale, spatial granularity and expression structure, resulting in inconsistent modeling input and difficulty in fusion; second, traditional machine learning and lightweight neural networks have difficulty processing high-dimensional and complex features across modalities and tasks, and cannot achieve simultaneous linkage output for anomaly identification, strategy push and customer warning; third, knowledge graphs and expert rules have not been effectively integrated into the model, resulting in a lack of semantic relevance and causal interpretability in strategy recommendations. Summary of the Invention

[0004] The present invention provides an electric power marketing data analysis method based on an AI big model, an AI analysis method that integrates semantic enhancement, multi-task collaboration, and strategy optimization capabilities.

[0005] The power marketing data analysis method based on the AI ​​big model includes the following steps: S1: Collect structured and unstructured data from the power marketing system, unify timestamps and spatial encodings through the spatiotemporal alignment module, and generate unified encoding data using a pre-trained word embedding model; S2: Input the uniformly encoded data into a pre-trained large model for the power sector, separate the static feature vector, the dynamic feature vector, and the semantic feature vector through a feature decoupling layer, and output a decoupled feature vector; S3: Entity linking the decoupled feature vector with the power knowledge graph, where the knowledge graph includes the electricity price policy entity and the equipment fault tree entity, and generating a fused feature vector through a graph attention mechanism; S4: Input the fused feature vector into a dynamic weight gating network, adjust the feature weight according to the real-time electricity price policy, and output abnormal user identification labels, demand response strategies and customer churn warning probability.

[0006] Optionally, the structured data includes user profiles and electricity consumption time series data, and the unstructured data includes customer service work order texts and equipment images.

[0007] Optionally, the pre-trained word embedding model includes a four-channel encoder and a fusion output layer, specifically: The user profile encoder uses a two-layer fully connected network to convert the user profile numeric fields into user profile feature tensors; The time series data encoder is composed of a cascade of one-dimensional convolutional layers and LSTM layers, which converts the resampled power consumption time series data into a power consumption time series vector; The text encoder is based on the BERT architecture and introduces a substation encoding embedding layer. It uses the location-aware attention mechanism to generate a customer service ticket text vector based on the customer service ticket text with spatiotemporal anchor point labels. The image encoder uses the ResNet-18 backbone network, with a spatial coding fusion module added to its input layer to convert device images with spatial coding and timestamps into device image vectors. The fusion output layer uses residual connections to concatenate the user profile feature tensor, electricity consumption time series vector, customer service ticket text vector, and device image vector into unified encoded data.

[0008] Optionally, the S1 includes: S11: Collect structured and unstructured data from the power marketing system. Structured data includes user profiles and electricity consumption time series data, and unstructured data includes customer service ticket texts and equipment images. Output the original multi-source data. S12: Performing a spatiotemporal alignment operation on the original multi-source data, specifically including resampling the power consumption time series data based on a 15-minute time stamp, adding a spatiotemporal anchor tag containing the spatial code of the work order location and the work order creation timestamp to the customer service work order text, and adding the spatial code of the shooting location and the shooting timestamp to the device image, and outputting the aligned multi-source data; S13: The aligned multi-source data is input into a pre-trained word embedding model for processing. The user profile is converted into a user profile feature tensor through a two-layer fully connected network. The resampled electricity consumption time series data is processed in cascade by a one-dimensional convolutional layer and an LSTM layer to generate an electricity consumption time series vector. The customer service work order text with spatiotemporal anchor point labels is generated into a customer service work order text vector through a location-aware attention mechanism containing a substation code embedding layer. The device image with spatial coding and timestamp is generated into a device image vector through a ResNet-18 network with an additional spatial coding fusion module. Finally, the user profile feature tensor, electricity consumption time series vector, customer service work order text vector and device image vector are spliced ​​into unified coding data in the fusion output layer.

[0009] Optionally, the pre-trained large power domain model includes a cascaded three-layer structure: a multimodal fusion layer, a feature decoupling layer, and a semantic anchoring module, wherein; The multimodal fusion layer is composed of a cross-modal attention mechanism. It receives the user profile feature tensor, power consumption time series vector, customer service ticket text vector, and device image vector in the unified encoded data, and generates a multimodal feature tensor through feature cross-gating. The feature decoupling layer includes three physically isolated parallel branch processing channels. The static feature branch uses a two-layer fully connected network to process user profile-related inputs. The dynamic feature branch uses a stack of time-series convolutional networks with expansion rates of [1, 2, 4] to process power consumption time series vectors. The semantic feature branch uses a BiLSTM-CRF sequence labeling model to process text OCR recognition results in customer service ticket text vectors and device image vectors. The semantic anchoring module has a built-in semantic anchor word library in the power field, and performs keyword matching and standard fault code binding operations on the intermediate output of the semantic feature branch before outputting the decoupled feature vector.

[0010] Optionally, the S2 includes: S21: Input the uniformly encoded data output by S1 into the multimodal fusion layer, and perform feature cross-gating calculations on the user profile feature tensor, power consumption time series vector, customer service ticket text vector, and device image vector through the cross-modal attention mechanism to generate a multimodal feature tensor; S22: Input the multimodal feature tensor into the three-way hardware isolation processing channel of the feature decoupling layer: the static feature branch extracts the user's credit rating and electricity consumption category through a two-layer fully connected network to form a static feature vector; the dynamic feature branch extracts the electricity consumption fluctuation pattern through a stack of temporal convolutional networks with an expansion rate of [1, 2, 4] to form a dynamic feature vector; and the semantic feature branch processes the customer service ticket text vector and the device image OCR recognition result through a BiLSTM-CRF sequence labeling model to form an original semantic vector. S23: Input the original semantic vector into the semantic anchoring module, match the standard fault code in the semantic anchor vocabulary in the power field for binding enhancement, output the enhanced semantic feature vector, and finally splice the static feature vector, dynamic feature vector and semantic feature vector into a decoupled feature vector according to the channel dimension.

[0011] Optionally, the S3 includes: S31: Match the static feature vectors in the decoupled feature vectors with the electricity price policy entity in the power knowledge graph to the electricity consumption category. At the same time, topologically link the equipment fault keywords in the semantic feature vector to the equipment fault tree entity to generate entity link feature pairs. The electricity price policy entity includes the real-time floating electricity price rule, and the equipment fault tree entity includes the fault propagation path matrix. S32: Based on the entity link feature pair, extract the historical execution effect vector of the electricity price policy entity and the maintenance knowledge vector of the equipment fault tree entity from the power knowledge graph, load the feature attributes of adjacent entities through a breadth-first graph traversal algorithm, and output a graph structure enhancement feature that integrates historical policy data and maintenance experience; S33: The graph structure enhanced features and the original decoupled feature vector are input into the graph attention mechanism together, wherein the static feature vector and the electricity price policy entity feature calculate the strategy decision weight, the dynamic feature vector and the load curve pattern entity calculate the timing association weight, and the semantic feature vector and the equipment fault tree entity calculate the fault propagation weight, and a fusion feature vector is generated through the gated residual connection.

[0012] Optionally, the dynamic weight gating network comprises a cascaded three-layer structure: a feature weighting layer, a multi-task decision layer, and a strategy optimization layer; wherein The feature weighting layer consists of a policy response coefficient generator and a feature gating unit. It receives real-time electricity price policy input and generates credit rating weight coefficients, load sensitivity weight coefficients, and policy response weight coefficients. It multiplies the static feature vectors, dynamic feature vectors, and graph structure enhancement features in the fusion feature vector with the corresponding weight coefficients and then concatenates them into a weighted feature tensor. The multi-task decision layer consists of three physically isolated parallel branch processing channels. The abnormal user identification branch uses a two-layer fully connected network to process the static feature components in the weighted feature tensor and output abnormal user identification labels. The demand response strategy branch integrates a rule engine and an LSTM time series decision maker to process dynamic feature components and graph structure enhancement features to generate demand response strategies. The customer churn warning branch uses the Cox proportional risk model to analyze the historical behavior feature components in the weighted feature tensor and output the customer churn warning probability. The strategy optimization layer has a built-in strategy collaboration module. When the abnormal user identification label is high risk and the customer churn warning probability exceeds the preset threshold, the customer care measure code is added to the demand response strategy.

[0013] Optionally, the S4 includes: S41: Input the fused feature vector output by S3 and the real-time electricity price policy into the feature weighting layer of the dynamic weight gating network. The credit rating weight coefficient, load sensitivity weight coefficient, and policy response weight coefficient are calculated through the policy response coefficient generator. The static feature vector, dynamic feature vector, and graph structure enhancement feature are multiplied by the corresponding weight coefficients using the feature gating unit and then concatenated into a weighted feature tensor. S42: The weighted feature tensor is input into the three-way hardware isolation processing channel of the multi-task decision layer. The abnormal user identification branch processes the static feature components in the weighted feature tensor through a two-layer fully connected network to output an abnormal user identification label. The demand response strategy branch processes the dynamic feature components and graph structure enhancement features through a rule engine and an LSTM time series decision maker to generate a demand response strategy. The customer churn warning branch analyzes the historical behavior feature components through a Cox proportional risk model to output a customer churn warning probability. S43: The abnormal user identification label, demand response strategy and customer churn warning probability are input into the strategy optimization layer. When the abnormal user identification label is high risk and the customer churn warning probability exceeds the preset threshold, the customer care measure code is added to the demand response strategy through the strategy collaboration module, and finally the optimized abnormal user identification label, demand response strategy and customer churn warning probability are output.

[0014] Beneficial effects of the present invention: This invention, through the spatiotemporal alignment module and pretrained word embedding model in S1, achieves the first unified encoding of multimodal data, including user profiles, electricity usage time series data, customer service ticket text, and device images. The four-channel architecture (fully connected network, convolutional + LSTM, BERT position-aware model, and ResNet-18 image encoder) works synergistically to establish a unified reference frame for structured and unstructured data based on timestamps and spatial coordinates. This provides a highly consistent and expressive data foundation for subsequent model input, thereby improving the generalization and robustness of data-driven models in marketing scenarios.

[0015] This method establishes semantic links between the decoupled static, dynamic, and semantic feature vectors and the electricity pricing policy, load curve pattern, and equipment fault tree entities, respectively. Graph traversal and graph attention mechanisms are then used to enhance context and weight semantics, significantly improving the model's understanding of policy adaptability and fault transmission mechanisms. Compared to traditional black-box modeling approaches based on pure neural networks, this method introduces entity semantic anchoring and topological semantic enhancement mechanisms, providing a more interpretable logical path for abnormal user identification and policy recommendation.

[0016] This invention, through a dynamic weighted gating network, establishes a three-level structure integrating a feature weighting layer, a multi-task decision layer, and a policy optimization layer. This not only dynamically adjusts the influence of various features in abnormal user identification, demand response decision-making, and customer churn warning, but also automatically triggers additional customer care measures under high-risk conditions. This solution not only improves identification accuracy and intervention response rates, but also significantly enhances the power marketing system's capabilities in user retention, load regulation, and service personalization, with excellent application scalability and scenario adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention; Figure 2 Schematic diagram of the S3 process of an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0020] like Figure 1-Figure 2 As shown in FIG, the power marketing data analysis method based on the AI ​​big model includes the following steps: S1: Collect structured and unstructured data from the power marketing system, unify the timestamps and spatial encodings through the spatiotemporal alignment module, and use a pre-trained word embedding model to generate unified encoding data. Specifically: S11, Raw multi-source data acquisition: By calling the open interfaces and database services of the power marketing system, the required structured and unstructured data are obtained. Structured data includes user profiles and electricity consumption time series data. User profiles are provided by the system's user information management module and include user ID, electricity price category, access type, energy consumption characteristics, and contract capacity. Electricity consumption time series data is collected by smart meters or centralized collection terminals and automatically aggregated into the data center at a 15-minute granularity through the power data collection system. Unstructured data includes customer service ticket text and equipment images. Customer service ticket text is obtained from historical records in the customer service system and covers user consultations, fault reports, and processing records, including text descriptions, submission times, and associated substation information. Equipment images are sourced from on-site inspection terminals, image acquisition modules, or drone inspection platforms and include on-site images of distribution cabinets, transformers, and terminal facilities, accompanied by automatically recorded shooting timestamps and location coordinates.

[0021] S12, spatiotemporal alignment of multi-source data: To unify the time and space identifiers of data from different sources, the electricity consumption time series data is timestamped in units of 15 minutes to ensure that the electricity consumption behavior of each user is comparable within a unified time window. For customer service work order texts, a "spatiotemporal anchor tag" is appended to the original text data. The tag consists of the work order creation timestamp and the location where the work order occurred. The location of the occurrence is mapped to the specific longitude and latitude through the substation or substation ID associated with the work order. The equipment image is appended with "spatial coding" and "timestamp tags" through the shooting timestamp recorded by the image acquisition terminal and the location data provided by the embedded GPS module, forming image data with clear spatiotemporal attributes. This ensures that structured data and unstructured data have a unified time reference system and spatial coordinate system, and outputs aligned multi-source data.

[0022] S13, unified coded data generation: The spatiotemporally aligned multi-source data is fed into a pre-trained word embedding model. The model consists of four independent channels and a fusion output layer, each processing different data types. User profiles are first processed through a user profile encoder, a submodule consisting of a two-layer fully connected network, which converts numeric fields such as contract capacity and voltage level into user profile feature tensors. Electricity consumption time series data is fed into a time series data encoder consisting of a cascade of one-dimensional convolutional layers and LSTM layers, which extracts short-term load fluctuations and long-term trend features to generate electricity consumption time series vectors. Customer service ticket text is fed into a text encoder based on the BERT architecture. This encoder incorporates a "substation code embedding layer" and, combined with added spatiotemporal anchor labels, uses a location-aware attention mechanism to extract key semantics and spatial location relevance from the text, generating a customer service ticket text vector. Device images are processed through an image encoder with an additional spatial encoding fusion module at the input layer. This encoder uses a ResNet-18 backbone network to extract spatial structure and texture information, generating device image vectors. The four types of encoded vectors are concatenated in the fusion output layer using a residual connection mechanism, ultimately outputting unified coded data as input for subsequent large-scale model processing.

[0023] S2: Input the uniformly coded data into the pre-trained large model for the power sector, separate the static feature vector, dynamic feature vector, and semantic feature vector through the feature decoupling layer, and output the decoupled feature vector, specifically: S21, Multimodal Feature Tensor Generation: The uniformly encoded data output from step S1 is sequentially input into the multimodal fusion layer of the pre-trained large-scale power sector model. This layer consists of multiple cross-modal attention mechanisms. Specifically, the fusion layer receives four types of input: user profile feature tensors, electricity consumption time series vectors, customer service ticket text vectors, and device image vectors. To exploit the complementary information and coupling relationships between various modalities, the fusion layer constructs cross-attention channels between each pair of input modalities. For example, a temporal-semantic correlation channel is established between the electricity consumption time series vector and the customer service ticket text vector, and a static attribute-spatial state coupling channel is established between the user profile feature tensor and the device image vector. After completing all cross-modal channel attention calculations, the fusion layer uses a feature cross-gating mechanism to dynamically adjust and fuse the modal feature tensors from different sources, suppressing redundant information and enhancing the contribution of key modalities. The final output is a unified multimodal feature tensor for subsequent feature decoupling.

[0024] S22, decoupled feature vector generation: The multimodal feature tensor is input into the feature decoupling layer in the large model. The decoupling layer consists of three physically isolated parallel processing channels: static feature branch, dynamic feature branch, and semantic feature branch. The static feature branch focuses on expressing user profile features and long-term stable attributes. This branch receives the user profile-related portion of the multimodal feature tensor, performs nonlinear transformation and dimensionality reduction processing through a two-layer fully connected network, extracts static information related to the user's credit rating and electricity usage category, and outputs a static feature vector. The dynamic feature branch is used to capture the changing trends and short-term fluctuation patterns of user electricity consumption behavior. This branch extracts the portion of the multimodal feature tensor derived from the electricity consumption time series vector. Using a stacked temporal convolutional network with a progressively expanding expansion rate, it constructs a multi-scale time window perception mechanism. It extracts dynamic behavior features, including intraday load spikes and periodic fluctuations, layer by layer, and outputs a dynamic feature vector. The semantic feature branch is aimed at expressing semantic information in customer service ticket text and device images. First, the customer service ticket text vector and the OCR recognition result text of the device image are extracted from the multimodal feature tensor, and then input into the BiLSTM-CRF sequence labeling model based on the combination of bidirectional long short-term memory network and conditional random field model. Sequence context modeling and entity boundary recognition are performed to extract text elements such as fault type, location description, and urgency in the repair event, and output the original semantic feature vector.

[0025] S23, Semantic Anchoring and Feature Concatenation: The original semantic feature vector is fed into the semantic anchoring module, which has a built-in semantic anchor vocabulary for the power sector, including standardized fault type terms, typical O&M terminology, and predefined entity class labels. The semantic anchoring module performs keyword matching on the original semantic feature vector. If a high-confidence match with a vocabulary item is detected, the corresponding standard fault code is bound and the feature representation capability of the semantic channel is enhanced. The module ultimately outputs the enhanced semantic feature vector.

[0026] At this point, the static feature vector, dynamic feature vector and enhanced semantic feature vector have all been extracted. The three are spliced ​​according to the channel dimension to form a structured decoupled feature vector, providing a high-quality feature foundation for subsequent knowledge graph fusion and downstream task processing.

[0027] S3: Entity linking is performed between the decoupled feature vector and the power knowledge graph, where the knowledge graph includes the electricity price policy entity and the equipment fault tree entity. A fused feature vector is generated through the graph attention mechanism. Specifically: S31, Entity Link Feature Pair Generation: Using the decoupled feature vector output by S2 as input, the static feature vector and semantic feature vector serve as the primary link objects. Feature matching and topological binding are performed with different entities in the power knowledge graph. Specifically, the electricity category label in the static feature vector is parsed and attribute fields are matched with the electricity pricing policy entity in the power knowledge graph. Based on the electricity category (e.g., residential, commercial, agricultural, etc.), the entity is linked to the corresponding electricity pricing policy entity node. Each electricity pricing policy entity contains multiple attribute fields, including information such as real-time floating price rules, time-of-use price thresholds, and policy effective time periods. Simultaneously, equipment fault keywords are extracted from the semantic feature vector after processing by the semantic anchoring module. Based on the keyword content, these keywords are linked to specific fault nodes in the equipment fault tree entity in the power knowledge graph. The equipment fault tree entity is organized topologically, internally containing component structure, common fault types, fault propagation paths, and associated maintenance records. Its core structure is the fault propagation path matrix. After completing the matching of the above static feature vectors with electricity price policy entities, and the semantic feature vectors with equipment fault tree entities, entity link feature pairs are generated as the initial input for graph fusion.

[0028] S32, Graph Structure Enhanced Feature Extraction: Based on the aforementioned entity-link feature pairs, further entity extension information is extracted from the power knowledge graph to enhance the contextual knowledge representation of the original features. For the electricity pricing policy entity, its associated historical execution records are retrieved and the corresponding historical execution effect vector is extracted. This vector reflects the user response rate, energy saving effect, and load regulation effectiveness of different electricity pricing policies during past implementation. For the equipment fault tree entity, its associated maintenance knowledge vector is loaded. This vector is extracted from historical work orders, fault diagnosis processes, and component replacement records, covering common repair operations, response time, and cost estimates. To further expand the entity semantic boundaries, a breadth-first graph traversal algorithm is used. Starting from the current entity node in the power knowledge graph, the feature attributes of adjacent first-level nodes (such as policy supporting measures and anomaly propagation chains) are recursively loaded to construct a multi-hop graph context. Ultimately, the link path constructed by the static feature vector and the electricity pricing policy entity, and the topological path constructed by the semantic feature vector and the equipment fault tree entity, are uniformly encoded as graph structure enhanced features, providing structured input for the subsequent attention mechanism.

[0029] S33, fused feature vector generation: The decoupled feature vector generated by S2 and the graph structure enhanced features generated by S32 are jointly input into the graph attention mechanism to cross-aggregate multi-source information. The graph attention mechanism realizes dynamic association calculation between different features by constructing multi-channel attention paths. Specifically, it constructs a policy decision weight path between the static feature vector and the electricity price policy entity features to evaluate the sensitivity of user attributes to the performance of specific electricity price policies; constructs a time-series association weight path between the dynamic feature vector and the load curve pattern entity in the knowledge graph to analyze the degree of match between the user's current electricity consumption behavior in the known typical curve; and constructs a fault propagation weight path between the semantic feature vector and the equipment fault tree entity to model the impact chain between potential multi-level faults. After completing the attention calculation of the above three types of paths, the original semantic backbone of the decoupled feature vector is retained through the gated residual connection mechanism, and the contextual semantics of the graph structure enhanced features are introduced to finally generate a more expressive fused feature vector, providing a semantically sufficient input foundation for the subsequent decision network.

[0030] S4: Input the fused feature vector into the dynamic weight gating network, adjust the feature weight according to the real-time electricity price policy, and output abnormal user identification labels, demand response strategies, and customer churn warning probability. Specifically: S41, Weighted Feature Tensor Generation: The fused feature vector output from S3 and the real-time electricity pricing policy are input into the feature weighting layer of the dynamic weight gating network. This layer first calls the policy response coefficient generator to perform correlation modeling based on the input electricity pricing policy content (including current policy type, effective time period, and step thresholds) with the static attributes and behavioral characteristics of the user profile. This generates three weight coefficients: a credit rating weight coefficient, a load sensitivity weight coefficient, and a policy response weight coefficient. Subsequently, the feature gating unit extracts the static feature vector, dynamic feature vector, and graph structure enhancement features from the fused feature vector, and performs inter-channel weighted multiplications with the corresponding weight coefficients. Specifically, the static feature vector is multiplied by the credit rating weight coefficient to highlight the influence of the user's credit rating on policy identification; the dynamic feature vector is multiplied by the load sensitivity weight coefficient to enhance responsiveness to changes in electricity usage behavior; and the graph structure enhancement features are multiplied by the policy response weight coefficient to reflect the influence of entity semantics in the knowledge graph on policy linkage. These three weighted feature vectors are concatenated along the channel dimension to form a unified weighted feature tensor, which serves as input to the subsequent decision layer.

[0031] S42, multi-task output generation: The weighted feature tensor is fed into the multi-task decision layer of the dynamic weight gating network. This layer consists of three physically isolated parallel processing channels, each responsible for a different task: The abnormal user identification branch is used to identify potentially high-risk users. This branch extracts static feature components from the weighted feature tensor and inputs them into a two-layer fully connected network. Leveraging the nonlinear mapping capability of the feature space, it conducts a comprehensive analysis of the user's electricity usage category, contract capacity, change frequency, and other behaviors. It ultimately outputs abnormal user identification labels, which can be classified as high-risk, medium-risk, or normal users. The demand response strategy branch focuses on strategy formulation and response optimization. This branch simultaneously extracts dynamic feature components and graph structure enhancement features from the weighted feature tensor. It first uses an integrated rule engine to perform preliminary strategy screening (such as peak-valley shifting, load reduction, and peak-shaving incentives). This is then input into the LSTM time series decision maker, which leverages its memory and prediction capabilities to generate a specific demand response strategy, including strategy type, response time window, and expected load adjustment amount. The customer churn warning branch monitors customer behavior trends. This branch extracts historical behavioral feature components from the weighted feature tensor, including customer behavior cycles, feedback frequency, and engagement rates. These components are then fed into a Cox proportional hazards model. Combining current and historical data features, the model estimates the medium- and long-term customer exit risk and outputs a customer churn warning probability, which is used to quantify the degree of user churn propensity.

[0032] S43, strategy optimization and output: The output results of the above three branches: abnormal user identification labels, demand response strategies and customer churn warning probabilities are synchronously input into the strategy optimization layer of the dynamic weighted gating network. This layer has a built-in strategy collaboration module to coordinate the output results of each task and improve the refinement and care of the response strategy. When the dual conditions of the abnormal user identification label being detected as high risk and the customer churn warning probability exceeding the preset threshold are met, a "customer care measure code" is automatically added to the generated demand response strategy. This code indicates that humanized intervention needs to be inserted into the subsequent service process, such as manual customer service return visits, targeted subsidy recommendations, and usage guidance push. Finally, the optimized abnormal user identification labels, demand response strategies and customer churn warning probabilities are output as the direct execution basis of the intelligent marketing control system.

[0033] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0034] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. The power marketing data analysis method based on AI big model is characterized by: The following steps are involved: S1: Collect structured and unstructured data from the power marketing system, unify timestamps and spatial encodings through the spatiotemporal alignment module, and generate unified encoding data using a pre-trained word embedding model; S2: Input the uniformly encoded data into a pre-trained large model for the power sector, separate the static feature vector, the dynamic feature vector, and the semantic feature vector through a feature decoupling layer, and output a decoupled feature vector; S3: Entity linking the decoupled feature vector with the power knowledge graph, where the knowledge graph includes the electricity price policy entity and the equipment fault tree entity, and generating a fused feature vector through a graph attention mechanism; S4: Input the fused feature vector into a dynamic weight gating network, adjust the feature weight according to the real-time electricity price policy, and output abnormal user identification labels, demand response strategies and customer churn warning probability.

2. The power marketing data analysis method based on AI big model according to claim 1 is characterized in that: The structured data includes user profiles and electricity consumption time series data, and the unstructured data includes customer service work order texts and equipment images.

3. The power marketing data analysis method based on AI big model according to claim 2 is characterized in that: The pre-trained word embedding model includes a four-channel encoder and a fusion output layer, specifically: The user profile encoder uses a two-layer fully connected network to convert the user profile numeric fields into user profile feature tensors; The time series data encoder is composed of a cascade of one-dimensional convolutional layers and LSTM layers, which converts the resampled power consumption time series data into a power consumption time series vector; The text encoder is based on the BERT architecture and introduces a substation encoding embedding layer. It uses the location-aware attention mechanism to generate a customer service ticket text vector based on the customer service ticket text with spatiotemporal anchor point labels. The image encoder uses the ResNet-18 backbone network, with a spatial coding fusion module added to its input layer to convert device images with spatial coding and timestamps into device image vectors. The fusion output layer uses residual connections to concatenate the user profile feature tensor, electricity consumption time series vector, customer service ticket text vector, and device image vector into unified encoded data.

4. The power marketing data analysis method based on AI big model according to claim 3 is characterized in that: Said S1 comprises: S11: Collect structured and unstructured data from the power marketing system. Structured data includes user profiles and electricity consumption time series data, and unstructured data includes customer service ticket texts and equipment images. Output the original multi-source data. S12: Performing a spatiotemporal alignment operation on the original multi-source data, specifically including resampling the power consumption time series data based on a 15-minute time stamp, adding a spatiotemporal anchor tag containing the spatial code of the work order location and the work order creation timestamp to the customer service work order text, and adding the spatial code of the shooting location and the shooting timestamp to the device image, and outputting the aligned multi-source data; S13: The aligned multi-source data is input into a pre-trained word embedding model for processing. The user profile is converted into a user profile feature tensor through a two-layer fully connected network. The resampled electricity consumption time series data is processed in cascade by a one-dimensional convolutional layer and an LSTM layer to generate an electricity consumption time series vector. The customer service work order text with spatiotemporal anchor point labels is generated into a customer service work order text vector through a location-aware attention mechanism containing a substation code embedding layer. The device image with spatial coding and timestamp is generated into a device image vector through a ResNet-18 network with an additional spatial coding fusion module. Finally, the user profile feature tensor, electricity consumption time series vector, customer service work order text vector and device image vector are spliced ​​into unified coding data in the fusion output layer.

5. The power marketing data analysis method based on AI big model according to claim 4 is characterized in that: The pre-trained large power domain model includes a cascaded three-layer structure: a multimodal fusion layer, a feature decoupling layer, and a semantic anchoring module, wherein; The multimodal fusion layer is composed of a cross-modal attention mechanism. It receives the user profile feature tensor, power consumption time series vector, customer service ticket text vector, and device image vector in the unified encoded data, and generates a multimodal feature tensor through feature cross-gating. The feature decoupling layer includes three physically isolated parallel branch processing channels. The static feature branch uses a two-layer fully connected network to process user profile-related inputs. The dynamic feature branch uses a stack of time-series convolutional networks with expansion rates of [1, 2, 4] to process power consumption time series vectors. The semantic feature branch uses a BiLSTM-CRF sequence labeling model to process text OCR recognition results in customer service ticket text vectors and device image vectors. The semantic anchoring module has a built-in semantic anchor word library in the power field, and performs keyword matching and standard fault code binding operations on the intermediate output of the semantic feature branch before outputting the decoupled feature vector.

6. The power marketing data analysis method based on AI big model according to claim 5 is characterized in that: The S2 includes: S21: Input the uniformly encoded data output by S1 into the multimodal fusion layer, and perform feature cross-gating calculations on the user profile feature tensor, power consumption time series vector, customer service ticket text vector, and device image vector through the cross-modal attention mechanism to generate a multimodal feature tensor; S22: Input the multimodal feature tensor into the three-way hardware isolation processing channel of the feature decoupling layer: the static feature branch extracts the user's credit rating and electricity consumption category through a two-layer fully connected network to form a static feature vector; the dynamic feature branch extracts the electricity consumption fluctuation pattern through a stack of temporal convolutional networks with an expansion rate of [1, 2, 4] to form a dynamic feature vector; and the semantic feature branch processes the customer service ticket text vector and the device image OCR recognition result through a BiLSTM-CRF sequence labeling model to form an original semantic vector. S23: Input the original semantic vector into the semantic anchoring module, match the standard fault code in the semantic anchor vocabulary in the power field for binding enhancement, output the enhanced semantic feature vector, and finally splice the static feature vector, dynamic feature vector and semantic feature vector into a decoupled feature vector according to the channel dimension.

7. The power marketing data analysis method based on AI big model according to claim 6 is characterized in that: The S3 includes: S31: Match the static feature vectors in the decoupled feature vectors with the electricity price policy entity in the power knowledge graph to the electricity consumption category. At the same time, topologically link the equipment fault keywords in the semantic feature vector to the equipment fault tree entity to generate entity link feature pairs. The electricity price policy entity includes the real-time floating electricity price rule, and the equipment fault tree entity includes the fault propagation path matrix. S32: Based on the entity link feature pair, extract the historical execution effect vector of the electricity price policy entity and the maintenance knowledge vector of the equipment fault tree entity from the power knowledge graph, load the feature attributes of adjacent entities through a breadth-first graph traversal algorithm, and output a graph structure enhancement feature that integrates historical policy data and maintenance experience; S33: The graph structure enhanced features and the original decoupled feature vector are input into the graph attention mechanism together, wherein the static feature vector and the electricity price policy entity feature calculate the strategy decision weight, the dynamic feature vector and the load curve pattern entity calculate the timing association weight, and the semantic feature vector and the equipment fault tree entity calculate the fault propagation weight, and a fusion feature vector is generated through the gated residual connection.

8. The power marketing data analysis method based on AI big model according to claim 7 is characterized in that: The dynamic weight gating network comprises a cascaded three-layer structure: a feature weighting layer, a multi-task decision layer, and a strategy optimization layer; wherein The feature weighting layer consists of a policy response coefficient generator and a feature gating unit. It receives real-time electricity price policy input and generates credit rating weight coefficients, load sensitivity weight coefficients, and policy response weight coefficients. It multiplies the static feature vectors, dynamic feature vectors, and graph structure enhancement features in the fusion feature vector with the corresponding weight coefficients and then concatenates them into a weighted feature tensor. The multi-task decision layer consists of three physically isolated parallel branch processing channels. The abnormal user identification branch uses a two-layer fully connected network to process the static feature components in the weighted feature tensor and output abnormal user identification labels. The demand response strategy branch integrates a rule engine and an LSTM time series decision maker to process dynamic feature components and graph structure enhancement features to generate demand response strategies. The customer churn warning branch uses the Cox proportional risk model to analyze the historical behavior feature components in the weighted feature tensor and output the customer churn warning probability. The strategy optimization layer has a built-in strategy collaboration module. When the abnormal user identification label is high risk and the customer churn warning probability exceeds the preset threshold, the customer care measure code is added to the demand response strategy.

9. The power marketing data analysis method based on AI big model according to claim 8 is characterized in that: The S4 includes: S41: Input the fused feature vector output by S3 and the real-time electricity price policy into the feature weighting layer of the dynamic weight gating network. The credit rating weight coefficient, load sensitivity weight coefficient, and policy response weight coefficient are calculated through the policy response coefficient generator. The static feature vector, dynamic feature vector, and graph structure enhancement feature are multiplied by the corresponding weight coefficients using the feature gating unit and then concatenated into a weighted feature tensor. S42: The weighted feature tensor is input into the three-way hardware isolation processing channel of the multi-task decision layer. The abnormal user identification branch processes the static feature components in the weighted feature tensor through a two-layer fully connected network to output an abnormal user identification label. The demand response strategy branch processes the dynamic feature components and graph structure enhancement features through a rule engine and an LSTM time series decision maker to generate a demand response strategy. The customer churn warning branch analyzes the historical behavior feature components through a Cox proportional risk model to output a customer churn warning probability. S43: The abnormal user identification label, demand response strategy and customer churn warning probability are input into the strategy optimization layer. When the abnormal user identification label is high risk and the customer churn warning probability exceeds the preset threshold, the customer care measure code is added to the demand response strategy through the strategy collaboration module, and finally the optimized abnormal user identification label, demand response strategy and customer churn warning probability are output.

Images