Power market package configuration particle swarm optimization method considering user satisfaction

Abstract

The invention provides a power market package configuration particle swarm optimization method considering user satisfaction, and relates to the field of power market retail operation, and the method comprises the steps: collecting user power consumption behaviors, interaction records and social media texts, employing a BERT-LSTM model to recognize emotion intensity in real time, dynamically adjusting the satisfaction model weight, and constructing an emotion-satisfaction mapping mechanism; an anti-fact explanation technology is introduced, a'minimum disturbance 'package adjustment direction is generated for a negative emotion user, an explanation gradient vector is formed through aggregation, a particle swarm optimization algorithm is driven to evolve in the direction where the satisfaction degree is most easily improved in speed updating, and rapid convergence and satisfaction degree peak value improvement are achieved; and finally, a high-satisfaction and explainable power package scheme is output, and personalized explanation suggestions are pushed to the user, so that the user stickiness is remarkably improved, the complaint rate is reduced, and the market competitiveness of an electricity selling company is enhanced.

Description

Technical Field

[0001] This invention relates to the field of electricity market retail operations, and more specifically to a particle swarm optimization method for electricity market package configuration that takes into account user satisfaction. Background Technology

[0002] As the retail electricity market gradually opens up, electricity sales companies need to design diverse packages for a massive number of users. Most existing technologies aim to minimize electricity purchase costs or maximize sales revenue, using fixed weights to convert user load characteristics into objective scores, which are then solved using linear programming or traditional particle swarm optimization algorithms. However, this "cost-oriented" model ignores users' subjective experiences, leading to low package acceptance, high complaint rates, and significant user churn.

[0003] In recent years, some studies have incorporated user satisfaction into their optimization objectives. However, satisfaction models are generally static linear weighted, meaning that once the weights are set, they remain unchanged for a long period, failing to reflect real-time changes in user sentiment due to fluctuations in electricity prices, service incidents, or social media opinions. Furthermore, existing particle swarm optimization relies solely on random cognitive and social terms during the search process, resulting in aimless particle flight, slow convergence, and a lack of interpretability regarding "how to fine-tune the package to satisfy users," making it difficult to respond to inquiries from regulators and users.

[0004] In summary, the industry urgently needs a method for configuring electricity packages that can sense user emotions in real time, dynamically adjust satisfaction weights, and introduce explainable guidance during the optimization process, in order to improve user satisfaction, reduce complaint rates, and enhance the market competitiveness of electricity sales companies. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a particle swarm optimization method for electricity market package configuration that takes into account user satisfaction. This method solves the problems of lack of emotion perception, blind search, and lack of explanation in existing electricity retail package optimization. It dynamically adjusts the satisfaction weight to improve user satisfaction, convergence speed, and solution interpretability.

[0006] The technical solution adopted in this invention is to provide a particle swarm optimization method for electricity market package configuration that considers user satisfaction. By collecting user electricity consumption behavior, interaction records, and social media text, the method uses a BERT-LSTM model to identify the intensity of emotions in real time, dynamically adjusts the weights of the satisfaction model, constructs an emotion-satisfaction mapping mechanism, introduces counterfactual explanation technology, generates the "minimum perturbation" package adjustment direction for users with negative emotions, aggregates them to form an explanatory gradient vector, and drives the particle swarm optimization algorithm to evolve along the direction that is easiest to improve satisfaction during speed updates, achieving rapid convergence and improvement in peak satisfaction, and finally outputting a highly satisfactory and explainable electricity package plan, and pushing personalized explanation suggestions to users.

[0007] In a preferred embodiment, the present invention provides an emotion perception and dynamic satisfaction model. By collecting multi-source data, the system analyzes the user's emotional state using a BERT-LSTM model to calculate the emotion intensity and constructs a linkage mapping relationship between emotion and satisfaction weights. When a user's negative emotion is detected to be high, the system automatically adjusts the satisfaction calculation weight, increases the weight of electricity price sensitivity, and decreases the weight of green electricity preference. This allows the package optimization target to reflect the user's real feelings in real time and significantly improves the market acceptance of the solution.

[0008] In a preferred embodiment, the present invention also provides a counterfactual interpretation optimization algorithm, which constructs a counterfactual analysis result for users with negative emotions that "the minimum package adjustment can significantly improve satisfaction", aggregates the counterfactual adjustment directions of multiple users to form an optimization guidance direction representing "the easiest way to improve user experience", and introduces this guidance direction as an independent driving factor into the particle swarm search process, driving the particle swarm optimization algorithm to evolve along the direction that is easiest to improve satisfaction in the speed update, thereby achieving rapid convergence and an increase in peak satisfaction. Attached Figure Description

[0009] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a diagram of the operational architecture of the method of this invention; Figure 2 This is a flowchart of the operation of the method of the present invention. Detailed Implementation

[0010] To better understand the purpose, system architecture, and functional implementation of this embodiment, the embodiments and features described herein can be combined with each other without conflict. The exemplary embodiments disclosed herein will be described below with reference to the accompanying drawings, including specific technical details disclosed to aid understanding; however, these details should be considered exemplary rather than restrictive. Therefore, those skilled in the art should understand that various improvements and adjustments can be made to the embodiments described herein without departing from the scope and core ideas of the invention. Similarly, for clarity, detailed descriptions of well-known technologies, functions, and structures are omitted in the following description.

[0011] Example 1 Figure 1 This is a diagram of the operational architecture of the method of this invention.

[0012] like Figure 1 As shown, a particle swarm optimization method for configuring electricity market packages considering user satisfaction includes the following steps: S110. Data acquisition and preprocessing: Establish standardized data access pipelines and adopt multi-level fusion solutions to address the differences in time granularity of different data sources. S120. Emotion perception and weight adjustment: Establish a historical emotion distribution database, conduct relative evaluation based on dynamic benchmarks, and adjust satisfaction weights based on an interpretable rule engine. S130. Optimize model construction and execution, perform cluster analysis on users, identify user groups with similar characteristics, build a personalized satisfaction model, and set multiple convergence conditions to optimize the particle swarm algorithm. S140. Explanation generation and user interaction: For users with negative emotions, highly customized explanation content is generated, and user interaction is carried out using multi-channel reach and retry mechanism; S150 System monitoring and anomaly handling: Establish a comprehensive system health monitoring dashboard, automatically trigger alarm processes when anomalies occur, and design a hierarchical anomaly response mechanism to maintain normal system operation.

[0013] According to an embodiment of the present invention, in step S110, smart meter data is collected in real time from the distribution automation system every 15 minutes via the IEC 61850 protocol, including parameters such as load curves and voltage fluctuations. Customer service call recordings are obtained through an enterprise-level call center API, automatically transcribed into text, and timestamped. APP interaction logs collect user behavior data through tracking technology, including package browsing time, price comparison frequency, and complaint button clicks. Social media data is obtained through a dedicated web crawler tool for the power industry, focusing on local forums and Weibo topics. The collected data undergoes preliminary cleaning at the edge gateway to remove duplicate records, fill in missing values, standardize the time format, and establish a data quality scoring mechanism to ensure the reliability of the data input to the model.

[0014] To address the differences in time granularity across different data sources, a multi-level fusion scheme is employed. Load data at the 15-minute level is used as the basic time unit. Customer service calls are aggregated according to morning and evening peak hours, APP behavior is statistically analyzed daily, and social media activity is statistically analyzed hourly. A time alignment algorithm maps the data from each source onto a unified timeline to form three-dimensional data. For sentiment recognition, a feature engineering system is constructed, using abnormal load fluctuations as negative sentiment trigger signals, such as load fluctuations exceeding a threshold for three consecutive periods. Viewing time on price comparison pages in the APP exceeding 90 seconds is marked as price-sensitive, and posts on social media containing keywords such as "electricity bills are too high" or "electricity prices are unreasonable" are included in the sentiment intensity calculation. A comprehensive sentiment index is generated by weighting and summing these features to avoid bias from a single indicator.

[0015] According to an embodiment of the present invention, in step S120, a historical emotion distribution database is established, and baseline intervals are divided according to user type (residential / commercial), season (summer / winter), and electricity consumption period (weekday / weekend). A sliding window mechanism is also introduced to calculate the moving average of emotion intensity over the past 7 days, avoiding misjudgments caused by a single extreme event. When an emotion intensity is detected to be above 0.7 for three consecutive cycles, a high-priority optimization process is triggered; otherwise, the process enters the regular optimization channel.

[0016] The satisfaction weight adjustment follows an interpretable rule engine. When the emotion intensity is >0.7, the system automatically activates the following rule: The weighting of electricity price sensitivity has been increased from the baseline value of 1.0 to 1.5, reflecting users' extreme concern about costs; The weighting of green electricity preference has decreased from a baseline of 1.2 to 0.8, indicating that environmental concerns have given way to economic considerations. The weight of service convenience has been increased from the baseline value of 1.0 to 1.3, emphasizing the importance of simplifying the operation process; The stability guarantee weight has been increased from the baseline value of 1.0 to 1.4, highlighting users' urgent need for power supply reliability.

[0017] The adjustment rules are stored in the rule engine, allowing administrators to dynamically modify them based on market changes. During the promotion period of green energy policies, the lower limit of the green electricity preference weight can be temporarily increased to balance policy guidance and users' actual needs.

[0018] According to an embodiment of the present invention, in step S130, before constructing the optimization model, cluster analysis needs to be performed on users to identify user groups with similar characteristics. Using the K-means algorithm combined with dimensions such as electricity consumption behavior (peak-to-valley ratio, average monthly load), social attributes (housing type, income level estimation), and historical emotional patterns, users are divided into 5-8 typical groups. A dedicated satisfaction function is established for each group, where the parameters are trained from representative samples of the group. Residents in older communities may be more concerned about electricity price stability than the proportion of green electricity, while residents in emerging technology parks value the proportion of green electricity and service response speed more.

[0019] In particle swarm optimization, multiple convergence conditions are set to prevent overfitting. In addition to the basic iteration limit, the following termination conditions are added: The satisfaction improvement rate is considered to be close to the local optimum when the global optimal solution satisfaction improvement rate is less than 0.01 in 5 consecutive iterations. To maintain diversity, the distribution entropy of the particle population is monitored. When the entropy value is lower than the threshold, a small number of random particles are forcibly injected to avoid premature convergence. Cost constraints apply, with a maximum allowable cost increase of no more than 5%. If this is exceeded, the system will automatically revert to the previous feasible solution.

[0020] An adaptive inertia weight mechanism is introduced. In the early stages of the algorithm, the inertia weight is used to promote global search. As the iteration progresses, the weight is gradually reduced to 0.4 to enhance local exploitation capabilities and achieve a dynamic balance between exploration and utilization.

[0021] According to an embodiment of the present invention, in step S140, the system generates a three-level explanation based on the user profile. The result layer clearly informs the user of the monthly electricity cost changes and expected satisfaction improvement after the package adjustment; the reason layer explains the basis for the adjustment; and the value layer emphasizes long-term benefits. The explanation text supports multiple tone styles, using concise and lively language for younger users and plain and easy-to-understand expressions for older users. Visualization components are provided for the explanation content, with bar charts showing changes in electricity cost composition and line graphs displaying the predicted satisfaction trend, helping users understand intuitively. The explanation content is pushed via SMS, and user confirmation feedback is used not only to update the package but also to optimize the model, associating the user's acceptance or rejection decision with the corresponding package parameters to train a classifier to predict user preferences.

[0022] According to an embodiment of the present invention, step S150, establishing a comprehensive system health monitoring dashboard, includes: Data stream monitoring: Real-time display of data acquisition success rate, latency, and number of abnormal alarms for each data source; Emotion model monitoring: Tracking emotion recognition accuracy, confidence distribution, and the number of false cases; Optimization process monitoring: Displays the convergence trajectory of the particle swarm, the satisfaction change curve, and the computational resource utilization rate; User feedback monitoring: Statistics and interpretation of SMS delivery success rate, user confirmation rate, and complaint conversion rate.

[0023] When any metric becomes abnormal, including data collection success rate <90% or continuous decline in user confirmation rate, an alarm process is automatically triggered to notify the operations and maintenance team to intervene and investigate.

[0024] This invention designs a tiered anomaly response mechanism to handle anomalies occurring during monitoring, including: Minor anomaly: A brief interruption of a single data source will trigger historical data completion to maintain normal system operation; Moderate anomaly: If the emotion model prediction error is >15%, switch to the backup model and start the model retraining process. Critical anomaly: Cloud service is unavailable. Automatically downgraded to independent operation at the edge, continuing to provide basic package recommendations using preset weights and simplified algorithms.

[0025] Simultaneously, a dual-active data center architecture is established, with primary and backup centers synchronizing configurations and model parameters in real time to ensure business continuity. During major holidays or severe weather, contingency plans are pre-loaded, sentiment monitoring frequency is increased, and more preferential package options are reserved to address potential fluctuations in user sentiment.

[0026] By connecting with the power grid dispatch system, real-time electricity prices and renewable energy generation forecasts are obtained to ensure that package pricing conforms to market principles. When connecting with the power trading center, the authenticity of green electricity trading certificates is verified to ensure the reliability of green electricity ratio data. When connecting with the marketing management system, information such as user contract changes and capacity adjustments is synchronized to avoid discrepancies between package recommendations and actual electricity consumption. Cross-system data exchange is achieved through standardized API interfaces, information security is ensured by adopting the OAuth 2.0 protocol, a data sharing whitelist system is established, and access permissions are strictly controlled.

[0027] The system has a built-in regulatory report generation function, which regularly outputs various reports that meet the requirements of the National Energy Administration, including: The generated "User Satisfaction Analysis Report" includes sentiment distribution maps by region and user type; Generate a "Green Energy Consumption Effectiveness Report" to quantify the emission reduction contribution of promoting green energy packages; Generate a "Package Adjustment Transparency Report" detailing the input parameters, calculation process, and output results for each optimization. All report generation processes are logged, allowing auditors to trace the entire decision-making chain for any package adjustment, thus meeting regulatory requirements for fairness and transparency.

[0028] Figure 2 This is a diagram of the operational architecture of the method of this invention.

[0029] like Figure 2 As shown, a comprehensive user behavior dataset is formed by collecting data through channels such as smart meters, customer service language-to-text conversion, APP interaction logs, and social media sentiment. After multi-source data cleaning and semantic alignment, abnormal meter load, meaningless text, and malicious social media content are removed. All data are calibrated to a 15-minute time slice. Text data is semantically annotated using BERT. The preprocessed data is input into the emotion recognition model BERT-LSTM to map multi-source text into continuous emotion intensity values. At the same time, in order to avoid instantaneous fluctuations, a time smoothing operation is introduced, as shown in Equation (1).

[0030] (1) in, For users At the present moment The smoothed emotional intensity value, For users At a historical moment The instantaneous emotion intensity recognition results By averaging the emotional observation window length across multiple time points, the system focuses on persistent dissatisfaction rather than momentary emotional fluctuations, thereby improving the stability of strategy adjustments.

[0031] Based on the emotion recognition results, it is determined whether the emotion threshold is exceeded. If it is exceeded, it is sorted according to the load ratio, renewal intention and historical complaint number. High-value and high-risk users are given priority to calculate personalized satisfaction. If the emotion threshold is not exceeded, it directly enters the benchmark weight process, as shown in the following formula (2).

[0032] (2) in, For the current moment The actual weights of each satisfaction dimension As the benchmark weight, This is the coefficient for amplifying or inhibiting emotions. The system uses a baseline weight to determine the user's emotions when the user's emotions are stable, ensuring the long-term rationality of the package evaluation. When the user's emotions exceed the threshold, the system automatically triggers weight adjustment, amplifying the factors that the user cares about most at the moment and appropriately weakening secondary or long-term factors.

[0033] By merging the results of personalized satisfaction and standard satisfaction into the optimization model, the objective function in the optimization model maximizes group satisfaction and minimizes operating costs, as shown in equation (3) below.

[0034] (3) in, Let be the decision variable, representing the vector of package configuration parameters to be optimized. For users For packages The satisfaction rating score This represents the total number of users currently participating in the evaluation. This is the satisfaction weighting coefficient. This represents the cost weighting coefficient. The optimization algorithm identifies the optimal package based on a comprehensive consideration of factors such as satisfaction and cost. Candidate solutions are ranked according to satisfaction priority, and the solution with the highest satisfaction that satisfies all constraints is selected as the optimal package, and the optimal package recommendation is output.

[0035] The system determines whether a user has high negative emotions. If so, a personalized explanation is generated, and a personalized SMS message is sent highlighting the electricity savings and peak-hour price reductions. If not, a general explanation is generated, and a general SMS message is sent, supplementing the description of the suitable user group for the plan. After the initial SMS push, if the user doesn't read it within 2 hours, an automatic push notification is sent via an app pop-up and a WeChat official account template message. If the user still hasn't confirmed within 4 hours, a customer service follow-up call is triggered for high-value users, recording their concerns and synchronizing them to the optimization model. If the user refuses to confirm, an automatic survey is pushed, recording the results immediately after the user's selection. If the user doesn't select anything, customer service calls back within 24 hours to collect additional information and mark it in the user profile. Weekly summaries of user confirmation data, reasons for non-confirmation, and changes in satisfaction after the plan takes effect are compiled, and the model's weight mapping rules and constraints are adjusted accordingly.

[0036] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0037] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A particle swarm optimization method for configuring electricity market packages considering user satisfaction, characterized in that, Includes the following steps: S1. Collect multi-source user data and use the BERT-LSTM model to calculate the intensity of user emotions in real time; S2. Dynamically adjust the weights of each dimension in the satisfaction model based on the intensity of emotion to construct an emotion-satisfaction mapping mechanism; S3. For users with negative emotions, generate the minimum perturbation package adjustment direction based on counterfactual interpretation technology and aggregate it into an explanatory gradient vector; S4. In the speed update of the particle swarm optimization algorithm, the explanatory gradient vector is introduced as a driving term to guide the particles to search in the direction that is most likely to improve satisfaction. S5. Output the optimized package plan and push personalized explanations and suggestions to users, and update the model based on user feedback.

2. The method according to claim 1, characterized in that, step In S1, the emotional intensity is updated every 15 minutes. If it exceeds the set threshold for multiple consecutive cycles, it is marked as a user with high negative emotions and counterfactual interpretation is performed first.

3. The method according to claim 1, characterized in that, In step S2, the emotion-satisfaction mapping mechanism is implemented through a piecewise linear mapping table, which is stored in the local database of the edge gateway and supports remote online updates.

4. The method according to claim 1, characterized in that, In step S2, the satisfaction model includes at least four dimensions: electricity price level, green electricity ratio, power constraint penalty, and service response time. The weight of each dimension is dynamically adjusted according to user sentiment.

5. The method according to claim 1, characterized in that, In step S3, the counterfactual explanation is solved by the projection gradient method, setting the maximum number of iterations and the step size, and limiting the perturbation amplitude within a set range to prevent sudden changes in the package.

6. The method according to claim 1, characterized in that, In step S4, the position vector of the particle in the particle swarm optimization algorithm includes four real-value dimensions: peak-valley price difference, green electricity ratio, value-added service level, and power factor reward, and has a corresponding value range.

7. The method according to claim 1 or 6, characterized in that, The inertia weights in the particle swarm optimization algorithm employ a linear descent strategy to explain the exponential decay of the strength coefficients of the driving terms over time.

8. The method according to claim 1, characterized in that, In step S5, the personalized explanation suggestion is pushed in the form of a quantitative SMS message, which includes the adjustment range of the green electricity ratio, changes in electricity costs and an estimated improvement in satisfaction, and provides a one-click confirmation interface.

9. The method according to claim 1, characterized in that, The method adopts an edge-cloud collaborative architecture, with emotion recognition and weight mapping performed on the edge gateway, and counterfactual solving and particle swarm optimization performed on the cloud GPU server, resulting in an end-to-end latency of less than 5 minutes.

10. The method according to claim 1, characterized in that, When the average emotional intensity of all users continues to exceed the set threshold, a new round of package optimization is automatically triggered, and an early warning notification is sent to the operations staff.

Images