User energy efficiency optimization strategy generation method and system based on deep learning

Abstract

The invention discloses a user energy efficiency optimization strategy generation method and system based on deep learning, particularly relates to the field of comprehensive energy service, and is used for solving the problem of strategy instability caused by frequent jump of current user portraits. Constructing a group structure chart to identify leading customers and target customers to mark response features, generating a strategy candidate set by the group structure chart and a trigger queue, translating the strategy candidate set into an action script, setting a write-back suppression coefficient and a cooling period threshold, and calculating write-back control parameters to constrain write-back amplitude and rhythm; and outputting a resource release plan, collecting and executing a return certificate, and updating the dynamic portrait set and the group structure chart according to a write-back control parameter, thereby effectively maintaining the stability of the portrait, and realizing a stable closed loop of user energy efficiency optimization.

Description

Technical Field

[0001] This invention relates to the field of integrated energy services, and more specifically, to a method and system for generating user energy efficiency optimization strategies based on deep learning. Background Technology

[0003] The profiling results not only directly guide resource allocation, demand response organization, and integrated energy transformation commitments, but are also immediately written back after the implementation of allocation actions to update potential ranking and value levels. Event-triggered updates track the impact of short-term actions, while periodic updates reorder medium- and long-term focus targets. Consequently, profiling experiences frequent high-priority adjustments in a short period, constantly reversing resource allocation order, and forcing marketing commitments and service schedules to be rewritten accordingly. The positioning of dominant customer groups and target customers frequently shifts; potential high-value customers considered key in one round are considered saturated in the next, and then re-marked as key in subsequent rounds. Group structure observation, dynamic profiling, operational decisions, resource allocation, and profiling write-back are all within the same closed loop. It is not easy for deep learning strategies to provide stable user energy efficiency optimization strategies within this closed loop, especially in scenarios where integrated energy services need to be implemented through actual contractual commitments and on-site response arrangements. Without consistent references, the credibility and continuity of the strategy are difficult to maintain.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for generating user energy efficiency optimization strategies based on deep learning. This method aggregates multi-source data using a unified timescale to generate a dynamic profile set and a trigger queue. It constructs a group structure diagram to identify dominant and target customers and annotates response characteristics. The group structure diagram and trigger queue are used to generate a strategy candidate set and translate it into an action script. Write-back suppression coefficients and cooling-off period thresholds are set to calculate write-back control parameters that constrain the write-back amplitude and rhythm. Finally, a resource deployment plan is output, execution verification is collected, and the dynamic profile set and group structure diagram are updated according to the write-back control parameters, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based method for generating user energy efficiency optimization strategies includes the following steps: S1: Aggregate multi-source energy consumption and business events at a unified time scale, generate a dynamic profile set, and establish event-triggered and time-triggered trigger queues; S2: Construct a group structure diagram based on a dynamic profile set, identify dominant customers and target customers, and label their respective response characteristics; S3: Generate a set of policy candidates using the group structure diagram and trigger queue as input, and translate the selected policies into action scripts containing action order and time duration; S4: Set the strategy writeback suppression coefficient and cooldown threshold. After the action script is executed, the writeback control parameters are generated by machine learning based on the cooldown performance and profile changes, and the writeback amplitude and rhythm are constrained accordingly. S5: Output resource deployment plan and action script, collect execution feedback, and update dynamic profile set and group structure diagram after the cooldown period expires according to write-back control parameters.

[0007] In a preferred embodiment, step S1 involves extracting records of business action activation commitment compensation agreements from the marketing business platform, extracting energy consumption records from the electricity collection platform, and extracting demand response instruction records from the demand response platform. All records are mapped to a Coordinated Universal Time (UTC) timeline. Time gaps are processed to generate a record stream. The record stream is grouped by user identifier. User attribute sets are extracted in various dimensions to form a dynamic profile set. The record stream is scanned to identify event-type actions and generate event trigger markers. Time trigger markers are generated through periodic calculations. Trigger queues are arranged according to constraint strength. The dynamic profile set and trigger queues are linked to verify consistency and output a traceable structure.

[0008] In a preferred embodiment, step S2 uses the tagged attributes of the dynamic profile set as input, calculates the intra-group and inter-group associations according to the degree of similarity between users in terms of attribute similarity, generates a similarity matrix, filters the association set through threshold filtering, uses user entities as nodes, and uses similarity, collaboration, and influence relationships as edges to form a group structure graph.

[0009] In a preferred embodiment, the node ranking is calculated by reading the tagged attributes to select the dominant customer, the offset distance of each node relative to the cluster center of the dominant customer is analyzed and calculated to select the target customer, the various tags are fixed to the corresponding nodes to form a response feature description, and the identified dominant customer and target customer are located and fixed in the group structure diagram.

[0010] In a preferred embodiment, step S3 starts with the set of nodes of the dominant customer and the target customer in the group structure diagram, reads the currently valid trigger markers in the trigger queue, matches the trigger markers with the set of nodes to filter out the range of objects that need immediate response and the response window, and extracts the action unit by combining the response features corresponding to the nodes.

[0011] In a preferred embodiment, the action unit for extracting response features corresponding to the combined node includes extracting executable action units from the selected object range and the response features corresponding to the combined node to generate a preliminary strategy. After sorting and regularizing the action sequence, sequential dependencies, executable time period length and corresponding business constraints, a strategy candidate set is formed. The strategies that pass the test in the strategy candidate set are translated into action scripts and distributed to the execution side and the marketing side for use.

[0012] In a preferred embodiment, step S4 establishes a cooldown period for each action script, records the actual execution on-site to form an on-site fulfillment trajectory, records the changes in the positions of the dominant customer and target customer in the dynamic profile set and group structure diagram to form a profile displacement trajectory, records the committed amount and actual delivery amount for each action in the action script according to the action sequence and time period length, calculates the execution fulfillment ratio sequence, analyzes the fluctuation amplitude of all values ​​of the execution fulfillment ratio sequence during the cooldown period, and obtains the fulfillment stability measure by calculating the entropy value of the sequence; generates a priority offset record sequence by comparing the changes in the sorting order, priority label and service tilt mark of the dominant customer and target customer in the dynamic profile set before and after the execution of the action script, then merges them according to the coverage of the group structure diagram, uses a proportional aggregation algorithm to quantify the degree of change diffusion and outputs the write-back impact rate.

[0013] In a preferred embodiment, a write-back stability coefficient is generated using the fulfillment stability metric and write-back impact rate as inputs. The write-back stability coefficient is then linearly combined with the strategy write-back suppression coefficient and the cooldown threshold to obtain write-back control parameters. These parameters are then used to limit the magnitude and rhythm of the impact of write-back on the dynamic profile set and the population structure graph.

[0014] In a preferred embodiment, step S5 uses the action script as a basis to list the resource deployment plan, collect on-site execution feedback, and after the cooldown period expires, uses the write-back control parameters as the constraint boundary to update the dynamic profile set with the actual behavior confirmed in the current round of resource deployment plan and execution feedback. The updated dynamic profile set and the group structure diagram are then used as inputs in the next round of step S1.

[0015] A deep learning-based user energy efficiency optimization strategy generation system includes: Dynamic profile aggregation: Aggregate energy consumption and business events from multiple sources at a unified time point to generate a dynamic profile set, and establish trigger queues for event triggering and time triggering. Group structure construction: Based on the dynamic profile set, a group structure diagram is constructed to identify the dominant customers and target customers, and their respective response characteristics are labeled; Strategy script generation: Generate a set of policy candidates using the group structure diagram and trigger queue as input, and translate the selected policies into action scripts that include action sequence and time period length; Writeback parameter control: Set the strategy writeback suppression coefficient and cooldown threshold. After the action script is executed, the writeback control parameters are generated by machine learning based on the cooldown performance and profile changes, and the writeback amplitude and rhythm are constrained accordingly. Resource plan update: Output resource deployment plan and action script, collect execution feedback, and update dynamic profile set and group structure diagram after the cooldown period expires according to write-back control parameters.

[0016] The technical effects and advantages of the user energy efficiency optimization strategy generation method and system based on deep learning of this invention are as follows: This invention aggregates multi-source energy consumption data using a unified timescale to generate a dynamic profile set and trigger queue, providing a consistent foundation for building a group structure graph. This ensures that user behavior characteristics and event triggers are analyzed in a linked manner within the same framework, thereby identifying dominant and target customers and labeling response characteristics to achieve a stable mapping of group relationships. Based on this, the group structure graph and trigger queue drive the generation of a strategy candidate set, which is then translated into action scripts to accurately extract executable action sequences, preventing strategies from deviating from actual resource endowments. Furthermore, a strategy write-back suppression coefficient and a cooldown period threshold are introduced. Write-back control parameters are calculated based on machine learning to constrain the amplitude and rhythm of action script execution results, suppressing short-term fluctuations from interfering with the profile and forming a closed-loop feedback mechanism. Finally, by outputting a resource allocation plan, collecting execution feedback, and updating the dynamic profile set and group structure graph according to the write-back control parameters, real behavioral data is accumulated, driving iterative optimization. This technical logic integrates multi-source data and dynamic write-back control, not only solving the problem of frequent profile jumps leading to decision reversals but also improving the credibility and continuity of strategies in integrated energy services, ensuring the accuracy of resource allocation and the reliability of business commitments in multi-energy coupled scenarios. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the user energy efficiency optimization strategy generation method based on deep learning according to the present invention.

[0018] Figure 2 This is a schematic diagram of the user energy efficiency optimization strategy generation system based on deep learning according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: Figure 1This invention presents a deep learning-based method for generating user energy efficiency optimization strategies, comprising: S1: Aggregate energy consumption and business events from multiple sources at a unified time scale, generate a dynamic profile set, and establish trigger queues for event triggering and time triggering.

[0021] S2: Construct a group structure map based on dynamic profile sets, identify dominant customers and target customers, and label their respective response characteristics.

[0022] S3: Generate a policy candidate set using the group structure graph and trigger queue as input, and translate the selected policies into action scripts that include action sequence and time period length.

[0023] S4: Set the strategy write-back suppression coefficient and cooldown threshold. After the action script is executed, the write-back control parameters are generated by machine learning based on the cooldown performance and profile changes, and the write-back amplitude and rhythm are constrained accordingly.

[0024] S5: Output resource deployment plan and action script, collect execution feedback, and update dynamic profile set and group structure diagram after the cooldown period expires according to write-back control parameters.

[0025] In the field of integrated energy services, the generation of user energy efficiency optimization strategies needs to address the complexity of fragmented multi-source data and real-time dynamic adjustments. This requires the construction of a highly stable data integration framework, using deep learning as the core to identify user energy consumption structure and resource endowment, thereby accurately guiding resource allocation and operational decisions. The frequent jumps in existing profiling methods weaken the credibility of the strategies. Therefore, step S1 aggregates multi-source energy consumption and business events through a unified timescale to generate a dynamic profile set and trigger queue, which serve as the sole input basis for all subsequent analyses, ensuring data consistency and traceability, and avoiding short-term fluctuations from interfering with the continuity of the overall optimization process.

[0026] S1.1 Time Alignment Processing: To achieve unified integration of multi-source data, records of business actions, effective commitments, and compensation agreements are first extracted from the marketing business platform; energy consumption records are extracted from the electricity consumption data collection platform; and demand response instruction records are extracted from the demand response platform. These records all contain original timestamps. Based on a single timescale, a timestamp normalization algorithm is used to map all records onto a unified Coordinated Universal Time (UTC) timeline, including: Each record is traversed, its timestamp format (e.g., ISO 8601) is parsed, and converted to a unified timestamp value. For example, the local time is adjusted to Coordinated Universal Time (UTC) to eliminate time zone differences. Subsequently, a linear interpolation algorithm is used to handle time gaps. For data with missing time points, interpolation is calculated using a linear function of adjacent records. For example, for an energy consumption sequence, if a time point is missing, the value of an adjacent time point is used for estimation to ensure the continuity of the record stream. The algorithm's execution steps are: initializing the time axis grid, sorting records chronologically, identifying gaps, and applying interpolation.

[0027] Output a set of records that can be directly compared on the same timeline. Each record contains a unified timestamp, source identifier, and content description such as energy consumption status value or demand response instruction type to avoid deviations caused by asynchrony.

[0028] S1.2 Tag Attribute Generation: Based on the time-aligned record stream, role merging is performed at the granularity of individual users, locations, and time periods. First, the record stream is traversed, grouped by user identifiers, such as user IDs, to aggregate related records. Then, the K-means clustering algorithm is used to classify the grouped records, extracting attributes according to five dimensions: user behavior characteristics, demand characteristics, response tendencies, resource endowment, and business sensitivity. This includes initializing K cluster centers, where K is preset to be the number of dimensions (5), calculating the similarity between records and centers using Euclidean distance, and iteratively updating the centers until convergence. After clustering, the center value of each cluster is calculated and a label is assigned. For example, user behavior characteristic labels are quantified using cosine similarity, generating a set of labeled attributes such as the percentage of transferable load.

[0029] This collection comprehensively describes the user's performance under the coupling of multiple energy sources, including cold, heat, electricity, and water, ensuring that the dynamic image collection captures all historical and real-time states without omitting any data at any granular level.

[0030] S1.3 Trigger Marker: During time alignment, energy-saving actions, demand response actions, marketing commitment actions, and contractual constraint actions are labeled. First, the record stream is scanned, and a rule-based labeling algorithm is used to identify event-type actions. For example, a rule is defined: if a record contains the keyword "demand response instruction" and the timestamp corresponds to a sudden event, it is marked as an event trigger. Each record is matched against a preset rule set, including regular expression matching, such as demand response for instruction identification. Subsequently, for routine verification actions with a fixed rhythm, such as monthly resource assessments, time trigger tags are generated, determined through periodic calculations; for example, the first day of each month is marked as a time trigger. All trigger tags are sorted chronologically and arranged using a priority queue algorithm based on constraint strength, maintained using a heap structure. When inserting a tag, the strength value is calculated, such as a preset threshold based on the action's impact range. If the strength value is greater than 0.5, the priority is higher. The execution logic includes building a max-heap; the time complexity of insertion and extraction operations is O(logn), where n is the number of tags, forming a trigger queue. This trigger queue explicitly records the timing of reassessing the profile and strategy position, serving as input along with the dynamic profile set to avoid subjective interference.

[0031] S1.4 Queue and Image Integration: The generated dynamic profile set and trigger queue are integrated. First, a hash mapping algorithm is used to establish the association. For example, using user ID and timestamp as keys, trigger tags are linked to the corresponding tagged attribute set in the dynamic profile set. This includes creating a hash table, calculating the hash value of the key, and storing pointers to ensure an average access time of O(1). Subsequently, the integration consistency is verified by traversing and checking to ensure that the timestamps of all trigger tags are aligned with the dynamic profile set without any discrepancies. This process outputs a traceable structure, supporting subsequent group structure graph construction and strategy generation, ensuring that the dynamic profile set serves as the sole record of the user's true profile status, and the trigger queue serves as the strict basis for strategy recalculation.

[0032] Through the detailed processing in step S1, multi-source energy consumption and business events are transformed into a highly consistent dynamic profile set and trigger queue. This provides a solid data foundation for user group identification and optimization strategies in integrated energy services, effectively alleviates the instability of strategies caused by the current profile jumps, and ensures the credibility and long-term continuity of decisions in real contract commitments and on-site response scenarios.

[0033] The generation of user energy efficiency optimization strategies relies on accurate segmentation and feature identification of user groups. Deep learning is used as the core to process dynamic user profiles, thereby constructing a stable group relationship model and guiding targeted resource allocation. Current profiling methods suffer from unstable group positioning, leading to decision reversals. Therefore, step S2 constructs a group structure diagram based on the dynamic user profile set, identifies dominant and target customers, and labels response characteristics. This serves as direct input for generating the strategy candidate set, ensuring the reliability of group boundaries and priority ranking, and preventing short-term changes from interfering with operational continuity.

[0034] S2.1 Association Generation: To construct a user group model, the tagged attributes in the dynamic profile set are used as input. Based on the similarity between users in terms of energy consumption structure, demand response capability, and resource endowment, the intra-group and inter-group relationships are calculated.

[0035] The algorithm iterates through the set of tagged attributes and uses a cosine similarity algorithm to quantify the similarity between users. For example, it takes the energy structure vector as input to generate a similarity matrix. Then, it filters associations using a threshold; for example, similarities greater than 0.7 are considered intra-group associations, and those less than 0.3 are considered inter-group associations. The output set of association relationships provides the foundation for subsequent graph construction, ensuring that the connections between user entities are based on real user profile data rather than subjective assumptions.

[0036] S2.2 Group Structure Diagram Formation: Based on the set of relationships, each user entity is regarded as a node in the graph, and the similarity, collaboration and influence relationships between nodes are regarded as edges, forming a traceable group structure graph.

[0037] In constructing the group structure graph, firstly, all user entities are extracted from the dynamic profile set as a node list, with each node initialized using a user identifier as a unique ID. Then, based on the previously generated set of relationships, the similarity matrix is ​​traversed to add edges. Similarity relationships are determined by an energy structure similarity threshold, collaborative relationships are quantified by shared response history (e.g., the number of overlapping demand response events), and influence relationships are evaluated using a causal inference model (e.g., the propagation effect of one user's response behavior on another user). Edge addition uses an adjacency list storage structure to ensure graph scalability and efficient traversal. Next, the PageRank algorithm is applied to calculate the importance score of each node. This algorithm iteratively updates... The score of a new node is equal to the weighted sum of the scores of its incoming nodes divided by its out-degree, until convergence, to identify highly active, high-value, or highly responsive nodes and select the top 10% of the nodes by score as dominant customers. Simultaneously, cluster center shift analysis is performed. First, K-means clustering is performed on the dominant customer nodes to determine the centers. Then, the Euclidean distances of the remaining nodes relative to these centers are calculated. Nodes with moderate distances and positive attribute growth trends are selected as target customers with potential for further incubation, resource allocation value, or a high willingness to follow. Finally, the connectivity and consistency of the graph are verified to ensure that all nodes and edges are aligned with the dynamic profile set, forming a complete group structure graph that provides group boundaries, priority ranking, and influence diffusion paths.

[0038] When identifying highly active, high-value, or highly responsive nodes as dominant customers in a graph structure, relevant tagged attributes are first retrieved from a dynamic profile set. For example, activity is quantified using user behavior characteristic tags, such as the average response frequency and number of interactions; value is assessed using resource endowment tags, such as the total contribution of shiftable and reduceable loads; and responsiveness is measured using demand response compliance tags, such as the percentage of historical response success rate. Subsequently, the PageRank algorithm is applied to calculate the ranking of each node. This algorithm iteratively calculates the importance score of nodes based on edge weights, where the score is updated to the weighted sum of the scores of its incoming nodes divided by its out-degree. After convergence, the top 10% of nodes by score are selected as dominant customers, ensuring that the identification is based on global influence rather than local attributes within the graph structure.

[0039] When identifying nodes with potential for further incubation, resource allocation value, or a high willingness to follow as target customers in a group structure diagram, the first step is to analyze tagged attributes. For example, incubation potential is assessed using demand characteristic tags, such as the growth rate of energy consumption structure changes; resource allocation value is quantified using business sensitivity tags, such as the responsiveness to compensation agreements; and willingness to follow is measured using response tendency tags, such as the degree of subordination in collaborative relationships. Subsequently, cluster center offset analysis is employed. This analysis calculates the offset distance of each node relative to the dominant customer cluster center, uses the Euclidean distance formula to evaluate the offset magnitude, and selects nodes with moderate offset and positive attribute growth to ensure that target customer identification focuses on potential conversion value rather than the current state.

[0040] S2.3 Response Feature Annotation: Focusing on two sets of customers—primary and target customers—the system reads resource endowment tags, movable load tags, load reduction tags, energy storage window tags, and demand response compliance tags from a dynamic profile set. These tags are then fixed onto corresponding nodes to form response feature descriptions. This includes matching node IDs with tagged attribute sets and assigning tags to nodes one by one; for example, movable load tags are added as numerical attributes. This labeling uses a key-value pair structure for storage, ensuring that response features are implemented at the node level and have direct scheduling significance, such as for load adjustment commands. The process outputs a labeled group structure diagram, supporting subsequent action-level planning.

[0041] S2.4 Diagram Integration with Clients: The identified dominant and target customers are embedded into the group structure diagram, and overall consistency is verified. A traversal check ensures that all node labels are aligned with the dynamic profile set without deviation. This integrated output is a complete group structure diagram, directly used for prioritization and resource allocation, ensuring that dominant and target customers are clearly located in the diagram, serving as the boundary basis for strategy generation.

[0042] Through the processing in step S2, the dynamic profile set is transformed into a group structure graph with labeled response features. This provides a grouping and prioritization framework for generating strategy candidate sets in integrated energy services, effectively stabilizing customer positioning, avoiding current jump problems, and ensuring the credibility and execution continuity of optimization strategies under real business contracts.

[0043] Generating user energy efficiency optimization strategies requires extracting executable elements from the group structure and using deep learning to drive strategy formulation, ensuring efficient and stable responses to dominant and target customers. Current methods are susceptible to changes in user profiles; therefore, step S3 uses the group structure graph and trigger queue as input to generate a candidate set of strategies and translates the selected strategies into action scripts, providing direct guidance for resource deployment. This ensures the verifiability of action sequence and duration, preventing decision reversals from disrupting business continuity.

[0044] S3.1 triggers the matching and filtering of objects: To generate targeted strategies, the process begins by starting with the node set of dominant and target customers in the group structure graph. It then reads currently valid trigger markers from the trigger queue, including event-triggered and time-triggered markers. This involves traversing the node set and matching trigger markers by user identifier. For example, event-triggered markers are associated with node response characteristics to filter out the scope of objects requiring immediate response, such as specific user groups, and the response window, such as within 24 hours after the trigger time. This matching uses a hash mapping algorithm to establish key-value pairs between markers and nodes. The output of this process defines the filtered object scope and response window, providing boundaries for action refinement.

[0045] S3.2 Initial Strategy for Action Unit Extraction and Generation: Based on the scope of the screened objects, the response characteristics corresponding to the combined nodes are used to extract executable action units. These action units include load balancing actions, load reduction actions, energy storage charging and discharging actions, service tilting actions, and compensation fulfillment actions. Response feature tags are scanned, such as the moveable load tag, to quantify the movement amount. A rule-based combination algorithm is used to generate instruction sequences node by node. For example, load reduction actions are linked with energy storage charging and discharging actions to form a preliminary strategy list. Each preliminary strategy records the action type and preliminary parameters.

[0046] The specific execution logic of generating instruction sequences node by node using a rule-based combination algorithm includes: first, traversing and filtering each node within the target range, reading its response feature labels such as load shifting labels, load reduction labels, energy storage window labels, and demand response compliance labels, and mapping these labels to conditions in a preset rule set. For example, rule 1 defines that if the load shifting label is greater than the threshold of 20%, a load shifting action instruction will be generated first; rule 2 defines that if the energy storage window label indicates that the available capacity exceeds 50%, an energy storage charging and discharging action instruction will be generated next. Subsequently, condition matching is applied according to the rule priority order to form a dependency chain. For example, a load reduction action is used as a prerequisite to trigger a compensation fulfillment action, ensuring that the sequence is conflict-free and checking the dependency loop through verification. Each instruction sequence records the action type, such as the specific parameters of the load shifting action, including the shift amount and the start point of the time period, as well as preliminary parameters such as the quantified value calculated based on the label. Finally, all node sequences are aggregated to form a preliminary strategy list, supporting subsequent sorting and regularization.

[0047] S3.3 Formation of the Policy Candidate Set: The initial strategies are sorted and organized. First, they are arranged in order according to action dependencies using a topological sorting algorithm, for example, load balancing actions are performed before compensation actions. Then, parameters are organized, integrating time period lengths such as action duration in hours, and business constraints such as the agreed-upon compensation cap, forming a strategy candidate set. This includes verifying acyclic dependencies, calculating that the total time period length does not exceed the response window, and ensuring that each strategy includes action order, sequential dependencies, executable time period length, and corresponding business constraints. This process outputs a strategy candidate set, ensuring that all strategies have priority and comparability.

[0048] In a scenario of integrated energy services, a user node of the dominant customer A triggers an event trigger flag. The target scope is customer A and its associated target customer B, with a response window of 4 hours. The initial strategy list includes three action units: load balancing action (20kW balancing, 1-hour time period, no dependency), load reduction action (15kW reduction, 2-hour time period, dependent on load balancing action), and compensation payment action (compensation amount calculated based on reduction, 1-hour time period, dependent on load reduction action). First, a dependency graph is constructed: load balancing action points to load reduction action, and load reduction action points to compensation payment action; the dependency is verified to be acyclic by checking for no circular paths in the graph. Subsequently, a topology sorting algorithm is used to arrange the order, resulting in the action order as load balancing action, load reduction action, and compensation payment action. When normalizing parameters, the total time period is integrated to 4 hours to ensure it does not exceed the response window; sequential dependencies are added, such as the load reduction action must be started after the load balancing action, and business constraints are added, such as a compensation agreement upper limit of 100 yuan / kW. The final strategy that forms the candidate set consists of the following action sequence [load shifting action (1 hour), load reduction action (2 hours), compensation fulfillment action (1 hour)], with sequential dependencies [none, dependent on the previous one, dependent on the previous one], a total executable time period of 4 hours, corresponding to the business constraint [compensation cap of 100 yuan / kW], and is sorted in the candidate set according to priority (e.g., based on value tags) to ensure comparability with other strategies for subsequent testing.

[0049] S3.4 Action Script Translation and Output: The strategies that pass the test in the strategy candidate set are compiled into action scripts. First, strategies that meet business constraints are selected and translated one by one into a structured description, including the order of actions, time period, responsibility (e.g., the executing department), response boundaries (e.g., user location restrictions), and compensation criteria (e.g., the amount calculation formula). A template-filling algorithm is used to generate script files, which are then directly distributed to the execution and marketing teams. This translation ensures that the action scripts are executable and verifiable, and supports subsequent write-back control.

[0050] Through the processing in step S3, the group structure diagram and trigger queue are transformed into action scripts with action sequence and time period length. This provides an operable framework for resource deployment in integrated energy services, effectively stabilizes strategy generation, avoids current flip-flop issues, and ensures optimization continuity under contract commitments and on-site arrangements.

[0051] The generation of user energy efficiency optimization strategies needs to maintain profile stability through a write-back mechanism, using deep learning to process execution feedback and ensure the closed-loop consistency of resource allocation. Current methods are prone to frequent jumps due to write-back, therefore step S4 sets a strategy write-back suppression coefficient and a cooldown period threshold, calculates write-back control parameters for the action script deployment results, and constrains the write-back amplitude and rhythm as boundaries for updating the dynamic profile set, ensuring the credibility and continuity of the strategy in real business scenarios.

[0052] S4.1 Cooldown Period Establishment and Trajectory Recording: To control the impact of write-back, a corresponding cool-down period is first established for each action script that has been issued and executed. This cool-down period is a preset fixed duration, such as 7 days. During this period, the actual execution status on site is continuously recorded, including the promised and actual delivery volume of each action, as well as the changes in the positions of the leading and target customers in the dynamic profile set and group structure diagram.

[0053] The specific execution logic includes initializing the cooldown timer, grouping and recording data according to the action script identifier, collecting execution data in real time to form the on-site fulfillment trajectory, and simultaneously monitoring the profile displacement trajectory, such as priority label offset. This record is stored in a time-series database to ensure complete trajectory recording, support subsequent measurement calculations, and avoid missing any execution deviations.

[0054] S4.2 Realization of Stability Measurement Calculation: Based on the on-site fulfillment trajectory during the cooling-off period, a fulfillment stability metric is calculated. First, for each action in the action script, the committed amount and actual delivery amount are recorded one by one according to their action sequence and time period length, and the execution fulfillment ratio sequence is calculated. Each execution fulfillment ratio is a dimensionless ratio obtained by dividing the actual delivery amount by the committed amount, forming a time-ordered sequence. Subsequently, the fluctuation amplitude of all values ​​of this execution fulfillment ratio sequence during the cooling-off period is analyzed. A concentration analysis algorithm is used to quantify the degree of concentration in the time dimension by calculating the entropy value of the sequence. The specific process includes traversing the sequence elements, calculating the probability distribution of each element, and then applying the entropy formula to calculate the overall uncertainty. A lower entropy value indicates that the sequence is concentrated and consistent, thus outputting a single value as the fulfillment stability metric. This value ranges from 0 to 1, where close to 1 represents highly stable execution, and close to 0 represents significant fluctuations and inconsistencies in the execution process.

[0055] S4.3 Writeback Impact Rate Calculation: Based on the profile displacement trajectory during the cooling-off period, the write-back impact rate is calculated. First, the specific changes in the ranking order, priority tags, and service tilt markers of the dominant and target customers in the dynamic profile set before and after the execution of the action script are compared. For each customer node, its ranking position offset, priority tag value difference, and service tilt marker state switch are recorded to form a priority offset record sequence. Subsequently, this priority offset record sequence is merged according to the coverage of the group structure map. A proportional aggregation algorithm is used to calculate the proportion of the offset record in the total number of group nodes. The specific process includes grouping the node set, calculating the average offset of each group and dividing it by the total number of nodes, and then aggregating the proportions of all groups to obtain the overall diffusion degree. A single value is output as the write-back impact rate, which ranges from 0 to 1, where close to 0 indicates stable and localized changes, and close to 1 indicates drastic and widespread changes.

[0056] S4.4 Write-back stationarity coefficient generation and control parameter formation: Using the stability metric and writeback impact rate as input features, a gradient boosting regression tree algorithm is used to generate a writeback stability coefficient. This algorithm iteratively approximates the target round by round based on the residuals, and outputs continuous values ​​to characterize the stability of the current execution round. Subsequently, the writeback stability coefficient is calculated together with a preset strategy writeback suppression coefficient (e.g., 0.8) and a cooldown threshold (e.g., a threshold of 0.5), and a linear combination algorithm is used to form the writeback control parameters. For example, the control parameters are equal to the stability coefficient multiplied by the suppression coefficient and compared with the threshold.

[0057] The strategy write-back suppression coefficient refers to a preset dimensionless numerical parameter, typically ranging from 0 to 1, used to suppress the excessive influence of write-back operations on the dynamic profile set and group structure graph. This coefficient is applied as a multiplication factor when calculating write-back control parameters. For example, by multiplying by the write-back stability coefficient, the update amplitude is reduced, thereby preventing drastic fluctuations in profile ranking caused by short-term execution results and ensuring the stability of the optimization strategy. The determination of this coefficient is based on empirical analysis of historical execution data. First, the write-back impact rate distribution after multiple rounds of action script deployment is collected. Statistical methods such as median are used to calculate and set an initial value. For example, if the historical average impact rate exceeds 0.5, the coefficient is set to 0.8 to enhance suppression. Subsequently, iterative adjustments are made in simulated scenarios through cross-validation to ensure that the coefficient minimizes the risk of decision reversal while maintaining profile consistency. Finally, it is solidified as a system constant or dynamically fine-tuned according to the specific needs of business domains such as integrated energy services.

[0058] The cooling-off threshold refers to a preset threshold parameter, typically expressed as a dimensionless ratio or time unit. It defines whether the execution results during the cooling-off period are sufficiently stable to allow write-back operations. This threshold serves as a benchmark when formulating write-back control parameters. For example, if the write-back stability coefficient is below the threshold, write-back is delayed, thereby controlling the write-back rhythm and preventing frequent updates from interfering with the group structure. The determination of this threshold is based on empirical evaluation of the cooling-off period trajectory. First, the fluctuation patterns of the on-site fulfillment trajectory and the profile displacement trajectory are analyzed. Threshold optimization algorithms, such as ROC curve analysis, are used to determine the set value. For example, a threshold of 0.5 is set for samples with a fulfillment stability metric below 0.6 to filter unstable execution. Subsequently, sensitivity tests are conducted on real business datasets, and the threshold is adjusted to balance update timeliness and stability. Finally, it is embedded into the system as a configuration parameter, supporting adjustments to higher values ​​based on the frequency of energy demand response events, such as during peak periods, to enhance control.

[0059] The generated write-back control parameters limit the impact magnitude and rhythm of write-back on the dynamic profile set and group structure map through quantization thresholds and temporal constraints. First, the write-back control parameters are a composite index, including an amplitude factor and a rhythm factor. The amplitude factor is calculated based on the product of the write-back stability coefficient and the strategy write-back suppression coefficient, and is used to limit the adjustment magnitude of each write-back on the sorting order, priority label, and service skew marker. For example, if the amplitude factor is less than 0.5, the write-back is only allowed to shift no more than 10% of the original position, thereby preventing a short-term large increase or decrease in the position of the dominant customer and target customer. Second, the rhythm factor is formed through a cooling-off period threshold comparison calculation and is used to control the write-back. The execution frequency is controlled by setting a threshold. For example, if the rhythm factor exceeds the cooldown threshold, the write-back is delayed until the start of the next cycle. This ensures that the write-back operation does not exceed a specified number of times within a preset time interval, such as once a day, to avoid frequent jumps that could disrupt the stability of the population structure graph. In practical applications, the write-back control parameters serve as the sole criterion for judgment. After the cooldown period expires, the decision on whether to perform a write-back, the magnitude of the adjustment, and the interval between rhythm changes are evaluated. For example, a threshold filtering algorithm is used to traverse priority offset records, applying only changes that match the magnitude factor, and scheduling the update queue according to the rhythm factor. This maintains the consistency of the full description of the dynamic profile set and the reliability of node positioning in the population structure graph, ensuring the continuity of the overall user energy efficiency optimization strategy within the closed loop.

[0060] In the construction of the gradient boosting regression tree model, a weak learner is first initialized as a base regression tree. This tree constructs decision nodes based on the input features, namely the realization stability metric and the write-back impact rate. Each node splits the dataset using the information gain criterion, and the tree depth is controlled between 3 and 5 layers to avoid overfitting. Subsequently, an ensemble method is used to add new regression trees round by round. Each round of new trees fits the residuals of the previous round's model to form an additive model, where the overall prediction output is a weighted sum of the predictions of all trees, ensuring that the model gradually approaches the true write-back stability coefficient target.

[0061] In the optimization process of the gradient boosting regression tree model, mean squared error is used as the loss function, and the residual is updated through the negative gradient direction. In each iteration, the residual between the current model prediction and the actual target is calculated as a pseudo-label to train a new tree. The learning rate parameter is introduced to scale the contribution of the new tree to control the step size. The iteration continues until the loss function converges or reaches the preset maximum number of iterations. The optimization objective is to minimize the sum of squared residuals on the training set. At the same time, an early stopping mechanism is applied to monitor the performance of the validation set to prevent overfitting.

[0062] In setting the parameters of the gradient boosting regression tree model, the number of trees is set to 100 to 200 rounds to balance computational efficiency and accuracy, the maximum tree depth is set to 3 to 5 to capture nonlinear relationships without introducing noise, the learning rate is set to 0.01 to 0.1 for incremental optimization to avoid oscillations, the minimum number of split samples is set to 2 to ensure leaf node stability, and the subsampling rate is set to 0.8 to introduce randomness and improve generalization ability. These parameters are tuned on historical execution data through grid search or cross-validation to ensure that the mean squared error of the model on the cooling-off period trajectory dataset is below the threshold of 0.05.

[0063] Through the processing in step S4, the action script deployment results are transformed into write-back control parameters, which provides amplitude and rhythm constraints for profile updates in integrated energy services, effectively suppressing current jump problems and ensuring the stability and long-term credibility of the strategy closed loop under contractual commitments and response arrangements.

[0064] The generation of user energy efficiency optimization strategies requires continuous iteration through a closed-loop feedback mechanism, using deep learning to process execution results and ensure stable updates of user profiles and graph structures. Current methods lack constraints during write-back, which can lead to inconsistent decisions. Therefore, step S5 outputs resource deployment plans and action scripts, collects execution feedback, and updates the dynamic profile set and group structure graph after the cooldown period according to write-back control parameters. This serves as the basis for closing and restarting the entire optimization process, ensuring the accumulation of genuine evidence of resource deployment and supporting the credibility and long-term sustainability of the strategy in real contract commitments and on-site response scenarios.

[0065] S5.1 Resource Deployment Plan Generation: To ensure the strategy's implementation, a systematic resource deployment plan is first developed based on the action script. This plan specifies the service team deployment arrangements, such as assigning specific operations teams to target customer locations; the integrated energy technology adaptation arrangements, such as matching energy storage equipment with user resource endowments; the demand response organization arrangements, such as scheduling load shifting timetables; and the compensation payment arrangements, such as the formula for calculating compensation amounts and payment cycles. Specifically, this includes: The process iterates through the sequence of actions in the action script, including load shifting, load reduction, energy storage charging / discharging, service tilting, and compensation fulfillment. These actions are grouped and categorized by responsible party. The execution side is responsible for on-site operational actions such as load adjustment, while the marketing side is responsible for fulfilling commitments such as compensation calculation. Subsequently, for each group, a systematic list of corresponding arrangements is created, including service team deployment arrangements (assigning specific operations teams to key or target customer locations), integrated energy technology adaptation arrangements (matching compatible solutions for equipment such as energy storage batteries with shiftable loads based on user resource endowment tags), demand response organization arrangements (developing scheduling timelines, such as coordinating the specific start and end times of load shifting within the response window), and compensation fulfillment arrangements (determining...). The compensation amount is calculated using a formula based on a fixed percentage of the actual reduction amount multiplied by the agreed unit price, with a specified payment cycle such as monthly settlement. To ensure consistency, the parameters of the action script are strictly compared during the grouping process. For example, if the action script specifies load balancing as the first priority for 2 hours, the resource deployment plan will arrange service teams to perform the balancing operation during that period, while recording business constraints such as the compensation cap not exceeding the agreed percentage threshold. This generation process uses a template mapping algorithm to directly map the structural elements of the action script to a predefined plan template. For example, key-value pair matching converts the action sequence into a plan sequence, ensuring unbiased output and forming a readable document. Finally, the generated resource deployment plan is output in document form and directly distributed to the operations and marketing departments.

[0066] S5.2 Execute receipt collection: Based on the execution of the resource deployment plan, the system collects execution receipts, which include on-site execution records such as actual load adjustments and timestamps, response completion records such as the percentage of completion of demand response instructions, compensation fulfillment records such as actual payment amounts and receipts, and compliance records with business constraints such as whether agreed boundaries have been exceeded. Specifically, this involves real-time monitoring of the execution process, comparing each item against the action script using sensors or log systems. For example, it compares the actual delivery volume with the promised volume for each action sequence, creating a comparison table. Subsequently, all records are aggregated, and data verification algorithms are used to check completeness. Missing items are filled through supplementary queries, ensuring that the execution receipts cover all elements of the action script. This collection process employs a distributed collection framework to ensure that the receipts serve as unique and reliable credentials, supporting subsequent accounting and auditing without relying on manual reports, thus preventing data tampering or omissions.

[0067] S5.3 Updates Dynamic Portrait Set and Group Structure Diagram: After the cooling-off period, using the write-back control parameters as constraints, the resource allocation plan for this round and the actual behaviors confirmed in the execution feedback are embedded into a dynamic profile set. First, user identifiers are matched, and the actual behaviors in the execution feedback, such as actual response compliance, are updated to the tagged attribute set. For example, if the execution feedback shows that the actual load reduction completion rate is higher than promised, the demand response compliance tag is adjusted upwards accordingly. Specifically, this includes applying the amplitude factor of the write-back control parameters to limit the update amplitude, for example, allowing only a 10% offset in tag values, and scheduling update timing according to a rhythm factor. Subsequently, the positions of leading and target customers are adjusted on the group structure graph. By recalculating node rankings and offset distances, for example, using the PageRank algorithm to iteratively update the leading customer score, the actual behaviors in the execution feedback, such as actual response completion rate and compliance indicators, are integrated into the response feature tags. The value of the updated priority tag is calculated using a weighted average algorithm; for example, the new priority tag equals the original tag value multiplied by the amplitude factor of the write-back control parameters plus the quantified contribution of the actual behavior, ensuring that the adjustment reflects actual execution without introducing drastic offsets. This update ensures that the dynamic image set captures the full real state, and the group structure graph reflects the latest group boundaries without introducing drastic swings.

[0068] S5.4 Output Integration and Loop Restart: The updated dynamic profile set and group structure diagram are then used as input for the next step, S1. First, update consistency is verified by traversing and checking to ensure all tags and nodes are aligned with the write-back control parameters, with no discrepancies. Then, the refresh rhythm is recorded, such as update counts limited by a rhythm factor. This integration process ensures the comprehensive accumulation of results from resource deployment plans and action scripts. Subsequent strategy generation relies on verified execution feedback, supporting iterative optimization of the entire closed loop and avoiding reliance on subjective interpretations.

[0069] Through the processing in step S5, the resource deployment plan and action script are transformed into traceable execution receipts, and are subject to write-back control parameters to update the dynamic profile set and group structure diagram. This provides a closed-loop feedback framework for optimization strategies in integrated energy services, effectively maintaining profile stability and ensuring decision continuity and business credibility in scenarios of multi-energy coupling and demand response.

[0070] Example 2: Figure 2 The present invention provides a user energy efficiency optimization strategy generation system based on deep learning, comprising: Dynamic profile aggregation: Aggregate energy consumption and business events from multiple sources at a unified time point to generate a dynamic profile set, and establish trigger queues for event triggering and time triggering.

[0071] Group structure construction: Based on the dynamic profile set, a group structure map is constructed to identify the dominant customers and target customers, and their respective response characteristics are labeled.

[0072] Strategy script generation: Generate a set of policy candidates using the group structure graph and trigger queue as input, and translate the selected policies into action scripts that include the action sequence and time period length.

[0073] Writeback parameter control: Set the strategy writeback suppression coefficient and cooldown threshold. After the action script is executed, the writeback control parameters are generated by machine learning based on the cooldown performance and profile changes, and the writeback amplitude and rhythm are constrained accordingly.

[0074] Resource plan update: Output resource deployment plan and action script, collect execution feedback, and update dynamic profile set and group structure diagram after the cooldown period expires according to write-back control parameters.

[0075] It should be noted that the system of the present invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting a variety of hardware environments and usage requirements.

[0076] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0077] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely to distinguish one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0078] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating user energy efficiency optimization strategies based on deep learning, characterized in that, Including the following steps: S1: Aggregate multi-source energy consumption and business events at a unified time scale, generate a dynamic profile set, and establish event-triggered and time-triggered trigger queues; S2: Construct a group structure diagram based on the dynamic profile set, identify the dominant customers and target customers, and label their respective response characteristics; S3: Generate a set of policy candidates using the group structure diagram and trigger queue as input, and translate the selected policies into action scripts containing action order and time duration; S4: Set the strategy writeback suppression coefficient and cooldown threshold. After the action script is executed, the writeback control parameters are generated by machine learning based on the cooldown performance and profile changes, and the writeback amplitude and rhythm are constrained accordingly. S5: Output resource deployment plan and action script, collect execution feedback, and update dynamic profile set and group structure diagram after the cooldown period expires according to write-back control parameters.

2. The user energy efficiency optimization strategy generation method based on deep learning according to claim 1, characterized in that: In step S1, records of business action activation commitments and compensation agreements are extracted from the marketing business platform, energy consumption records are extracted from the electricity consumption data collection platform, and demand response instruction records are extracted from the demand response platform. All records are mapped to the Coordinated Universal Time (UTC) timeline, time gaps are processed to generate a record stream, the record stream is grouped by user identifier, user attribute sets of various dimensions are extracted to form a dynamic profile set, the record stream is scanned, event-type actions are identified to generate event trigger markers, time trigger markers are generated through periodic calculations, and trigger queues are arranged according to constraint strength. The dynamic profile set and the trigger queue are linked to verify consistency and output a traceable structure.

3. The user energy efficiency optimization strategy generation method based on deep learning according to claim 1, characterized in that: In step S2, the tagged attributes of the dynamic profile set are used as input. The intra-group and inter-group associations are calculated according to the degree of similarity between users in terms of attribute similarity. A similarity matrix is ​​generated. The association set is filtered by threshold filtering. User entities are used as nodes, and similarity, collaboration, and influence relationships are used as edges to form a group structure graph.

4. The user energy efficiency optimization strategy generation method based on deep learning according to claim 3, characterized in that: Read the tagged attributes to calculate the node ranking and select the dominant customer. Analyze and calculate the offset distance of each node relative to the cluster center of the dominant customer to select the target customer. Solidify each tag to the corresponding node to form a response feature description. Solidify the identified dominant customer and target customer location into the group structure diagram.

5. The user energy efficiency optimization strategy generation method based on deep learning according to claim 1, characterized in that: Step S3 starts with the set of nodes of the dominant customer and the target customer in the group structure diagram, reads the currently valid trigger markers in the trigger queue, matches the trigger markers with the set of nodes to filter out the scope of objects that need immediate response and the response window, and extracts the action unit by combining the response features corresponding to the nodes.

6. The user energy efficiency optimization strategy generation method based on deep learning according to claim 5, characterized in that: The action unit for extracting response features corresponding to the combined node includes extracting executable action units from the selected object range and the response features corresponding to the combined node to generate a preliminary strategy. After sorting and regularizing the action sequence, sequential dependencies, executable time period length and corresponding business constraints, a strategy candidate set is formed. The strategies that pass the test in the strategy candidate set are translated into action scripts and distributed to the execution side and the marketing side for use.

7. The user energy efficiency optimization strategy generation method based on deep learning according to claim 1, characterized in that: Step S4 establishes a cooling-off period for each action script, records the actual on-site execution to form an on-site fulfillment trajectory, records the changes in the positions of the dominant customer and target customer in the dynamic profile set and group structure diagram to form a profile displacement trajectory, records the committed amount and actual delivery amount for each action in the action script according to the action sequence and time period length, calculates the execution fulfillment ratio sequence, analyzes the fluctuation amplitude of all values ​​of the execution fulfillment ratio sequence during the cooling-off period, and obtains the fulfillment stability measure by calculating the entropy value of the sequence; By comparing the changes in the ranking order, priority tags, and service tilt markers of dominant and target customers in the dynamic profile set before and after the execution of the action script, a priority offset record sequence is generated. Then, it is merged according to the coverage of the group structure map, and the degree of change diffusion is quantified using a proportional aggregation algorithm to output the write-back impact rate.

8. The user energy efficiency optimization strategy generation method based on deep learning according to claim 7, characterized in that: Using the realization stability metric and writeback impact rate as inputs, a writeback stability coefficient is generated. The writeback stability coefficient is then linearly combined with the strategy writeback suppression coefficient and the cooldown threshold to obtain writeback control parameters. These parameters are then used to limit the impact magnitude and rhythm of writeback in the dynamic profile set and the population structure diagram.

9. The user energy efficiency optimization strategy generation method based on deep learning according to claim 1, characterized in that: Step S5 uses the action script as a basis to list the resource deployment plan, collect on-site execution feedback, and after the cooldown period expires, uses the write-back control parameters as the constraint boundary to update the dynamic profile set with the actual behavior confirmed in the current round of resource deployment plan and execution feedback. The updated dynamic profile set and group structure diagram are then used as inputs in the next round of step S1.

10. A deep learning-based user energy efficiency optimization strategy generation system, used to implement the deep learning-based user energy efficiency optimization strategy generation method according to any one of claims 1-9, characterized in that, include: Dynamic profile aggregation: Aggregate energy consumption and business events from multiple sources at a unified time point to generate a dynamic profile set, and establish trigger queues for event triggering and time triggering. Group structure construction: Based on the dynamic profile set, a group structure diagram is constructed to identify the dominant customers and target customers, and their respective response characteristics are labeled; Strategy script generation: Generate a set of policy candidates using the group structure diagram and trigger queue as input, and translate the selected policies into action scripts that include action sequence and time period length; Writeback parameter control: Set the strategy writeback suppression coefficient and cooldown threshold. After the action script is executed, the writeback control parameters are generated by machine learning based on the cooldown performance and profile changes, and the writeback amplitude and rhythm are constrained accordingly. Resource plan update: Output resource deployment plan and action script, collect execution feedback, and update dynamic profile set and group structure diagram after the cooldown period expires according to write-back control parameters.

Images