Multi-objective coordination and feature clustering based dynamic benchmark load design method and system

Abstract

The application relates to the field of power systems and automation technology, and discloses a dynamic reference load design method and system based on multi-target cooperation and feature clustering, which comprises the following steps: obtaining user historical power consumption data and type labels, extracting shape and regularity features, performing feature fusion and target guided constraint clustering, and generating a user feature portrait library; receiving a regional level load regulation target issued by a power grid, combining the user feature portrait library, and obtaining a personalized node load baseline that is consistent with user behavior characteristics through a multi-target optimization model; based on the user feature portrait library and the real-time state of the power grid, a differentiated user utility function is constructed, and a dynamic incentive price is formulated for different user clusters; the personalized node load baseline and the dynamic incentive price are issued to the users, actual response data of the users is collected, the user feature portrait library is updated, and a closed-loop optimization process is formed; and the application realizes the scaling and efficient regulation of massive heterogeneous distributed loads.

Description

Technical Field

[0001] This invention relates to the field of power systems and their automation technology, specifically to a dynamic baseline load design method and system based on multi-objective collaboration and feature clustering. Background Technology

[0002] In the context of current power system construction, how to efficiently and on a large scale mobilize massive distributed load-side resources and balance the volatility of a high proportion of renewable energy has become a core challenge for grid operation. Demand response is a key means to address this challenge, but existing technological solutions have significant bottlenecks in terms of implementation scale and sustainability.

[0003] Current mainstream demand response solutions are divided into price-based demand response (PDR) and incentive-based demand response (IDR). Price-based demand response guides users through indirect price signals, which suffers from signal lag, is prone to load inversion (peak-valley imbalance), and exhibits a "rebound effect." Furthermore, dynamic pricing is limited by policy and user acceptance. Incentive-based demand response relies primarily on customer baseline load (CBL) estimation methods, which suffers from heavy computational burden, fairness concerns, and data scarcity in extreme weather or new user scenarios, making it difficult to support routine, lightweight deployment for tens of millions of users.

[0004] To achieve deep integration of grid regulation needs with users' commercial and utility goals, the Location-based Customer Baseline Load (LCDL) mechanism has emerged as a breakthrough solution. Through its "goal-oriented" and "network-coupled" design, it addresses issues such as inaccurate guidance signals, neglect of network constraints, and unfair settlement benchmarks inherent in traditional solutions. This mechanism abandons the traditional model, with system operators calculating and publishing ideal load curves (LCDLs) for user groups in different geographical locations. Guidance is achieved using a two-layer Stackelberg game model, iteratively adjusting incentive prices and response quantities to balance the economics, privacy, and interests of all parties. Simultaneously, a hierarchical guidance and network constraint decomposition scheme is employed. Transmission System Operators (TSOs) distribute transmission-level LCDLs to Distribution System Operators (DSOs), who then decompose them into distribution-level LCDLs based on local grid conditions and distribute them to end users. The algorithm incorporates grid power flow constraints, enabling precise scheduling of demand response resources.

[0005] However, the existing location-based customer benchmark load mechanism mainly focuses on grid-side regulation demand, treating users as homogeneous entities that simply respond to economic incentives, which has significant shortcomings. First, the target model is too simplistic, only considering grid-side regulation demand and failing to integrate diverse optimization objectives from the user side. The core demands of industrial, commercial, and residential users differ significantly, and this homogeneous guidance severely reduces user participation. Second, user electricity consumption behavior characteristics are vague, failing to fully utilize multi-dimensional features such as volatility and peak-valley differences obtained from machine learning clustering. It uses load following degree as the sole basis for settlement and fails to develop differentiated incentive strategies based on user electricity consumption characteristics.

[0006] In summary, the existing location-based customer benchmark load mechanism is insufficient to stimulate multi-dimensional interaction among power generation, grid, load, and storage, and cannot fully unleash the regulation potential of massive distributed load resources. There is an urgent need to build a collaborative guidance system that takes into account both grid demand and the diverse goals and behavioral characteristics of users, so as to provide technical support for the construction of a highly resilient new power system. Summary of the Invention

[0007] To overcome the shortcomings of the existing technology, the present invention aims to provide a dynamic benchmark load design method and system based on multi-objective collaboration and feature clustering. This addresses the technical problem that the existing location-based customer benchmark load mechanism designs load benchmarks and incentive strategies solely based on grid-side regulation needs, without considering the diverse optimization objectives and multi-dimensional electricity consumption behavior characteristics of the user side. It adopts a homogeneous guidance and a single economic incentive model, resulting in low user participation, insufficient response accuracy, and difficulty in achieving large-scale, refined, and efficient regulation of massive heterogeneous distributed loads.

[0008] This invention is achieved through the following technical solution: In a first aspect, the present invention provides a dynamic baseline load design method based on multi-objective collaboration and feature clustering, comprising: Obtain users' historical electricity consumption data and type tags, extract morphological and regular features, and generate a user feature profile library through feature fusion and target-guided constraint clustering. The system receives regional load regulation targets from the power grid and, in conjunction with the user feature profile database, obtains personalized node load profiles that fit user behavior characteristics through multi-objective optimization model decomposition. Based on the user feature profile database and the real-time status of the power grid, we construct differentiated user utility functions and formulate dynamic incentive prices for different user clusters. The personalized node load baseline and dynamic incentive price are distributed to users, user actual response data are collected, and the user feature profile database is updated to form a closed-loop optimization process.

[0009] Preferably, the specific process of obtaining users' historical electricity consumption data and type tags, extracting morphological and regular features, and generating a user feature profile library through feature fusion and target-guided constraint clustering includes: The daily load curves of users are subjected to density-based noisy spatial clustering algorithm for morphological feature extraction, abnormal electricity consumption days are eliminated, and a morphological feature vector containing the main typical curve, morphological richness, main cluster concentration and noise ratio is constructed. Regularity features of user daily electricity consumption sequences and main typical curves are quantified to construct regularity feature vectors that include autocorrelation coefficients, curve consistency, and weekday-weekend differences. The morphological feature vector and the regularity feature vector are standardized and then fused. The GMM algorithm with a target consistency penalty term is used for clustering to generate user feature profiles.

[0010] Furthermore, the parameter adaptation process of the noisy spatial clustering algorithm includes: Calculate the dynamic time-normalized distance between load curves, plot the k-distance curve and take the inflection point value as the neighborhood radius, and set the minimum number of points to K+1, where K is the preset number of nearest neighbors.

[0011] Furthermore, in the GMM algorithm with a target consistency penalty term, the calculation of user posterior probability introduces a target type consistency indicator function to guide users with similar behavioral characteristics and consistent target types to be grouped into the same cluster.

[0012] Preferably, the objective function of the multi-objective optimization model includes a global target tracking term, a shape fitting term, and a regulation effort term, wherein the shape fitting weight and the regulation effort weight are determined based on the dynamic mapping of user feature profiles.

[0013] Furthermore, the constraints of the multi-objective optimization model include global objective constraints, user power limit constraints, adjustment rate constraints, dynamic power limit constraints based on adjustable potential, user-level comfort constraints, and relative fairness constraints within the cluster.

[0014] Preferably, the personalized node load profile that aligns with user behavior characteristics adopts a distributed-centralized hybrid solution strategy, the specific process of which is as follows: Preliminary guidelines are obtained by pre-decomposing the target region based on user characteristics and solving small-scale optimization problems within the cluster. Summarize the initial cluster plan, check the power flow constraints and target deviations, fine-tune the cluster target and rebalance to obtain a personalized node load profile that fits the user's behavior characteristics.

[0015] Preferably, the differentiated user utility functions include a cost-driven utility function for industrial users, a comfort-constrained utility function for commercial users, and a comprehensive trade-off utility function for residential users. The parameters of the cost-driven utility function for industrial users, the comfort-constrained utility function for commercial users, and the comprehensive trade-off utility function for residential users are dynamically determined based on user feature profiles.

[0016] Preferably, the dynamic incentive price consists of a base price, a feature adjustment item, and a potential incentive item, wherein the dynamic incentive price satisfies the total incentive budget constraint.

[0017] Secondly, the present invention also provides a dynamic baseline load design system based on multi-objective collaboration and feature clustering, comprising: The feature profiling module is used to obtain users' historical electricity consumption data and type labels, extract morphological and regular features, and generate a user feature profile library through feature fusion and target-guided constraint clustering. The guideline decomposition module is used to receive the regional load regulation target issued by the power grid, and combine it with the user feature profile library to obtain personalized node load guidelines that fit the user behavior characteristics through multi-objective optimization model decomposition. The incentive customization module is used to construct differentiated user utility functions based on the user feature profile library and the real-time status of the power grid, and to formulate dynamic incentive prices for different user clusters. The closed-loop optimization module is used to distribute the personalized node load baseline and dynamic incentive price to users, collect actual user response data, update the user feature profile database, and form a closed-loop optimization process.

[0018] Compared with the prior art, the present invention has the following beneficial technical effects: This invention provides a dynamic benchmark load design method based on multi-objective collaboration and feature clustering. Through the collaborative implementation of user feature extraction, constrained clustering, personalized guideline decomposition, and differentiated incentive pricing, it significantly improves user participation and response accuracy. Offline feature clustering accurately identifies user electricity consumption habits and intrinsic decision-making goals; dynamic weight mapping aligns the load guideline with user habits and reduces the regulation burden; and customized utility functions and incentive prices ensure that guiding signals are highly aligned with user interests, driving users from passive response to proactive optimization. Simultaneously, this invention accurately quantifies the predictability and stability of user behavior, assigning refined regulation tasks to users with different characteristics and providing targeted incentives, deeply exploring the regulation potential of heterogeneous loads, and achieving a dual improvement in the quality and quantity of regulation potential. Furthermore, personalized guidelines and high-precision responses can accurately match renewable energy output fluctuations, improving the capacity for clean energy absorption; the hierarchical decomposition strategy incorporates network constraints, avoiding line overload and enhancing grid operation safety; and differentiated incentives reduce ineffective expenditures, lower regulation costs, and improve the economic efficiency of grid operation. Meanwhile, the utility model that aligns with user decision-making logic promotes rapid game convergence, clustering and fairness constraints ensure the balance of user obligations, the layered distributed architecture reduces computational complexity, and it is highly inclusive of new users, providing scalable and lightweight technical support for large-scale deployment with tens of millions of users. Attached Figure Description

[0019] Figure 1 This is a flowchart of the dynamic baseline load design method based on multi-objective collaboration and feature clustering in an embodiment of the present invention; Figure 2 This is a framework diagram of the dynamic baseline load design method based on multi-objective collaboration and feature clustering in an embodiment of the present invention; Figure 3 This is a flowchart of the user feature extraction and clustering module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the daily load variation curves of a typical residential user on multiple dates in an embodiment of the present invention; Figure 5 This is a schematic diagram of the GMM user profile clustering results constructed based on user behavior features in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the changes in the user baseline load curve before and after demand response optimization in an embodiment of the present invention. Figure 7 This is a schematic diagram illustrating the overall change in regional total load before and after demand response optimization in an embodiment of the present invention; Figure 8 This is a schematic diagram of the dynamic baseline load design system based on multi-objective collaboration and feature clustering in an embodiment of the present invention. In the diagram: 1. Feature profiling module; 2. Directrix decomposition module; 3. Incentive customization module; 4. Closed-loop optimization module. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0021] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] The purpose of this invention is to provide a dynamic benchmark load design method and system based on multi-objective collaboration and feature clustering, in order to solve the technical problems of existing location-based customer benchmark load mechanisms that only design load benchmarks and incentive strategies based on grid-side regulation needs, without considering the multi-dimensional optimization objectives and multi-dimensional electricity consumption behavior characteristics of users, and adopting homogeneous guidance and a single economic incentive mode, resulting in low user participation, insufficient response accuracy, and difficulty in achieving large-scale, refined, and efficient regulation of massive heterogeneous distributed loads.

[0023] The present invention will now be described in further detail with reference to the accompanying drawings: Example 1 See Figure 1 and Figure 2 In one embodiment of the present invention, a dynamic baseline load design method based on multi-objective collaboration and feature clustering is provided, comprising: Step 1: Obtain users' historical electricity consumption data and type tags, extract morphological and regular features, and generate a user feature profile library through feature fusion and target-guided constraint clustering. Specifically, according to Figure 3 As shown, the specific process is as follows: S1. The Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm is used to extract morphological features from the user's daily load curve, eliminate abnormal electricity consumption days, and construct a morphological feature vector that includes the main typical curve, morphological richness, main cluster concentration, and noise ratio. Among them, the set of daily load curves for D consecutive days of input users User registration type tags A noisy spatial clustering algorithm based on dynamic time warping distance is used for processing. Adaptive determination of neighborhood radius using the k-distance inflection point method With minimum number of points =K+1, cluster the load curves and identify abnormal noise days, then extract the main typical curves. morphological richness Cluster concentration Noise ratio To construct morphological feature vectors; where, The number of non-noise curves; Among them, for users DBSCAN may output From a typical morphological cluster (some users may only have one main cluster), extract the following morphological features to construct a vector. : The principal canonical curve is obtained by taking the median of all curves within the cluster at the corresponding time point and extracting the cluster center curve of the cluster with the most members (the principal cluster). .

[0024] Morphological richness: The value itself, as a feature, reflects the diversity of users' electricity consumption patterns (such as whether there is a clear "weekday-weekend" dual mode).

[0025] Cluster concentration: This refers to the proportion of the number of main cluster curves to the total number of non-noise curves, reflecting the consistency of power consumption patterns.

[0026] Noise ratio: This directly reflects the frequency or unpredictability of abnormal user electricity consumption behavior.

[0027] S2, quantify the regularity features of the user's daily electricity consumption sequence and the main typical curve, and construct a regularity feature vector including autocorrelation coefficient, curve consistency, and weekday-weekend differences; Among them, based on daily electricity consumption sequence The expressions for calculating the autocorrelation coefficients with a lag of 1 day and 7 days are as follows:

[0028] in, For users The autocorrelation coefficient of the daily electricity consumption series with a 1-day lag indicates that the electricity consumption on adjacent days is highly correlated and has a strong regularity. For users In the Total electricity consumption per day; For users exist Average daily electricity consumption within a day; D represents the user. The total number of historical days (length of the daily electricity consumption sequence); L is the lag days, which can be 1 or 7.

[0029] Calculate the load curves and the main typical curve for all non-noise days. The mean and coefficient of variation of Dynamic Time Warping (DTW) are expressed as follows:

[0030]

[0031] in, For users The smaller the average DTW distance between the load curve and the main typical curve for all normal days (non-noise days), the more stable the user's daily electricity consumption pattern is. The coefficient of variation is the distance to DTW. The smaller the value, the better the consistency of the user's daily electricity consumption pattern and the smaller the fluctuation. A set of non-noise curves For its quantity, It is the dynamic time-warped distance between two curves, used to measure shape similarity. It is the load curve vector for a certain day. User The main typical curve, It is a standard deviation operator used to calculate the average daily load curves for weekdays and weekends, and to calculate the DTW distance between the two. The greater the distance, the stronger the influence of social routines on the user.

[0032] S3. After standardizing the morphological feature vector and the regularity feature vector, they are fused together, and clustering is performed using a Gaussian Mixture Model (GMM) with a target consistency penalty term to generate a user feature profile.

[0033] Among them, the extracted morphological feature vectors and regular eigenvectors After standardization (Z-score), the data are concatenated to form a high-dimensional comprehensive feature vector of the user. .

[0034] A Gaussian mixture model is used for the final user segmentation. The Gaussian mixture model provides probabilistic soft classification, which better reflects the fact that user behavior may exhibit intermediate states. Its objective function is to maximize the log-likelihood:

[0035] in, , These are the weights, mean, and covariance matrix of the k-th Gaussian component, respectively. N K represents the total number of users; K is the preset number of clusters in the Gaussian mixture model, i.e., the number of user clusters.

[0036] In the Expectation-Maximization Algorithm (EM) for Gaussian mixture models, a goal consistency penalty term is introduced. We compute the algorithm for each user in the E-step. Posterior probability of belonging to each cluster k At that time, not only feature likelihood is considered, but also the target type tag registered by the user. (Industrial / Commercial / Residential) and the dominant target type of cluster k Consistency.

[0037]

[0038] in, For indicator functions; This is the penalty coefficient; For users Belongs to the Update the posterior probability of each Gaussian component; For the first The mixing weights of Gaussian components; User feature vector Gaussian components The probability density function under; It is the first The mean vector of Gaussian components; For the first The covariance matrix of Gaussian components, For users The type label declared during registration (industrial / commercial / residential); The dominant target type for the k-th cluster.

[0039] This operation guides GMM to group users with similar behavioral characteristics (morphology + regularity) and consistent underlying business goals into the same Gaussian component with a higher probability, thereby realizing the integration of data-driven and domain knowledge and constrained clustering with consistent feature and target types, generating a user feature profile library.

[0040] Ultimately, K user clusters are obtained, each described by a set of Gaussian distribution parameters and having a dominant optimization objective type.

[0041] Step 2: Receive the regional load regulation target issued by the power grid, and combine it with the user feature profile database to obtain a personalized node load profile that fits the user behavior characteristics through multi-objective optimization model decomposition. Specifically, the parameter adaptation process of the noisy spatial clustering algorithm includes: Calculate the dynamic time-normalized distance between load curves, plot the k-distance curve and take the inflection point value as the neighborhood radius, and set the minimum number of points to K+1, where K is the preset number of nearest neighbors.

[0042] Specifically, in the GMM algorithm with target consistency penalty term, the user posterior probability calculation introduces a target type consistency indicator function to guide users with similar behavioral characteristics and consistent target types to be grouped into the same cluster.

[0043] Specifically, the objective function of the multi-objective optimization model includes a global target tracking term, a shape fitting term, and a regulation effort term, wherein the shape fitting weight and the regulation effort weight are determined based on the dynamic mapping of user feature profiles.

[0044] The constraints of the multi-objective optimization model include global objective constraints, user power limit constraints, adjustment rate constraints, dynamic power limit constraints based on adjustable potential, user-level comfort constraints, and relative fairness constraints within the cluster.

[0045] In this embodiment, the specific process is as follows: Assume the total load regulation target issued by the upper-level power grid to this region (managed by the Distribution System Operator (DSO)) for each time period t within the scheduling cycle is... The DSO needs to break it down into its jurisdiction. Personalized baseline for each user .

[0046] The optimized model in this embodiment is:

[0047]

[0048]

[0049] in, For users During the period Personalized load guidelines; This represents the total number of time periods within the scheduling cycle. For time period Global target tracking weights; For the upper-level power grid to issue to this region during the time period Total load regulation target; It represents the total number of users participating in demand response within the region; For users The shape fits the weight; This refers to the deviation in shape fit. For users Adjustment effort weight; To adjust for effort deviation; For users During the period The forecast baseline load.

[0050] The objective function of the multi-objective optimization model includes a global target tracking term, a shape fitting term, and a regulation effort term. Global target tracking item: Penalizes deviations between the total load of the region and the target value to ensure that the grid regulation needs are met. The weighting is based on time periods, which can be increased during critical adjustment periods, such as when renewable energy output drops sharply or when network congestion is severe, to enhance the tracking accuracy of regional targets.

[0051] Shape fit item: Measuring users exist Standard value for the time period Its principal typical curve The degree of deviation is designed to make the guideline shape closer to user habits. Weighting based on shape.

[0052] Adjusting effort items: Measurement baseline and user forecast baseline load The deviation represents the adjustment effort that the user needs to make. To adjust the effort weight.

[0053] The specific expressions for the shape fit weight and the adjustment effort weight are as follows:

[0054]

[0055] in, For users The coefficient of variation for load curve consistency; the smaller this value, the more stable the user's electricity consumption pattern. Therefore, This value is negatively correlated with the fact that for users whose behavior is more stable, the personalization guidelines should be more strictly required to conform to their inherent patterns, because their habits are difficult to change.

[0056] For users The basic adjustment willingness coefficient indicates that the higher the value, the stronger the inherent resistance of users to load adjustments. Therefore, This value is negatively correlated with the absolute adjustment effort required from users with low willingness to adjust.

[0057] This is a scaling parameter used to adjust the sensitivity of the feature's influence.

[0058] To fit the global form with the preference coefficient The setting is determined by the DSO based on the urgency of the current adjustment task. When the value is close to 1, comfort is prioritized; when it is close to 0, the goal is achieved with the minimum total adjustment.

[0059] The constraints in this embodiment are as follows: (1) Global objective constraints:

[0060] in, To ensure that the total load regulation target issued by the upper-level power grid to this region in time period t is strictly equal to the regional target, this is the fundamental requirement of the decomposition, guaranteeing the realization of the power grid-side regulation demand.

[0061] (2) User power limit constraints:

[0062] in, For users The minimum permissible power consumption; For users Maximum permissible power consumption; Consider the user's physical power consumption capacity, such as the upper and lower limits of the power of electrical equipment and the contracted capacity, to ensure that the benchmark is enforceable.

[0063] (3) Adjustment rate constraint:

[0064] in, For users Maximum allowable load change rate between adjacent time periods.

[0065] Considering the physical inertia of user load regulation, avoid sudden load changes between adjacent time periods to improve the smoothness and operability of the baseline.

[0066] (4) Dynamic constraints on upper and lower power limits based on adjustable potential:

[0067] The dynamic limit is determined by both user characteristics and baseline load.

[0068]

[0069] in, It is the maximum potential for load reduction / increase based on historical user data and equipment information; For calling coefficients; For users During the period The dynamic minimum power limit; For users During the period Dynamic maximum power limit; In this invention, For users with high noise levels and abnormal behavior, their potential for being invoked will be significantly reduced.

[0070] (5) User-level comfort and fairness constraints:

[0071]

[0072] in, The total number of time periods. For users The principal typical curve in the time period The value, For users The acceptable average deviation from the threshold is determined by the stability characteristics in the user profile. For the first A set of user clusters. This constraint requires that the relative adjustment rates of users within the same cluster be as equal as possible to ensure fairness; For users During the period The forecast baseline load.

[0073] Specifically, the personalized node load profile, which aligns with user behavior characteristics, employs a distributed-centralized hybrid solution strategy. The specific process is as follows: Preliminary guidelines are obtained by pre-decomposing the target region based on user characteristics and solving small-scale optimization problems within the cluster. Summarize the initial cluster plan, check the power flow constraints and target deviations, fine-tune the cluster target and rebalance to obtain a personalized node load profile that fits the user's behavior characteristics.

[0074] In this embodiment, in each feature cluster Internally, it performs rapid pre-decomposition based on cluster aggregation characteristics and adjustment tasks.

[0075] Power distribution system operators will have overall regional targets Based on the historical electricity consumption ratio of the cluster and the average regulation capacity of the cluster, the targets for each cluster are initially decomposed.

[0076] Cluster coordinator receives Solve the cluster optimization problem:

[0077]

[0078] in, It is a cluster During the period The target load (assigned by the upper layer); Weights are assigned to fit the form of user i. This refers to the deviation in shape fitting. The adjustment effort weight for user i; To adjust for effort deviation.

[0079] By treating the cluster optimization problem as a small-scale problem that can be solved quickly, a preliminary directrix scheme can be obtained.

[0080] The power distribution system operator compiles preliminary plans from all clusters, checks for constraints such as overall network power flow security, and identifies any instances of exceeding limits or deviations from the overall target. Global coordinated optimization is then initiated, but at this point, the decision variables are simplified to the aggregated adjustment amounts of each cluster. This reduces the complexity of the problem by adjusting task transfer amounts between clusters, rather than focusing on baselines for all users. Based on the coordination results, the objectives for each cluster are fine-tuned. It also notifies the cluster coordinator to perform internal reallocation.

[0081] Step 3: Based on the user feature profile database and the real-time status of the power grid, construct differentiated user utility functions and formulate dynamic incentive prices for different user clusters; Specifically, the differentiated user utility functions include cost-driven utility functions for industrial users, comfort-constrained utility functions for commercial users, and comprehensive trade-off utility functions for residential users. The parameters of each utility function are dynamically determined based on user feature profiles.

[0082] Specifically, the dynamic incentive price consists of a base price, feature adjustment items, and potential incentive items, and the dynamic incentive price satisfies the total incentive budget constraint.

[0083] The specific process in this embodiment is as follows: Establish a mapping function from user characteristics to game parameters. Key parameters include: Basic adjustment willingness coefficient:

[0084] in, For users The basic adjustment willingness coefficient; These are the weighting coefficients. For users noise ratio, For users The coefficient of variation of the consistency of the load curve.

[0085] This coefficient comprehensively reflects the user's inherent sensitivity to load adjustments. A high noise-to-noise ratio indicates more behavioral abnormalities, poor predictability, and greater difficulty in adjustment. A high (consistency coefficient of variation) indicates large fluctuations in daily electricity consumption and poor stability. The weighted sum of these two factors constitutes the basic assessment of the difficulty of regulation.

[0086] Adjusting the cost sensitivity coefficient (taking industrial users as an example):

[0087] in, This is a cost-sensitive baseline value for industrial users. For users The basic adjustment intention coefficient; multiplied by Incorporate behavioral characteristics; For users The autocorrelation coefficient of the daily electricity consumption series lagged by 7 days (reflecting weekly periodicity); multiplied by This is to take into account users with strong weekly patterns ( (Approaching 1) Temporarily adjusting the production plan is more costly.

[0088] Sensitivity coefficient to loss of comfort (taking business users as an example):

[0089] in, This is a basic value sensitive to the comfort level of business users, and it is a time-sensitive function (e.g., the value is higher during peak business hours). It is a time-sensitive function (e.g., it takes a higher value during peak business hours); For users Cluster concentration (reflecting consistency in power consumption patterns); multiplied by This is because users with low cluster concentration may be more sensitive to changes in comfort.

[0090] Based on the above parameters, this embodiment constructs a user utility function: Industrial (IND) users - cost-driven:

[0091] in, For industrial users Total utility; For the user's cluster During the period Incentive price; For users During the period Load adjustment amount; For users Cost sensitivity coefficient; For users During the period The forecast baseline load; For users During the period The actual load; It is a time period The real-time market electricity price (RMB / kWh).

[0092] Users adjust the load Obtain incentive benefits However, cost changes may occur due to adjustments in production plans (based on real-time electricity prices). (Calculation). Utility maximization is the pursuit of the optimal trade-off between economic incentives and production costs.

[0093] Commercial (COM) users - comfort-constrained:

[0094] in, For business users Total utility; For users During the period The comfort loss sensitivity coefficient; This is used to explain that a loss of comfort only occurs when the load is reduced, and does not occur when the load is increased.

[0095] The incentive benefits are the same as those for industrial users; the second term is the cost of comfort loss, expressed as a quadratic term for load reduction. The discomfort it causes, and the sensitivity level Values ​​vary over time, with higher values ​​observed during peak business hours.

[0096] Residential (RES) - Comprehensive Trade-off:

[0097] in, For residential users Total utility; The cost coefficient is adjusted linearly. It is the quadratic discomfort coefficient; It is the green preference coefficient, representing the degree of importance users attach to the consumption of renewable energy; This is a green labeling function, taking the value 1 during periods of high renewable energy availability and 0 otherwise. This function simultaneously considers: incentive benefits, linear adjustment costs (proportional to the absolute value of the adjustment), quadratic discomfort, and green contribution benefits. This is a green labeling function; it is 1 for periods with high renewable energy activity and 0 otherwise.

[0098] In this embodiment, a dynamic incentive pricing mechanism based on cluster characteristics and grid demand is used.

[0099] in, This is the base price calculated based on system balance costs (such as wind and solar curtailment costs, reserve costs, etc.).

[0100] = For feature adjustment terms, and cluster average adjustment difficulty Positive correlation; higher incentives are given to clusters that are difficult to regulate in order to compensate for their participation costs. For clusters The average basic willingness to adjust coefficient of all users within the system;

[0101] This is a potential incentive term used to explore the cluster's regulatory potential. For clusters in time periods The greater the projected adjustable potential, the higher the incentive. ,Right now This encourages clusters with historically low response reliability to improve their performance; Scaling factor for potential incentive terms Constraints: DSO pricing must meet budget constraints. ,in For the total incentive budget, This represents the cluster response volume predicted based on the user utility function.

[0102] Step 4: Distribute the personalized node load baseline and dynamic incentive price to users, collect actual user response data, update the user feature profile database, and form a closed-loop optimization process.

[0103] Specifically, the generated personalized node load baseline Corresponding cluster incentive price The data is distributed to end users, who respond to the load based on their own utility functions. The system collects the actual electricity consumption and response data of users, updates the user electricity consumption characteristics, clustering results and utility parameters based on the feedback data, and iteratively optimizes the user feature profile library to provide data support for the next round of baseline decomposition and incentive pricing, thereby achieving continuous closed-loop optimization.

[0104] To verify the feasibility, robustness, and engineering applicability of this method in a real load data environment, this experiment used smart meter data from a residential community in Anning District, Lanzhou City, Gansu Province in March 2025, which included load records from 500 households over 30 consecutive days at 15-minute sampling intervals.

[0105] The overall solution is based on real residential load data. Through load clustering, typical curve extraction, user behavior feature analysis, user profile construction, multi-objective optimization, and dynamic incentive feedback, it realizes personalized load baseline generation and dynamic optimization control.

[0106] The two code snippets respectively implement: realdata_method.py -- This mainly performs real data reading, DTW clustering, main cluster identification, typical curve extraction, and basic behavior analysis.

[0107] baseline.py -- Based on the aforementioned results, it further implements user profile construction, dynamic weight mapping, multi-objective optimization, dynamic incentive pricing, and closed-loop feedback mechanism to form a complete methodology.

[0108] Therefore, this embodiment not only implements the core algorithm logic, but also completes the entire engineering chain from real data input to final optimized output, and has strong practical application feasibility.

[0109] The experimental data uses real residential user electricity consumption data files. The data structure conforms to the actual data collection methods of smart meters in the current power system. Therefore, the data source has a strong realistic basis and can truly reflect the electricity consumption behavior characteristics of residential users. It includes 500 users, with 30 days of data collected from each user, totaling 15,000 records. This indicates that the experiment is not based on small-scale simulated data, but on large-scale real user data for analysis. In actual operation, the code can stably complete: multi-user loop; DTW distance matrix construction; cluster analysis; GMM modeling; multi-objective optimization; indicating that the solution has engineering feasibility at a certain scale. From the original load curve, we can observe: obvious morning and evening peaks; behavioral differences among users; large load fluctuations for some users; and obvious periodic patterns for some users. These characteristics are consistent with real residential load patterns, indicating that the data has high reliability and can support the verification of the patented method.

[0110] This case study utilizes the dynamic baseline load design method based on multi-objective collaboration and feature clustering provided in this embodiment. The specific process is as follows: A complete data preprocessing workflow was built into the code to ensure effective operation. The code automatically recognizes the 96-dimensional time load column, avoiding fixed field dependencies and adapting to different Excel file formats, thus improving project adaptability. We will subsequently convert the load column to a numeric type and implement missing value filling and negative value truncation.

[0111] This scheme uses the Dynamic Time Warping (DTW) algorithm to calculate the distance between load curves. DTW can align similar fluctuation characteristics at different time points by dynamically stretching the time axis, thereby more accurately reflecting the true similarity between residents' load behaviors.

[0112] In the code implementation, the user's multi-day 96-dimensional load curve is first read, then the DTW distance matrix is ​​constructed, and load behavior clustering is completed using DBSCAN. DBSCAN can automatically identify density clusters without pre-specifying the number of clusters, and can also automatically identify abnormal load days, making it more suitable for the random fluctuation characteristics of residential load scenarios.

[0113] Experimental results show that DTW clustering can effectively distinguish different types of resident behavior patterns, with relatively stable main cluster identification results, and most users exhibiting obvious dominant behavior patterns. Furthermore, abnormal load days can be automatically identified as noise points, indicating that the method has good robustness to random disturbances. Therefore, based on practical results, the DTW clustering scheme demonstrates strong engineering feasibility.

[0114] After completing DTW clustering, the system further identifies the main behavioral patterns in user load. Because residential load exhibits occasional abnormal behaviors, such as holiday gatherings, prolonged absences, or extreme weather, directly using the average load curve is easily affected by outliers, leading to distortion of typical behaviors.

[0115] Therefore, this case study uses the "largest cluster + median curve" approach to extract typical load curves. Specifically, the program first identifies the cluster with the largest number of samples in the clustering results, and then takes the median of all curves within the cluster at different time points to generate the user's typical load curve.

[0116] Compared to the traditional average method, the median can effectively suppress the influence of outliers, making the final typical curve more stable and better reflecting the long-term real behavioral characteristics of users.

[0117] Experimental results show that most users exhibit distinct dominant behavioral patterns, with a high proportion of dominant clusters, indicating that while residential load fluctuates, it maintains a relatively stable overall electricity consumption pattern. Furthermore, the median curve closely reflects users' daily behavioral characteristics without significant abnormal distortions. Therefore, this method can effectively construct a stable and reliable personalized load baseline.

[0118] according to Figure 4 The figure shows the daily load variation curves of a typical residential user on multiple dates. The horizontal axis represents 96 15-minute sampling times within a day, and the vertical axis represents the normalized load value. The figure clearly shows a relatively stable electricity consumption pattern for the residential user: lower load in the early morning, a significant morning peak around 7:00 to 9:00 AM, and a higher evening peak around 6:00 to 9:00 PM. The overall trend of the curves across different dates is relatively consistent, with only minor fluctuations in local periods, indicating strong periodicity and regularity in this user's behavior. This result demonstrates that real residential load data has been successfully read and preprocessed, effectively supporting subsequent user profiling analysis, extraction of the main typical curve, and demand response optimization modeling.

[0119] To further characterize the differences in resident user behavior, the system extracts multidimensional behavioral features from the clustering results, including: shape_richness (richness of behavioral patterns); main_cluster_ratio (the proportion of stable main clusters—used to reflect the stability of user behavior); noise_ratio (the proportion of abnormal behavior—used to reflect the degree of random fluctuation of users); autocorrelation function (ACF) (used to characterize periodic regularity); DTW average distance; DTW coefficient of variation (used to measure the complexity of behavior); intraday fluctuation characteristics, etc.

[0120] These features can comprehensively describe residents' load behavior patterns from multiple perspectives. Experimental results show significant differences in stability, periodicity, and volatility among different users. For example, some users exhibit obvious stable periods, while others show strong random fluctuations. This indicates that the constructed feature system can effectively reflect the heterogeneity of real residents' behavior, providing a reliable foundation for subsequent user profiling and personalized optimization.

[0121] After behavioral features are extracted, the system uses a Gaussian Mixture Model (GMM) to construct user profiles. Compared to traditional K-Means clustering, GMM can describe the category to which a user belongs through a probability distribution, making it more suitable for handling data scenarios with overlapping behavioral characteristics, such as resident workload. The program first standardizes all features, then uses a Gaussian Mixture Model to perform user clustering, and automatically generates user labels based on the clustering characteristics, including: stable users; users with strong periodic patterns; users with high volatility; and multi-mode users, etc.

[0122] Experimental results show that different user categories can form relatively clear cluster boundaries in the two-dimensional PCA space, and the user feature space has already shown a certain clustering structure, indicating that there are behavioral differences among different resident users, which can support the formulation of subsequent personalized demand response strategies. This demonstrates that GMM can effectively distinguish different resident behavioral patterns. Furthermore, the generated user tags have a high degree of consistency with actual load behavior, exhibiting good interpretability.

[0123] according to Figure 5 The figure shows the clustering results of the GMM user profile built based on user behavior characteristics. First, a user behavior feature space is constructed using indicators such as load volatility, autocorrelation, DTW distance, and principal cluster percentage. Then, Principal Component Analysis (PCA) is used to map the high-dimensional features onto a two-dimensional plane for visualization. The figure shows that different categories of users have formed a certain degree of clustering structure in the feature space, with some user clusters having clear boundaries and others exhibiting some overlap. This indicates that there are indeed differentiated electricity consumption behaviors among residential users, and also reflects the strong similarity of the actual residential load data itself. Overall, the results demonstrate that the GMM clustering model can effectively classify user behavior characteristics, thus providing support for subsequent personalized baseline construction and dynamic demand response strategy formulation. Therefore, this method can achieve refined user profiles, providing important support for subsequent differentiated demand response.

[0124] Because there are significant differences in the stability and adjustment capabilities of different resident users' behaviors, the system further constructs a dynamic weight mapping mechanism to map user behavior characteristics to behavior fit weights and adjustment effort weights in the optimization objective. Specifically, the more stable the behavior, the higher the behavior fit weight; the greater the fluctuation, the higher the adjustment effort cost. This mechanism enables differentiated adjustment strategies for different users, thus avoiding excessive deviations in user behavior caused by using a uniform optimization approach. Experimental results show that dynamic weights can effectively reflect differences in user behavior, ensuring that the optimization results both meet the regional load adjustment objective and maintain the original behavioral patterns of users, thereby improving the acceptability of user participation in demand response.

[0125] according to Figure 6 The figure shows the changes in the user baseline load curve before and after demand response optimization. The baseline curve represents the user's original baseline load, the optimized curve represents the target load after dynamic optimization, and the response curve represents the user's response under the incentive mechanism. As can be seen from the figure, the optimized load curve is significantly lower than the original baseline curve during peak hours, while the actual response curve generally follows the optimization target well, indicating that the dynamic optimization mechanism can effectively regulate the residential load. This result demonstrates that the system has successfully completed key steps such as baseline generation, dynamic weight mapping, multi-objective optimization, and user response feedback, verifying the feasibility of the dynamic demand response control process in the patented solution.

[0126] After completing the user profile and dynamic weight construction, the system further establishes a multi-objective optimization model. The optimization objectives mainly include: regional total load peak shaving objective; user behavior alignment objective; and user adjustment effort objective. The regional objective is used to achieve the overall peak shaving demand of the power grid, while the user objective is used to ensure that the optimized curve maintains the original user behavior characteristics and avoids unreasonable load distortion. The program uses a limited-memory Broyden-Fletcher-Goldfarb-Shanno with Bounds (L-BFGS-B) algorithm with boundary constraints for optimization, and sets upper and lower load bound constraints to ensure that the optimization results meet physical feasibility.

[0127] according to Figure 7The figure shows the overall change in regional total load before and after demand response optimization. The blue curve represents the original regional total load, the orange curve represents the optimized regional total load, and the green dashed line represents the target load curve set by the system. The results show that during both morning and evening peak hours, the optimized regional total load is significantly lower than the original load curve, the overall curve is smoother, and it gradually approaches the target load level. This indicates that the constructed dynamic demand response mechanism can effectively achieve regional-level peak shaving and valley filling, reduce the pressure on the power grid during peak hours, and improve the rationality of load allocation. These results further verify the good engineering feasibility and application value of the patented method in real residential electricity consumption data scenarios.

[0128] Experimental results show that the optimized user load curve achieves the regional load shaving target while maintaining the original behavioral patterns. Furthermore, the optimized curve is relatively smooth, without drastic oscillations or unreasonable abrupt changes, indicating that the optimization model has good stability and feasibility. To further enhance user participation in demand response, the system dynamically generates personalized price curves based on user behavior characteristics. Incentive prices are increased during evening peak hours to guide users to proactively reduce peak load. Because different users have varying behavioral stability and adjustment capabilities, the dynamically generated price levels will also differ.

[0129] Experimental results show that dynamic pricing mechanisms can reflect individual user differences and form targeted demand response incentive strategies, thereby improving user participation and response efficiency.

[0130] To verify the overall stability of the system, the program further simulated user response behavior and performed closed-loop feedback analysis on the optimization results. The system used Mean Absolute Error (MAE) to evaluate the deviation between the optimized curve and the actual user response curve. Experimental results show that the error between the optimized curve and the response curve is small, indicating that the constructed personalized load baseline can adapt well to changes in residents' actual behavior. Simultaneously, the system can continuously correct the optimization results through closed-loop feedback, improving the accuracy of subsequent adjustments. Therefore, this scheme not only achieves static load optimization but also possesses continuous dynamic adjustment capabilities, demonstrating high practical engineering application value.

[0131] This embodiment, based on real residential load data, fully implements the entire process of user behavior identification, cluster analysis, user profile construction, multi-objective optimization, dynamic incentives, and closed-loop feedback described in the patent. The effectiveness of the method is verified using real data. Experimental results show that this method can effectively identify residential user behavior characteristics, achieve personalized load baseline optimization, and maintain user behavior consistency while meeting regional load regulation targets, demonstrating significant engineering application value and potential for wider adoption.

[0132] This embodiment includes a multi-dimensional electricity consumption feature extraction and fusion clustering method for load-side users. This includes a specific method based on dynamic time warping distance and adaptive DBSCAN clustering algorithm to extract morphological feature vectors such as the principal typical curve, morphological richness, main cluster concentration, and noise ratio from the user's daily load curve. It also includes a specific method based on the user's daily electricity consumption sequence and principal typical curve to calculate time-series regularity feature vectors such as daily electricity consumption autocorrelation (lag of 1 day and 7 days), load curve consistency (mean and coefficient of variation), and weekday-weekend differences. Furthermore, it includes a method that standardizes and concatenates morphological and regularity features, and then uses a Gaussian mixture model that incorporates user registration type labels (industrial / commercial / residential) as soft constraints for user grouping. The core of this method lies in the clustering mechanism that combines "data-driven feature similarity" and "domain knowledge goal consistency" embodied in the constrained clustering formula.

[0133] In this embodiment, the personalized decomposition method for node load profiles based on feature clustering results constructs a personalized load profile optimization model with the objectives of regional adjustment, profile shape conforming to typical user curves, and minimizing user adjustment efforts. The formulas and methods for dynamically mapping user features to the optimization model weights are used. Based on the aforementioned clustering results, a two-layer solution strategy of "inter-cluster coordination - intra-cluster optimization" is adopted to decompose the large-scale optimization problem into specific implementation processes and methods for cluster-level task allocation and intra-cluster parallel optimization.

[0134] In this embodiment, the personalized incentive mechanism design method that integrates user target type and behavioral characteristics quantifies users' electricity consumption behavior characteristics into key parameters of the user utility function through mapping functions (such as linear weighting and nonlinear combination). For different types of users, such as industrial (cost-driven), commercial (comfort-constrained), and residential (comprehensive trade-off), a differentiated utility function model integrating economic incentives, adjustment costs, comfort loss, and green contributions is constructed. Simultaneously, a dynamic incentive price model and method are developed based on user characteristic clusters, including a base price, characteristic adjustment terms, and potential incentive terms.

[0135] Example 2 according to Figure 8 As shown, this embodiment also provides a dynamic baseline load design system based on multi-objective collaboration and feature clustering, including: Feature profiling module 1 is used to acquire users' historical electricity consumption data and type labels, extract morphological and regular features, and generate a user feature profile library through feature fusion and target-guided constraint clustering. The guideline decomposition module 2 is used to receive the regional load adjustment target issued by the power grid, and combine it with the user feature profile library to obtain a personalized node load guideline that fits the user behavior characteristics through multi-objective optimization model decomposition. Incentive customization module 3 is used to construct differentiated user utility functions based on user feature profiles and the real-time status of the power grid, and to formulate dynamic incentive prices for different user clusters. The closed-loop optimization module 4 is used to send the personalized node load baseline and dynamic incentive price to users, collect actual user response data, update the user feature profile library, and form a closed-loop optimization process.

[0136] In summary, the dynamic baseline load design method and system based on multi-objective collaboration and feature clustering provided in this embodiment significantly improves user participation willingness and response accuracy. This invention accurately identifies users' inherent electricity consumption habits and intrinsic decision-making goals through offline feature clustering. When generating personalized load baselines, dynamic weight mapping ensures the baselines proactively align with the habits of stable users and reduces the burden on users with low adjustment willingness. Simultaneously, the incentive mechanism quantifies users' own goals and behavioral characteristics into utility function parameters, making incentives tailored to each user. Therefore, the guidance signal (baseline + incentive) received by the user is highly coordinated with their own habits and interests, fundamentally transforming response behavior into proactive optimization, thereby inevitably stimulating higher participation willingness and achieving higher response accuracy and comfort due to the clear and easily executable goals.

[0137] Secondly, this invention deeply explores and refines the heterogeneous load-side adjustment potential: traditional methods treat users as homogeneous objects, failing to differentiate the quality of their adjustability potential. This invention innovates on the predictability and stability of user behavior. In guideline decomposition and incentive pricing, the system, based on these characteristics, assigns more refined adjustment tasks tailored to the habits of users with stable and highly predictable behavior, and provides a reliability premium incentive. This allows for more full and reliable utilization of massive heterogeneous load resources without significantly increasing user discomfort, achieving a dual improvement in both the "quality" and "quantity" of adjustment potential.

[0138] Furthermore, it enhances the overall economy, security, and renewable energy absorption capacity of the power grid: First, personalized guidelines and high-precision responses enable load adjustments to more accurately match the spatiotemporal distribution and output fluctuations of renewable energy, thereby directly reducing wind and solar curtailment and improving clean energy absorption. Second, a hierarchical decomposition strategy based on feature clustering and network constraints considers user-level dynamic power constraints and cluster-level power flow security at the source of guideline generation, proactively avoiding line overload risks caused by unreasonable load allocation and enhancing the safety of power grid operation. Finally, differentiated and precise incentives reduce ineffective incentive expenditures, while improved incentive and response accuracy lowers the overall compensation cost required to achieve the same regulation objective. This systematically improves the overall economy and security of power grid operation.

[0139] Finally, the system's collaborative efficiency, fairness, and scalability are improved: At the game theory level, because the user utility model more closely reflects their actual decision-making logic, the pricing strategy of the Distribution System Operator (DSO) can more effectively guide expected behavior, reducing repeated games and strategy conflicts caused by information mismatch, and promoting faster convergence of the Stackelberg game to an efficient equilibrium point. Regarding fairness, data-driven clustering and the constraint of similar relative adjustment rates within clusters ensure that users with similar characteristics and goals bear similar relative obligations, achieving fairness within the group. In engineering implementation, the hierarchical distributed architecture based on feature clustering decomposes the massive user optimization problem into smaller, more manageable components, significantly reducing computational and communication complexity. Simultaneously, the framework is highly inclusive of new users, requiring only their inclusion in existing feature clusters or the formation of new clusters to integrate them into the system, providing a scalable and lightweight technical path to support routine deployments with tens of millions of users.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A dynamic baseline load design method based on multi-objective collaboration and feature clustering, characterized in that, include: Obtain users' historical electricity consumption data and type tags, extract morphological and regular features, and generate a user feature profile library through feature fusion and target-guided constraint clustering. The system receives regional load regulation targets from the power grid and, in conjunction with the user feature profile database, obtains personalized node load profiles that fit user behavior characteristics through multi-objective optimization model decomposition. Based on the user feature profile database and the real-time status of the power grid, we construct differentiated user utility functions and formulate dynamic incentive prices for different user clusters. The personalized node load baseline and dynamic incentive price are distributed to users, user actual response data are collected, and the user feature profile database is updated to form a closed-loop optimization process.

2. The dynamic baseline load design method based on multi-objective collaboration and feature clustering according to claim 1, characterized in that, The specific process of obtaining users' historical electricity consumption data and type tags, extracting morphological and regular features, and generating a user feature profile library through feature fusion and target-guided constraint clustering includes: The daily load curves of users are subjected to density-based noisy spatial clustering algorithm for morphological feature extraction, abnormal electricity consumption days are eliminated, and a morphological feature vector containing the main typical curve, morphological richness, main cluster concentration and noise ratio is constructed. Regularity features of user daily electricity consumption sequences and main typical curves are quantified to construct regularity feature vectors that include autocorrelation coefficients, curve consistency, and weekday-weekend differences. The morphological feature vector and the regularity feature vector are standardized and then fused. The GMM algorithm with a target consistency penalty term is used for clustering to generate user feature profiles.

3. The dynamic baseline load design method based on multi-objective collaboration and feature clustering according to claim 2, characterized in that, The parameter adaptation process of the noisy spatial clustering algorithm includes: Calculate the dynamic time-normalized distance between load curves, plot the k-distance curve and take the inflection point value as the neighborhood radius, and set the minimum number of points to K+1, where K is the preset number of nearest neighbors.

4. The dynamic baseline load design method based on multi-objective collaboration and feature clustering according to claim 2, characterized in that, In the GMM algorithm with a target consistency penalty term, the calculation of user posterior probability introduces a target type consistency indicator function to guide users with similar behavioral characteristics and consistent target types to be grouped into the same cluster.

5. The dynamic baseline load design method based on multi-objective collaboration and feature clustering according to claim 1, characterized in that, The objective function of the multi-objective optimization model includes a global target tracking term, a shape fitting term, and a regulation effort term, wherein the shape fitting weight and the regulation effort weight are determined based on the dynamic mapping of user feature profiles.

6. The dynamic baseline load design method based on multi-objective collaboration and feature clustering according to claim 5, characterized in that, The constraints of the multi-objective optimization model include global objective constraints, user power limit constraints, adjustment rate constraints, dynamic power limit constraints based on adjustable potential, user-level comfort constraints, and relative fairness constraints within the cluster.

7. The dynamic baseline load design method based on multi-objective collaboration and feature clustering according to claim 1, characterized in that, The personalized node load profile that matches user behavior characteristics adopts a distributed-centralized hybrid solution strategy, and the specific process is as follows: Preliminary guidelines are obtained by pre-decomposing the target region based on user characteristics and solving small-scale optimization problems within the cluster. Summarize the initial cluster plan, check the power flow constraints and target deviations, fine-tune the cluster target and rebalance to obtain a personalized node load profile that fits the user's behavior characteristics.

8. The dynamic baseline load design method based on multi-objective collaboration and feature clustering according to claim 1, characterized in that, The differentiated user utility functions include a cost-driven utility function for industrial users, a comfort-constrained utility function for commercial users, and a comprehensive trade-off utility function for residential users. The parameters of the cost-driven utility function for industrial users, the comfort-constrained utility function for commercial users, and the comprehensive trade-off utility function for residential users are dynamically determined based on user feature profiles.

9. The dynamic baseline load design method based on multi-objective collaboration and feature clustering according to claim 1, characterized in that, The dynamic incentive price consists of a base price, a feature adjustment item, and a potential incentive item, wherein the dynamic incentive price satisfies the total incentive budget constraint.

10. A dynamic baseline load design system based on multi-objective collaboration and feature clustering, characterized in that, include: The feature profiling module is used to obtain users' historical electricity consumption data and type labels, extract morphological and regular features, and generate a user feature profile library through feature fusion and target-guided constraint clustering. The guideline decomposition module is used to receive the regional load regulation target issued by the power grid, and combine it with the user feature profile library to obtain personalized node load guidelines that fit the user behavior characteristics through multi-objective optimization model decomposition. The incentive customization module is used to construct differentiated user utility functions based on the user feature profile library and the real-time status of the power grid, and to formulate dynamic incentive prices for different user clusters. The closed-loop optimization module is used to distribute the personalized node load baseline and dynamic incentive price to users, collect actual user response data, update the user feature profile database, and form a closed-loop optimization process.

Images