Electric power marketing intelligent decision engine based on multi-modal artificial intelligence
By integrating image, text, time series, and structured data through a multimodal artificial intelligence decision engine, the dynamic and personalized strategy generation and self-optimization of the power marketing system have been realized. This solves the problem of insufficient utilization of unstructured information in the existing system and improves the accuracy and adaptability of the strategy.
Patent Information
- Application Number
- CN202511645880.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-11
- Publication Date
- 2026-02-10
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing power marketing system lacks comprehensive utilization of unstructured information such as images, text, and time series. The strategy generation method is static and rigid, making it difficult to adapt to complex and ever-changing individual needs and market environment. The feedback mechanism is lagging and lacks dynamic optimization capabilities.
The power marketing intelligent decision engine adopts multimodal artificial intelligence. Through data access module, semantic modeling module, task scheduling module, dynamic fusion module and strategy generation module, it integrates image, text, time series and structured behavioral data, and uses Transformer and reinforcement learning to achieve accurate initial prediction and personalized fine-tuning of strategy, and build a closed-loop optimization path.
It enhances the breadth and depth of the system's semantic perception of user status and external environment, possesses individual adaptability and self-learning capabilities, generates more accurate and efficient electricity marketing strategies, and adapts to complex user groups and rapidly changing market conditions.
Smart Images
Figure CN121504207A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electricity marketing technology, and more particularly to an intelligent decision engine for electricity marketing based on multimodal artificial intelligence. Background Technology
[0002] Against the backdrop of the "dual-carbon" strategy and the construction of a new power system, electricity marketing is undergoing a transformation from experience-driven to data-intelligent driven. In current technological solutions, some power companies have already attempted to introduce artificial intelligence technology, mainly applied to power dispatch optimization, equipment inspection, and customer service, achieving some success. In the field of electricity marketing, existing systems generally adopt analysis methods based on structured data (such as electricity consumption, user profiles, and geographic information), combined with rule engines or traditional machine learning models for basic user segmentation, strategy recommendation, and price adjustment.
[0003] Furthermore, some studies have attempted to apply reinforcement learning algorithms to optimize electricity pricing strategies, achieving dynamic adjustment and feedback learning of strategies in simulated environments. However, most of these methods are still limited to single data modalities or specific scenarios, lacking systematic integrated design. Current mainstream electricity marketing systems are centered on structured data, lacking comprehensive utilization of unstructured information such as images, text, and time series data. In actual operation, they mostly use static strategy templates, lacking the ability to link with changes in user behavior in real time.
[0004] Existing solutions typically rely on structured user data, using traditional machine learning methods such as clustering and classification to segment users, and then combining this with rule engines to set marketing strategy templates. After strategy execution, the effectiveness is evaluated using static statistical indicators. This type of solution suffers from several technical shortcomings: First, it has limited data utilization dimensions, supporting only structured information and unable to handle unstructured data such as images, text, and time series, leading to information blind spots in the characterization of user behavior and states. Second, the strategy generation method is static and rigid, lacking intelligent modeling capabilities and struggling to adapt to complex and ever-changing individual needs and market environments. Third, the feedback mechanism is lagging and crude, unable to dynamically track and optimize the strategy's effectiveness in real time, resulting in a lack of closed-loop evolution capabilities for the overall strategy.
[0005] Therefore, there is an urgent need for an intelligent decision-making engine for electricity marketing based on multimodal artificial intelligence. Summary of the Invention
[0006] To address the aforementioned issues, this invention provides an intelligent decision-making engine for electricity marketing based on multimodal artificial intelligence.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0008] A smart decision-making engine for electricity marketing based on multimodal artificial intelligence, comprising:
[0009] The data access module is used to access existing data sources within the power company.
[0010] The semantic modeling module is used to set up dedicated deep encoders based on data sources to perform semantic modeling and generate feature vectors for each modality.
[0011] The task scheduling module is used to dynamically adjust the participation weights of each modality feature vector in the fusion process according to the task type.
[0012] The dynamic fusion module is used to fuse feature vectors from different modalities through a multi-layer fusion network with Cross-Attention as the core, based on the participating weights, to generate a unified semantic expression vector.
[0013] The policy generation module receives a unified semantic representation vector and outputs policy parameters.
[0014] The strategy execution and feedback module is used to push or recommend strategies based on existing marketing system platforms, and automatically collect user response behavior and power-side operation indicators to form a strategy execution feedback dataset.
[0015] Preferably, the data source includes image data, text data, time-series data, and structured behavioral data.
[0016] Preferably, the image class includes photos of electricity meters and equipment nameplates, the text class includes user repair records and customer service dialogues, the time series class includes electricity load and electricity price curves, and the structured behavioral data includes user profiles and past response records.
[0017] Preferably, the image data uses Swing Transformer to extract visual semantics, supplemented by OCR to recognize meter readings, device numbers, and scanned documents; the text data uses a customized power industry big data model to extract user intent, emotional attitudes, and service keywords; the time series data uses i Transformer to model user load changes and electricity price fluctuation trends; and the structured behavioral data uses DeepFM to model user attributes and behavioral interactions.
[0018] Preferably, the task types include electricity price push, strategy recommendation, and response prediction.
[0019] Preferably, the workflow of the strategy generation module includes:
[0020] Based on a unified semantic representation vector, an initial policy is output through a Transformer network;
[0021] After generating the initial policy, the system initiates policy fine-tuning based on reinforcement learning according to the user's actual response.
[0022] During the fine-tuning process, the PPO algorithm is adopted to automatically analyze the click-through rate, dwell time, and conversion effect of the pushed content, and combine the grid load and price fluctuations to adjust the key parameters in the initial strategy and generate optimized strategy parameters.
[0023] Preferably, the initial strategy includes push time period, electricity price range, and recommendation type.
[0024] Preferably, the system also includes a strategy optimization module, which receives a strategy execution feedback dataset and fine-tunes the optimized strategy parameters online using the FTRL algorithm.
[0025] Preferably, the system also includes a structural path reconstruction module, which is used to periodically backtest and evaluate the effectiveness of recent strategy execution and summarize the correlation data between user behavior, strategy response, and business metrics.
[0026] Preferably, after periodically reviewing the effectiveness of recent strategy implementation and summarizing the correlation data between user behavior, strategy response, and business metrics, the method further includes:
[0027] When the system detects that the click-through rate of a certain type of user is continuously declining or the conversion rate is lower than the set threshold, it will trigger the retraining of the strategy model based on the PPO algorithm. By relearning the state behavior patterns, the system will make overall corrections and optimizations to the strategy logic.
[0028] The beneficial effects of this invention are as follows:
[0029] This application integrates four modalities of data: image, text, time series, and structured behavior. It employs OCR+SwinTransformer to extract image information, a large model to model text semantics, iTransformer to process time series changes, and DeepFM to represent user behavior profiles. Furthermore, it dynamically integrates multi-source features through a task-aware modal scheduling mechanism and a Cross-Attention cross-modal fusion structure. Compared to traditional solutions that rely solely on structured data or static fusion methods, this approach enhances the breadth and depth of the system's semantic perception of user status, device status, and external environment, supporting more comprehensive and context-aware intelligent decision-making.
[0030] This system employs a two-stage decision-making mechanism combining Transformer and reinforcement learning. It utilizes the Transformer model to achieve accurate initial strategy prediction, and then uses the PPO algorithm to fine-tune key strategy variables (such as price adjustment magnitude and push timing) to maintain overall strategy stability while possessing high individual adaptability. This mechanism enables the system to continuously generate more accurate, efficient, and differentiated electricity marketing strategies in the face of complex user groups and rapidly changing market conditions.
[0031] By constructing a dual-path policy update mechanism based on feedback signals, the system can continuously perform fine-tuning and retraining operations during operation: on the one hand, FTRL is introduced to achieve rapid parameter adjustment in high-frequency scenarios; on the other hand, the experience replay mechanism and PPO are used to complete the structural update of the policy network, enabling the policy to have proactive evolution capabilities. This closed-loop optimization path overcomes the problems of static and easily aging existing models, enabling the system to maintain a "stable, robust, and scalable" state in actual operation for a long time. Attached Figure Description
[0032] Figure 1 This is a block diagram of the intelligent decision-making engine for electricity marketing based on multimodal artificial intelligence in a specific embodiment of the present invention. Detailed Implementation
[0033] Please see Figure 1 As shown, this invention relates to an intelligent decision-making engine for power marketing based on multimodal artificial intelligence. This invention provides an intelligent decision-making engine for power marketing based on multimodal artificial intelligence, which addresses the current practical problems of power companies such as low user identification accuracy, slow response to strategy push, and weak feedback utilization capabilities. It designs functional modules such as perception, fusion, decision-making, and feedback optimization to form a deployable and iterative intelligent marketing system.
[0034] The system takes multi-source heterogeneous data, including images, text, time series data, and structured behavioral data, as input. It combines deep semantic modeling, a task-oriented fusion mechanism, a two-stage policy generation method (combining supervised learning and reinforcement learning), and an adaptive optimization process based on user behavior feedback to form a complete business closed loop of "data perception—policy generation—feedback update." This engine can be embedded into an enterprise marketing platform, possessing capabilities such as online policy output, personalized user push notifications, and rapid self-learning adjustments. It is suitable for various marketing tasks, including electricity price adjustments, service recommendations, and load guidance.
[0035] Compared with existing technical solutions that mainly rely on structured data and static rules, this invention achieves joint modeling of unstructured information, dynamic and personalized adjustment of strategies, and optimization path driven by high-frequency feedback. This improves the system's responsiveness and execution accuracy in complex market environments and has good engineering feasibility and promotional value.
[0036] Specifically, an intelligent decision-making engine for electricity marketing based on multimodal artificial intelligence includes:
[0037] 1. Data access module, used to access existing data sources within the power company, specifically including:
[0038] The system first connects to four existing data sources within the power company: image data (such as meter photos and equipment nameplates), text data (such as user repair records and customer service conversations), time-series data (such as electricity load and electricity price curves), and structured behavioral data (such as user profiles and past response records). Data access can be achieved online or in batches through standard interfaces, ensuring compatibility with the company's existing marketing platform or electricity information system.
[0039] Specifically, data source identification and classification:
[0040] The system first scans and registers existing data sources within the power company, and then categorizes them according to their data type and interface protocol.
[0041] Data types mainly include:
[0042] Structured data: Tabular data from relational databases (such as Oracle, MySQL), such as case studies, electricity bills, and contract information.
[0043] Time-series data: Electricity consumption data, load data, voltage and current data from time-series databases (such as InfluxDB) or IoT platforms.
[0044] Unstructured text data: Customer service tickets, inspection reports, and policy documents from file systems or NoSQL databases.
[0045] Spatial data: Geographic information on substation locations and transmission line corridors from the GIS system.
[0046] Adapter mode access:
[0047] Configure a dedicated data adapter for each type of data source.
[0048] For database-related applications: The adapter encapsulates the corresponding JDBC / ODBC driver or dedicated connection library, which is responsible for establishing connections, executing SQL queries or calling stored procedures, and reading result sets.
[0049] For API interface class sources: the adapter is responsible for handling authentication (such as OAuth, API Key), constructing HTTP requests, and parsing the returned JSON / XML data.
[0050] For file-based sources: the adapter is responsible for monitoring the specified directory and parsing files in formats such as CSV, Excel, and PDF.
[0051] For streaming data sources: the adapter is based on message middleware such as Kafka and MQTT to perform real-time data subscription and consumption.
[0052] Data standardization and caching:
[0053] Each adapter converts the extracted raw data into a unified internal data model. For example, it standardizes all timestamps to ISO 8601 format and all numerical units to international standard units.
[0054] After initial cleaning and standardization, the data will be published to a unified data bus or written to an intermediate data storage, so that downstream deep encoders can consume it on demand, thereby decoupling data access from data processing.
[0055] 2. Semantic modeling module, used to set up dedicated deep encoders based on data sources for semantic modeling, generating feature vectors for each modality, specifically including:
[0056] For different data types, the system employs dedicated deep encoders for semantic modeling: Image modalities utilize the Swin Transformer to extract visual semantics, supplemented by OCR recognition of key information such as meter readings, equipment numbers, and scanned documents; text modalities call upon a customized large-scale power industry model to extract user intent, emotional attitudes, and service keywords; time-series modalities use the i Transformer to model user load changes and electricity price fluctuation trends; and structured behavioral data is modeled using DeepFM to depict user attributes and behavioral interactions. All features are ultimately unified and converted into high-dimensional semantic vectors, providing input support for the downstream strategy generation module.
[0057] Specifically, the modeling method is as follows:
[0058] Modal-specific encoder design:
[0059] For structured / time series data:
[0060] One-dimensional convolutional neural networks or long short-term memory networks are used as the core encoder.
[0061] 1D-CNN can capture local power consumption patterns (such as morning and evening rush hours).
[0062] LSTM excels at learning long-term dependencies, making it ideal for analyzing seasonal and cyclical trends in electricity load.
[0063] The encoder takes normalized time-series data points as input and outputs a fixed-length feature vector that represents the deep patterns of the entire time series.
[0064] For unstructured text data:
[0065] Pre-trained language models are used as encoders, such as BERT, RoBERTa, or models fine-tuned specifically for power domain corpora.
[0066] Processing flow: Segment customer service work orders, reports, etc. into words -> input into BERT model -> extract the output vector corresponding to the [CLS] tag or perform mean pooling on all word vectors as the semantic feature vector of the entire text.
[0067] For spatial data:
[0068] A graph neural network or a two-dimensional convolutional neural network is used as the encoder.
[0069] GNN: Models the power grid structure (substations, lines, user nodes) as a graph. GNN learns the embedding representation of nodes and the entire graph through message passing, which can effectively capture the topological relationship of the power grid.
[0070] 2D-CNN: If map data is converted into raster images, 2D-CNN can be used to extract spatial features.
[0071] Training and fine-tuning:
[0072] These deep encoders can be pre-trained on large-scale general datasets (such as ImageNet for CNN, Wikipedia for BERT) and then fine-tuned end-to-end on downstream tasks specific to electricity marketing.
[0073] For example, using "user churn prediction" or "high-value user identification" as a supervision signal, backpropagating the error, and adjusting the parameters of all encoders makes the extracted feature vectors most discriminative for the task.
[0074] For example, in a scenario where a user uploads an image of the electricity meter, the system can use OCR to identify the current electricity reading and device number, and then correlate and analyze it with its historical load curve and the last repair record to generate a complete user state vector, which serves as the context input for strategy generation.
[0075] 3. Task scheduling module, used to dynamically adjust the participation weights of each modality feature vector in the fusion process according to the task type, specifically including:
[0076] The feature vectors encoded by each modality are input to the task-aware scheduling module. This module dynamically adjusts the participation weight of each modality feature in the fusion process according to the current marketing task type (such as electricity price push, strategy recommendation, and response prediction), which is one of the key mechanisms for achieving task-differentiated processing in this invention. The system determines the importance distribution of various types of information under a specific task through an attention matching mechanism between the task vector and each modality vector. For example, in the device recognition task, image and structured modalities will receive higher weights; in the task of processing user feedback and satisfaction prediction, text and temporal modalities are prioritized.
[0077] Specifically, including:
[0078] Attention-based weight adjustment:
[0079] Input: Feature vectors from each deep encoder;
[0080] Calculate the attention score: Input each feature vector into a small feedforward neural network and calculate a scalar score;
[0081] e i =u T tanh(Wv i +b);
[0082] Where Wv is the feature vector of the i-th mode. i b and u are learnable parameters.
[0083] Normalized weights: The attention scores of all modalities are normalized to a probability distribution in which the sum of the weights is 1 using the Softmax function.
[0084]
[0085] Weighted fusion: The normalized weights are multiplied by the original feature vectors and then summed to obtain the final fusion vector.
[0086] Weight adjustment based on gating mechanism:
[0087] Inspired by the LSTM / GRU gating mechanism, a gating vector can be designed to control the flow of information between different modes.
[0088] 4. Dynamic fusion module, used to fuse feature vectors from different modalities through a multi-layer fusion network centered on Cross-Attention, based on the participating weights, to generate a unified semantic representation vector. Specifically, it includes:
[0089] The system constructs a multi-layered fusion network with Cross-Attention as its core to achieve intermodal context completion and joint modeling, such as text-guided image semantic interpretation and structured behavioral information enhancement for understanding load curves. The unified high-dimensional semantic representation vector generated after fusion serves as input to the policy generation module, driving more precise push timing, electricity pricing strategies, or recommended content selection.
[0090] The policy generation module receives a unified semantic representation vector and outputs policy parameters, specifically including:
[0091] The strategy generation module consists of two phases: the initial strategy is generated by a Transformer network, which uses historical marketing sample data to build the input-output mapping relationship during training. The output includes basic strategy parameters such as push time period, electricity price range, and recommendation type. This module can be started even without initial samples and is suitable for large-scale user initial strategy generation scenarios.
[0092] After the initial strategy generation, the system initiates a reinforcement learning-based strategy fine-tuning process based on actual user responses. This stage employs the Proximal Policy Optimization (PPO) algorithm. The system automatically analyzes the click-through rate, dwell time, and conversion effects of electricity price adjustments in the pushed content. Combined with grid load and price fluctuations, it adjusts key parameters in the strategy, such as optimizing push timing, adjusting price fluctuation ranges, and refining recommended content. This fine-tuning process supports edge node operation, enabling rapid response to changes in user behavior and achieving strategy outputs that better meet individual needs. This represents a core shift in the system from rule enforcement to self-evolution.
[0093] For example, when dealing with a group of users who have clicked on push notifications multiple times but never converted, the system will gradually reduce the weight of their electricity price discounts through reinforcement learning, while shortening the policy push window to observe changes in their behavior, thereby achieving personalized adjustments.
[0094] 5. Strategy Execution and Feedback Module: This module is used to push or recommend strategies based on parameters through the existing marketing system platform, and automatically collect user response behavior and power-side operation indicators to form a strategy execution feedback dataset, specifically including:
[0095] The optimized strategy is pushed or recommended through the existing marketing system platform. The system automatically collects user response behavior (such as click behavior, page dwell time, complaint records, evaluation content, etc.) and power-side operation indicators (such as regional load level, grid stability, electricity price change response, etc.) to form a strategy execution feedback dataset.
[0096] Feedback data is standardized and modeled by the data access component to form multi-dimensional structured feedback indicators, which support the invocation of the strategy optimization module. The system integrates an abnormal comment filtering mechanism to identify and filter low-quality data such as extreme sentiment, incomplete characters, and duplicate comments. This mechanism, based on a combination of a rule engine and a sentiment shift model, ensures the retention of authentic and representative user behavior. The system supports sampling frequencies from minutes to hours and can be configured for real-time access or scheduled retrieval, ensuring the continuity and stability of strategy optimization.
[0097] 6. Strategy optimization module: The strategy optimization module is used to receive the strategy execution feedback dataset and fine-tune the optimized strategy parameters online using the FTRL algorithm.
[0098] To enhance the stability and self-evolution capability of the strategy model in long-term operation, a dual-path strategy optimization mechanism of "rapid fine-tuning + structural reconstruction" was designed.
[0099] The rapid fine-tuning path is suitable for scenarios with frequent fluctuations in user behavior or rapid changes in the environment. The system uses the FTRL (Follow-The-Regularized-Leader) algorithm to fine-tune key strategy parameters such as push time and electricity price adjustment range online without retraining the model. The adjustment results can take effect immediately. For example, when users frequently leave or give negative reviews in a short period of time, the system can fine-tune parameters such as push rhythm and electricity price range in real time to avoid the failure of marketing strategies. It does not require triggering full model training and ensures rapid response.
[0100] 7. The structural path reconstruction module is used to periodically review and evaluate the effectiveness of recent strategy execution, and summarize the correlation data between user behavior, strategy response, and business metrics. Specifically, it includes:
[0101] The structural reconfiguration path is applicable to situations where the policy network experiences long-term performance degradation or new policy tasks emerge. The system periodically backtests the effectiveness of recent policy execution, summarizing the correlation data between user behavior, policy response, and business metrics. When it is found that the click-through rate of a certain type of user continues to decline or the conversion rate falls below a set threshold, the system will trigger the retraining of the policy model based on the PPO algorithm. By relearning state behavior patterns, the system will make overall corrections and optimizations to the policy logic, thereby improving the accuracy of subsequent push notifications and user satisfaction.
[0102] The system scheduling module automatically determines which path to use based on factors such as strategy effectiveness metrics, model confidence, or feedback volatility. The quick fine-tuning path can be embedded in daily operations for rolling optimization, while the structural refactoring path is triggered primarily during off-peak hours to ensure model updates do not disrupt normal business operations. For example, when a strategy's click-through rate and satisfaction rate decline continuously, the system will automatically trigger the fine-tuning and refactoring process to achieve adaptive optimization of the strategy.
[0103] This invention proposes a multimodal artificial intelligence-driven intelligent decision-making engine for electricity marketing. Through deep fusion and intelligent modeling of multimodal information such as images, text, time-series data, and structured behavioral data, it constructs an integrated intelligent marketing system with strong semantic awareness, strategy self-optimization, and closed-loop learning capabilities. Compared to existing technologies, this system possesses the following significant technical advantages:
[0104] 1. High depth of multimodal data fusion and strong semantic expression capability.
[0105] The system integrates four modalities of data: image, text, temporal data, and structured behavioral data. It employs OCR+SwinTransformer to extract image information, a large-scale model to model text semantics, iTransformer to handle temporal variations, and DeepFM to represent user behavior profiles. Furthermore, it dynamically integrates multi-source features through a task-aware modal scheduling mechanism and a Cross-Attention cross-modal fusion structure. Compared to traditional solutions based solely on structured data or static fusion methods, this approach enhances the breadth and depth of the system's semantic awareness of user status, device status, and the external environment, supporting more comprehensive and context-aware intelligent decision-making.
[0106] 2. The intelligence and dynamic adaptability of strategy generation are significantly enhanced.
[0107] This system employs a two-stage decision-making mechanism combining Transformer and reinforcement learning. It utilizes the Transformer model for initial, accurate strategy prediction, followed by personalized fine-tuning of key strategy variables (such as price adjustment magnitude and push timing) using the Proximal Policy Optimization (PPO) algorithm. This approach maintains overall strategy stability while possessing a high degree of individual adaptability. This mechanism enables the system to continuously generate more precise, efficient, and differentiated electricity marketing strategies in the face of complex user groups and rapidly changing market conditions.
[0108] 3. The strategy system possesses self-learning and self-evolution capabilities, enhancing the stability of long-term strategy performance.
[0109] By constructing a dual-path policy update mechanism based on feedback signals, the system can continuously perform "fine-tuning + retraining" operations during operation: on the one hand, FTRL is introduced to achieve rapid parameter adjustment in high-frequency scenarios; on the other hand, the experience replay mechanism and PPO are used to complete the structural update of the policy network, enabling the policy to have proactive evolution capabilities. This closed-loop optimization path overcomes the problems of static and easily aging existing models, enabling the system to maintain a "stable, robust, and scalable" state in actual operation for a long time.
[0110] 4. Possesses excellent engineering deployability and commercialization capabilities, saving deployment and maintenance costs.
[0111] Each functional module of this system supports independent deployment and distributed computing. The core model adopts the mainstream deep learning framework (PyTorch), and the backbone models such as OCR and the Qingyuan large model have reusable pre-trained versions, significantly reducing the threshold for modeling and training. The system architecture is compatible with the existing marketing system platform of power companies, and can be connected and launched without large-scale architectural adjustments. It has the advantages of low-cost deployment and high scalability, and is suitable for large-scale customized strategy deployment scenarios with "one thousand people, one policy".
[0112] Example 2
[0113] Based on the above embodiment 1, this application further includes:
[0114] I. Data Acquisition and Fusion
[0115] Multimodal data fusion layer: This layer is responsible for collecting and preprocessing data from heterogeneous data sources.
[0116] Internal structured data: Through data interfaces, user profiles, historical electricity consumption, electricity bills, payment records, business expansion application process data, etc. are obtained from the power marketing business system.
[0117] External unstructured text data: Through web crawling modules, targeted text information related to power policies, regional economic development, and industry trends is extracted from social media, news websites, and government reports.
[0118] Visual and spatiotemporal data: Access weather data (temperature, humidity, solar radiation intensity) from meteorological departments, as well as information such as urban construction and industrial zone heat maps extracted from satellite or drone remote sensing images.
[0119] AI Intelligent Analysis Center: This is the "brain" of the system, responsible for in-depth mining of the fused multimodal data.
[0120] The user profiling module employs deep learning models (such as the Transformer architecture) to model users' electricity consumption behavior sequences and identify their electricity consumption patterns (such as "peak users," "valley users," and "stable users"). Simultaneously, it utilizes natural language processing technology to analyze users' service tickets and customer service call records, extracting sentiment tendencies and key demands to form a dynamic and refined 360° user profile.
[0121] Market Situation Awareness Module: Combining external text data (such as "expansion of an industrial park" or "introduction of new energy vehicle subsidy policy"), this module uses topic modeling and named entity recognition technology to automatically identify market opportunities and risks. For example, it identifies the discussion heat of emerging industries such as "photovoltaics" and "energy storage" and predicts potential incremental risks.
[0122] Multimodal correlation analysis module: Establishes cross-modal correlation models. For example, by combining meteorological data (temperature) with historical electricity consumption data, a time-series forecasting model (such as LSTM or TCN) can be trained to accurately predict the regional load curve in the short term (e.g., the next 24 hours) and long term (e.g., the next quarter). Another example is linking remote sensing images (newly built buildings) with internal application data to predict future electricity application growth trends in the region.
[0123] II. Decision Generation
[0124] Based on the above analysis, the decision generation and optimization engine begins to operate. The following example, "Precise Recommendation of Energy Efficiency Packages for High-Value Industrial Users," illustrates the decision-making process:
[0125] Scenario Trigger: The AI Intelligent Analysis Center identified a large manufacturing user, A, whose user profile showed that it was "energy-intensive" and "sensitive to electricity prices." Furthermore, recent meteorological data predicted that the region would experience sustained high temperatures, and the load forecasting model indicated that the power grid was about to enter its peak period.
[0126] Strategy generation: The decision engine is triggered, and the following analysis results are invoked:
[0127] User profiles reveal the potential for adjustable user load (e.g., the ability to temporarily shut down certain non-core production lines).
[0128] Market observations revealed the existence of a "demand response" subsidy policy for industrial users.
[0129] By combining real-time electricity prices and load forecasts, the potential benefits of user participation in demand response and its impact on production can be calculated.
[0130] Multi-objective optimization: The engine performs multi-objective optimization calculations with the objectives of "maximizing user satisfaction," "maximizing grid peak shaving and valley filling benefits," and "maximizing power company revenue." It doesn't simply recommend a standard package, but dynamically generates a personalized "energy efficiency package combination," for example:
[0131] Auxiliary strategy: Match it with a high-efficiency frequency converter equipment preferential plan (through correlation analysis, identify its old equipment as the main energy consumption point).
[0132] Incentive strategy: Clearly inform participants of the direct electricity cost credits available for participating in this demand response, as well as the long-term energy-saving benefits of equipment upgrades.
[0133] Implementation and Feedback:
[0134] Strategy Execution: The generated decision-making solutions are accurately delivered through channels such as APP push notifications and account manager workbenches.
[0135] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A smart decision-making engine for electricity marketing based on multimodal artificial intelligence, characterized in that, include: The data access module is used to access existing data sources within the power company. The semantic modeling module is used to set up dedicated deep encoders based on data sources to perform semantic modeling and generate feature vectors for each modality. The task scheduling module is used to dynamically adjust the participation weights of each modality feature vector in the fusion process according to the task type. The dynamic fusion module is used to fuse feature vectors from different modalities through a multi-layer fusion network with Cross-Attention as the core, based on the participating weights, to generate a unified semantic expression vector. The policy generation module receives a unified semantic representation vector and outputs policy parameters. The strategy execution and feedback module is used to push or recommend strategies based on strategy parameters through the existing marketing system platform, and automatically collect user response behavior and power-side operation indicators to form a strategy execution feedback dataset.
2. The intelligent decision-making engine for electricity marketing based on multimodal artificial intelligence according to claim 1, characterized in that, The data sources include image data, text data, time-series data, and structured behavioral data.
3. The intelligent decision-making engine for electricity marketing based on multimodal artificial intelligence according to claim 2, characterized in that, The image class includes photos of electricity meters and equipment nameplates; the text class includes user repair records and customer service dialogues; the time series class includes electricity load and electricity price curves; and the structured behavioral data includes user profiles and past response records.
4. The intelligent decision-making engine for electricity marketing based on multimodal artificial intelligence according to claim 3, characterized in that, The image data uses Swing Transformer to extract visual semantics, supplemented by OCR to recognize meter readings, device numbers, and scanned documents. The text data calls a customized large-scale power industry model to extract user intent, emotional attitudes, and service keywords. The time series data uses iTransformer to model user load changes and electricity price fluctuation trends. The structured behavioral data uses DeepFM to model user attributes and behavioral interactions.
5. The intelligent decision-making engine for electricity marketing based on multimodal artificial intelligence according to claim 1, characterized in that, The task types include electricity price push, strategy recommendation, and response prediction.
6. The intelligent decision-making engine for electricity marketing based on multimodal artificial intelligence according to claim 1, characterized in that, The workflow of the strategy generation module includes: Based on a unified semantic representation vector, an initial policy is output through a Transformer network; After generating the initial policy, the system initiates policy fine-tuning based on reinforcement learning according to the user's actual response. During the fine-tuning process, the PPO algorithm is adopted to automatically analyze the click-through rate, dwell time, and conversion effect of the pushed content, and combine the grid load and price fluctuations to adjust the key parameters in the initial strategy and generate optimized strategy parameters.
7. The intelligent decision-making engine for electricity marketing based on multimodal artificial intelligence according to claim 6, characterized in that, The initial strategy includes the push period, electricity price range, and recommendation type.
8. The intelligent decision-making engine for electricity marketing based on multimodal artificial intelligence according to claim 1, characterized in that, It also includes a policy optimization module, which receives policy execution feedback datasets and fine-tunes the optimized policy parameters online using the FTRL algorithm.
9. The intelligent decision-making engine for electricity marketing based on multimodal artificial intelligence according to claim 1, characterized in that, It also includes a structure path reconstruction module, which is used to periodically backtest and evaluate the effectiveness of recent strategy execution and summarize the correlation data between user behavior, strategy response and business metrics.
10. The intelligent decision-making engine for electricity marketing based on multimodal artificial intelligence according to claim 9, characterized in that, The method for periodically retrospectively evaluating the effectiveness of recent strategy execution, and summarizing the correlation data between user behavior, strategy response, and business metrics, also includes: When the system detects that the click-through rate of a certain type of user is continuously declining or the conversion rate is lower than the set threshold, it will trigger the retraining of the strategy model based on the PPO algorithm. By relearning the state behavior patterns, the system will make overall corrections and optimizations to the strategy logic.