Load clustering method based on load adjustable potential

Abstract

The invention discloses a load clustering method based on load adjustable potential, and particularly relates to the technical field of power regulation and control, and the method comprises the steps: collecting the behavior and resource information of a load user, constructing feature vectors, carrying out the bi-clustering analysis, and extracting a user intersection with the behavior and resource dual matching; high-regulation potential users with high response reliability are screened out after identification; and combining the load pressure, resource saturation and historical scheduling feedback of the power grid, dynamically adjusting clustering parameters by using the trained machine learning model, optimizing the classification boundary and precision, and realizing accurate identification of a load adjustment user and stable support of a power grid scheduling strategy. According to the method, synchronous recognition of user response behaviors and resource controllability is achieved through two-dimensional clustering and intersection screening, the neural network is introduced to achieve dynamic self-adaptive adjustment of clustering parameters, the user screening reliability is improved in combination with a response confidence verification mechanism, and therefore the accuracy, practicability and stability of load adjustment resources are remarkably enhanced.

Description

Technical Field

[0001] This invention belongs to the field of power control technology, and specifically relates to a load clustering method based on load adjustability potential. Background Technology

[0002] With the continuous construction of new power systems and the integration of a high proportion of renewable energy, the power grid operation places higher demands on flexible and controllable load resources. As an important demand-side regulation method, user-side load response has become an important means to support the stable operation of the power system, alleviate dispatch pressure, and improve regulation capabilities, especially in critical scenarios such as regional extreme weather and high load peaks, where its role is particularly prominent.

[0003] For example, during periods of continuous high temperatures and concentrated industrial loads in summer, the dispatching system needs to quickly and accurately identify "high-potential users" from a massive pool of users who are willing to participate in the response and have the resource execution capabilities within a very short time. This is to avoid voltage fluctuations, frequency anomalies, or even power outages caused by response failures or misaligned control. However, in current practical applications, existing user clustering and identification methods are mostly based on static load data, single behavioral labels, or long-term historical average models, which are insufficient to characterize the dynamic response capabilities of users under sudden, high-frequency dispatching conditions.

[0004] Common methods often overlook the dynamic evolution of user behavior characteristics over time, environment, and temperature control strategies, lacking systematic modeling of behavioral volatility and response consistency. This leads to problems such as interrupted responses, delays, and large fluctuations in actual execution for some seemingly "responsive" users. Secondly, scheduling strategies often use "response willingness" as the primary clustering criterion, without effectively combining the access capabilities and resource status of user-side adjustable devices. This results in some users with aging equipment, unconnected interfaces, or limited energy storage being incorrectly classified into high-response groups, creating a disconnect between the "behavioral map" and the "resource map." Thirdly, most existing clustering methods use fixed parameter settings and do not incorporate real-time sensing mechanisms of grid operation status. They cannot dynamically adjust the clustering model based on scheduling pressure levels, resource saturation density, or past response performance, limiting the adaptability of control strategies. Furthermore, if users entering the adjustable set lack quantitative verification of their past scheduling performance's response success rate and execution completion rate, it can easily lead to serious consequences such as scheduling misinterpretation and strategy failure. Based on this, this invention proposes a load clustering method based on load adjustable potential to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a load clustering method based on load adjustability potential. Through a dual-dimensional feature clustering mechanism of behavior and resources, a response confidence verification mechanism, and an adaptive optimization strategy for clustering parameters, a dynamic clustering identification scheme with system awareness, verifiability of execution, and adaptability of strategy is constructed. This method is applicable to regional power grids during periods of high temperature and power shortage, during periods of sudden industrial load increases, or in areas with significant fluctuations in distributed power sources. It ensures that the identified users have high confidence, high response, and high controllability, thereby improving the ability of scheduling strategies to be implemented under extreme conditions.

[0006] Specifically, this invention relates to a load clustering method based on load adjustability potential, the specific steps of which are as follows:

[0007] Step S1: Collect the historical response behavior of various load users and the deployment information of their controllable power resources, and construct behavioral feature vectors and resource feature vectors respectively to represent the user's adjustment capabilities and behavioral patterns in multiple dimensions;

[0008] Step S2: Based on the behavioral feature vector, the first clustering algorithm with cluster radius as the core of control is used to cluster users in the behavioral dimension; based on the resource feature vector, the second clustering algorithm with density as the core discrimination criterion is used to cluster users in the resource capability dimension.

[0009] Step S3: Extract the intersection region of the first clustering result and the second clustering result, and identify the user groups identified in the intersection to obtain a set of loads with high adjustment potential, thereby screening out user resources with high response reliability under extreme high load scenarios;

[0010] Step S4: Based on the current load pressure level in the power grid operation, the saturation level of regional regulation resources, and the feedback effect of the previous round of load response scheduling, the key parameters of the two clustering algorithms are dynamically adjusted using a pre-trained machine learning model to achieve continuous optimization of clustering boundaries and classification accuracy, thereby improving the accuracy of load regulation and the stability of power grid scheduling.

[0011] In a preferred embodiment, the behavioral feature vector constructed in step S1 includes the following feature dimensions: load response start-up delay, response duration, unit adjustment load, historical response frequency and user response willingness level, and all behavioral feature data form a time series matrix by setting a sampling period, which serves as the basis for subsequent clustering input.

[0012] In a preferred embodiment, the resource feature vector constructed in step S1 includes the type of power adjustable device accessed by the load user, response interface compatibility, device availability status, maximum adjustable capacity and availability duration, and all resource feature items are input into the second clustering algorithm as a structured parameter matrix.

[0013] The construction of resource feature vectors also includes equipment historical failure rate and maintenance cycle index, which are then input into the second clustering algorithm after numerical mapping and standardization.

[0014] In a preferred embodiment, the first clustering algorithm for behavior dimension clustering in step S2 adopts the K-Medoids algorithm based on radius threshold. By using preset initial cluster centers and radius boundaries, the behavior feature vectors are classified by similarity to form multiple clusters.

[0015] In a preferred embodiment, the second clustering algorithm used for resource capacity dimension clustering in step S2 adopts the DBSCAN algorithm based on the density aggregation principle. This algorithm is based on the Euclidean distance between users' resource feature space, defines the neighborhood search radius and the minimum number of neighborhood samples as control variables, and forms high-density clusters for users with high density distribution characteristics of controllable power resources; low-density points outside the clusters are regarded as uncontrollable load users and are excluded in the subsequent intersection extraction step.

[0016] In a preferred embodiment, the intersection extraction operation in step S3 is completed by vector space position matching, that is, the user set output by the first clustering algorithm and the user set output by the second clustering algorithm are mapped in parallel, an alignment matrix is ​​established with user identifier as the dimension, and user groups that belong to both clusters are filtered out. The intersection user set has the characteristics of active behavior response and physical availability of resource devices, and is defined as the set to be filtered.

[0017] In a preferred embodiment, the machine learning model in step S4 is constructed based on an artificial neural network, using the power grid load pressure level, regional equipment saturation, and historical dispatch feedback effectiveness as input features, and outputting the optimal first clustering algorithm radius parameter and the second clustering algorithm neighborhood density control parameter.

[0018] In a preferred embodiment, the power grid load pressure level in step S4 is calculated based on the ratio between the real-time load and the historical peak load of the target area, and is divided into four levels. The signal is input to the learning model. The saturation of regional equipment is calculated by the ratio of the number of connected regulating devices per unit area or unit load capacity to the theoretical maximum configuration capacity. The historical dispatch feedback effectiveness is obtained by weighting the average response execution rate, average instruction arrival rate and average load change success rate of the user group in the previous multiple response dispatches, forming a comprehensive score of multi-dimensional feedback indicators.

[0019] In a preferred embodiment, the set to be screened needs to be verified for reliability. A response confidence threshold is set for the users in the intersection, i.e. the set to be screened. Only when the user confidence exceeds the preset standard is it considered as an adjustable user resource with high response reliability and selected into the set of loads with high adjustment potential.

[0020] In a preferred embodiment, reliability verification refers to:

[0021] The ratio of the number of successful responses to all scheduled tasks within a set period to the total number of tasks is obtained. If the ratio reaches or exceeds the preset response confidence threshold, the user is considered to have high response reliability and is included in the high adjustment potential load set; otherwise, the user is not included.

[0022] Compared with existing technologies, the beneficial effects are:

[0023] (1) This invention constructs a two-dimensional clustering structure to model the user's load response behavior characteristics and the characteristics of their access to adjustable resources, and uses the radius-based K-Medoids algorithm and the density-based DBSCAN algorithm for cluster analysis respectively. Through the intersection extraction mechanism, users with high consistency in both behavioral response and resource configuration are selected, ensuring that the identified users have both a positive willingness to participate and good behavioral response performance, as well as practically usable adjustment devices and physical execution capabilities. This effectively avoids the misjudgment problem of "positive behavior but uncontrollable resources", and significantly improves the reliability and scheduling accuracy of load resource call.

[0024] (2) Based on the current operating status of the power grid and historical dispatch feedback, this invention introduces an artificial neural network as a clustering parameter optimization model. It uses core indicators such as power grid load pressure level, regional equipment saturation, and historical dispatch response effectiveness as input features, and adaptively outputs the radius parameter for behavioral clustering and the neighborhood control parameter for resource clustering. Through this dynamic optimization mechanism, the clustering algorithm can flexibly adjust the screening boundary according to the actual system environment, possessing stronger state awareness and strategy flexibility. This avoids the problem of clustering distortion or screening imbalance in high-voltage scenarios caused by traditional fixed-parameter methods, thus improving the stability and real-time performance of the clustering model.

[0025] (3) This invention further introduces a response confidence verification mechanism after the intersection extraction of the two clusters, and performs statistical analysis on the actual execution records of each candidate user participating in scheduling within a set period. By calculating the ratio of the number of successful responses to the total number of scheduling attempts, and combining variables such as grid load pressure, equipment access saturation, and historical execution effects, a response confidence threshold is dynamically set. Only when the user's actual performance reaches this dynamic threshold can it be included in the final high-regulation-potential load set. This mechanism effectively shields user resources with unstable responses or distorted behavior, ensuring the credibility and controllability of the regulation resource pool, and providing a high-quality, low-failure-rate load dispatching basis for the power grid system under high-voltage, high-frequency, and high-risk scheduling scenarios. Attached Figure Description

[0026] Figure 1This is a flowchart illustrating a load clustering method based on load adjustability potential according to the present invention. Detailed Implementation

[0027] The following detailed description, in conjunction with the accompanying drawings, illustrates a specific implementation of the load clustering method based on load adjustability potential according to the present invention.

[0028] Reference Figure 1 The following examples were obtained:

[0029] Example 1:

[0030] The load clustering method based on load adjustability potential includes the following steps:

[0031] Step S1: Collect the historical response behavior of various load users and the deployment information of their controllable power resources, and construct behavioral feature vectors and resource feature vectors respectively to represent the user's adjustment capabilities and behavioral patterns in multiple dimensions;

[0032] Step S2: Based on the behavioral feature vector, the first clustering algorithm with cluster radius as the core of control is used to cluster users in the behavioral dimension; based on the resource feature vector, the second clustering algorithm with density as the core discrimination criterion is used to cluster users in the resource capability dimension.

[0033] Step S3: Extract the intersection region between the first and second clustering results, and identify the user groups in the intersection to obtain a high-adjustment-potential load set. This is used to screen user resources with high response reliability under extreme high-load scenarios. It should be noted that in this invention, multiple clusters are generated during both behavioral feature clustering and resource capability clustering. By identifying target clusters that respectively exhibit high consistency in response behavior and high controllable resource density as the screening basis, only users belonging to both target clusters are extracted for intersection. Intersection users must simultaneously meet two conditions: first, they belong to clusters with short response latency, stable behavior, and high response frequency in the behavioral clustering; second, they belong to clusters with stable device status, complete response interfaces, and strong adjustment capabilities in the resource clustering. By pre-screening and limiting the characteristics of the two clusters, it is ensured that the final intersection users not only have positive behavior and good response, but also have actual physical response capabilities. This excludes users who only perform well in some dimensions but lack response feasibility, avoiding situations where the intersection results include "non-responsive" or "uncontrollable" users.

[0034] Step S4: Based on the current load pressure level in the power grid operation, the saturation level of regional regulation resources, and the feedback effect of the previous round of load response scheduling, the key parameters of the two clustering algorithms are dynamically adjusted using a pre-trained machine learning model to achieve continuous optimization of clustering boundaries and classification accuracy, thereby improving the accuracy of load regulation and the stability of power grid scheduling.

[0035] The behavioral feature vector constructed in step S1 includes the following feature dimensions: load response initiation delay, response duration, unit adjustment load, historical response frequency, and user response willingness level. All behavioral feature data are used to form a time series matrix through a set sampling period, serving as the basis for subsequent clustering input. This behavioral feature vector is updated periodically based on a sliding window mechanism to ensure that the clustering model can reflect dynamic changes in user behavior in real time, thereby enhancing the accuracy and foresight in selecting users with high adjustment potential.

[0036] In identifying and classifying load users' regulation capabilities, relying solely on static behavioral data at a single moment cannot accurately reflect the users' response characteristics during actual operation. Especially given the increasing frequency of electricity demand response and the accelerating pace of regulation, users' response capabilities exhibit significant dynamic changes. Therefore, it is necessary to introduce a feature expression mechanism that can continuously reflect user behavior patterns.

[0037] By constructing a behavioral feature vector using load response initiation delay, response duration, unit load adjustment, historical response frequency, and user response willingness level as feature dimensions, this approach comprehensively characterizes user behavior from multiple aspects, including response efficiency, stability, flexibility, and participation enthusiasm. Furthermore, by setting a fixed sampling period, these features are constructed into a matrix structure in a time-series manner. This not only preserves the changing trends of behavioral indicators but also possesses good data organization, facilitating the subsequent clustering model's overall learning of time-series behavioral features. A sliding window mechanism is introduced to periodically update the behavioral feature vector, enabling the model to perceive real-time changes in user behavior, promptly eliminate abnormal response data, and dynamically correct user behavior profiles. This mechanism effectively improves the real-time performance and adaptability of behavioral clustering, thereby significantly enhancing the accuracy and foresight in identifying users with high adjustment potential and providing a more stable and reliable behavioral data foundation for power grid control.

[0038] The resource feature vector constructed in step S1 includes the type of power adjustable device connected to the load user, response interface compatibility, device availability status, maximum adjustable capacity, and available duration. All resource feature items are input into the second clustering algorithm as a structured parameter matrix. The construction of the resource feature vector also includes the historical failure rate of the equipment and maintenance cycle indicators. After numerical mapping and standardization, these are input into the second clustering algorithm to improve the depth and breadth of the resource clustering dimension, so that the selected high-adjustment-potential users not only have responsive behavior, but also stable execution capabilities.

[0039] In power load response management, relying solely on a user's response behavior information is far from sufficient to determine whether a user possesses controllability and regulation capabilities. The actual executability of the response capability largely depends on the physical execution capability, interface compatibility, and stability of the power regulation resources connected to the user. Therefore, to comprehensively and accurately assess a user's resource regulation capabilities, it is necessary to construct a resource feature vector covering multiple levels and introduce multi-dimensional parameters for cluster analysis. The construction of this resource feature vector not only considers the type of power regulation devices connected to the user but also incorporates compatibility information regarding whether the response interface can be recognized and controlled by the dispatch system. Furthermore, by combining key indicators such as the current operational availability status of the devices, maximum adjustable capacity, and duration of sustainable regulation, a complete picture of the user's adaptability and flexibility to dispatch tasks at the hardware resource level can be obtained.

[0040] Considering that some equipment may have hidden instability factors, such as frequent failures or excessive maintenance frequency, two indicators—historical equipment failure rate and maintenance cycle—were introduced into the resource feature construction. These two indicators were transformed into quantifiable data through numerical mapping and then standardized to ensure a unified computational dimension when combined with other structured features. This enhanced resource feature vector significantly improves the discrimination depth and classification accuracy of the second clustering algorithm for adjustable resources, ensuring that the ultimately selected high-potential users not only have strong responsiveness but also possess high stability and reliability at the equipment level, providing solid support for the dynamic control of the power system.

[0041] In step S2, the first clustering algorithm used for behavioral dimension clustering adopts the K-Medoids algorithm based on radius threshold. By using the preset initial cluster centers and radius boundaries, the behavioral feature vectors are classified by similarity to form multiple clusters. Users within each cluster have high consistency in response latency, duration and response stability indicators.

[0042] The set of user behavior feature vectors is denoted as X = {x1, x2, ..., x3}. n};x i This represents the behavioral feature vector of the i-th load user, belonging to R. d ;R d Let d be a real vector space of dimension d, where n∈N + , representing the total number of samples, which belongs to the set of positive integers; d∈N + , representing the dimension of each feature vector, is usually determined by the number of behavioral indicators; the number of clusters is K∈N. + The number of clusters is preset for the clustering process; the radius threshold is R∈R + , is a positive real number used to control the maximum allowable distance between cluster members and the center point; the set of center points is M = {m1, m2, ..., m}.K}, m k ∈X.

[0043] The first clustering algorithm execution steps include:

[0044] Step 211: Initialize the center points. Randomly select K samples from the sample set X as initial center points to form the initial center point set M.

[0045] Step 212: Assign samples to the center point. For any sample x i ∈X, calculate its relationship with each center point m k Euclidean distance ∈ M:

[0046]

[0047] x ij Indicates sample x i The value of m on the j-th behavioral feature dimension; kj Indicates the center point m k The value of D in the j-th feature dimension; ik ∈R + Indicates sample x i With center point m k The Euclidean distance. D ik When ≤R, it is considered as sample x i Can be attributed to the center point m k The cluster represented. If a sample satisfies the distance condition of multiple centroids, the centroid with the smallest distance is selected for classification; if none of the conditions are met, the sample is temporarily not assigned to any cluster and will be processed later.

[0048] Step 213: Update the centroid for each current cluster. Select a new center point from the sample with the smallest average intra-cluster distance.

[0049] From each cluster, select the sample point that minimizes the total distance between it and all samples in the same cluster, and use it as the center of the new round.

[0050] Step 214: Termination condition judgment. If the change of all center points compared with the previous round center points is less than the preset threshold, or the center points no longer change, then terminate the iteration; otherwise, return to step 212 and continue to perform new allocation and update.

[0051] Step 215: Output the clustering results, ultimately outputting the set of clusters {C1, C2, ..., C}. K Users within each cluster exhibit high consistency across the following behavioral characteristics: such as load response initiation delay, response duration, and stability of unit load variation.

[0052] Based on the radius constraint mechanism, the K-Medoids algorithm, while retaining the advantage of cluster centers being real sample points, further limits the cluster boundaries, thereby controlling the compactness and consistency of sample partitioning. For situations where user response behaviors exhibit significant differences or local anomalies, this algorithm can effectively eliminate noise influences, ensuring that the clustering results are more physically meaningful and practical for scheduling, laying a structured foundation for subsequent intersection filtering and identification of users with high adjustment potential.

[0053] In step S2, the second clustering algorithm used for clustering based on the resource capability dimension adopts the DBSCAN algorithm based on the density aggregation principle. This algorithm is based on the Euclidean distance between users' resource feature spaces, and defines the neighborhood search radius and the minimum number of neighborhood samples as control variables. Users with high density distribution characteristics of controllable power resources are formed into high-density clusters. Low-density points outside the clusters are regarded as uncontrollable load users and are excluded in the subsequent intersection extraction step, thereby improving the actual controllability and physical response feasibility of the final selected high-regulation-potential load set.

[0054] The user resource feature set is defined as: Z = {z (a) |a∈N + , 1≤a≤A};z (a) ∈R u Let R represent the resource feature vector of the a-th load user; u Describe a real vector space of dimension u; A∈N + This represents the total number of users in the resource feature set; each resource vector contains several dimensions of information, including the number of adjustable device types, capacity limit, interface availability, sustainable adjustment duration, and device failure rate. The neighborhood search radius is defined as ε∈R. + , representing the maximum allowed similarity distance in the resource feature space; the minimum number of neighborhood samples is defined as μ∈N + , representing the minimum number of neighbors required to form a dense cluster; the distance function uses a Euclidean metric and is defined as:

[0055] Represents the resource feature vector z (a) The component on the v-th feature dimension; z (b) ∈Z represents the resource vector of any other user in the set; δ(z) (a) ,z (b) )∈R + This represents the distance between two users in the resource feature space.

[0056] The second clustering algorithm executes the following steps:

[0057] Step S221: Neighborhood definition, for any user z (a)∈Z, define its

[0058] The ε-neighborhood is: That is, in the set and

[0059] z (a) All users within a distance of ε are considered its neighbors.

[0060] Step S222: Point type classification, if Then z (a) Marked as a core point; if However, there exists a core point z. (c) Make Then z (a) Marked as a boundary point; if z (a) If any of the above conditions are not met, the user is considered a noise point or a low-density user.

[0061] Step S223: Cluster generation. Starting from any unmarked core point, cluster all directly density-reachable users into a cluster along the density connection path until all core points belong to a cluster or are marked as boundary points. Each ultimately generated cluster... m∈N + Indicates the cluster number.

[0062] Step S224: Output the clustering results, outputting the set of all clusters {Q} (1) Q (2) Q (M) Each cluster represents a group of users who are highly similar in terms of resource capabilities. All users who are not assigned to any cluster (noise points) are considered individuals with insufficient resource controllability and are removed in the subsequent intersection filtering.

[0063] Compared to traditional centroid-based clustering algorithms, DBSCAN is more suitable for data spaces with uneven resource distribution or non-convex structures. Through dual control of the density parameter μ and the distance threshold ε, it can effectively identify key user sets that possess both actual resource adjustment capabilities and high access density. In scenarios involving adjustable power resources, this algorithm can automatically eliminate isolated users and uncontrollable load nodes, retaining only core resource clusters with structural controllability and physical response potential. This provides data assurance and a structural foundation for the accurate construction of subsequent load sets with high adjustment potential.

[0064] The intersection extraction operation in step S3 is completed through vector space position matching. That is, the user set output by the first clustering algorithm and the user set output by the second clustering algorithm are mapped in parallel. An alignment matrix is ​​established with user identifier as the dimension. User groups that belong to both clusters are selected. The intersection user set has the characteristics of active behavior response and physical availability of resource devices, and is defined as the set to be screened. The intersection user set has the characteristics of active behavior response and physical availability of resource devices, and therefore has high practical control value. After subsequent verification, it is given priority in the response scheduling strategy to improve the overall operational stability of the power regulation system.

[0065] The output of behavioral feature clustering is defined as follows:

[0066] This represents the p-th user belonging to the behavior clustering result; P represents the number of users in the behavior cluster.

[0067] The output of resource feature clustering is defined as follows:

[0068] Let q represent the q-th user belonging to the resource cluster result; Q represents the number of users in the resource cluster.

[0069] All users have a globally unique identifier ID(u) used to establish mapping relationships. User feature vectors do not directly participate in intersection determination; intersection is completed based on the identifier set.

[0070] Establish a two-dimensional alignment matrix for user identifiers:

[0071]

[0072] The matrix A, of size P×Q, represents the identifier matching between users in the behavior cluster and users in the resource cluster. The intersection set of users is then defined as follows:

[0073]

[0074] That is, selecting all users who exist in both clustering results and have the same identifier to form the intersection user set. Users in the intersection user set exhibit good response willingness and behavioral consistency in behavioral clustering, and possess high-density resource allocation adjustment and physical access capabilities in resource clustering. Because members of this set possess "dual feature consistency," they demonstrate high success rate, stable response behavior, and strong resource controllability in subsequent scheduling execution. Therefore, this set is defined as the "selection set," providing a candidate basis for the next step of "response confidence verification."

[0075] The machine learning model in step S4 is built based on an artificial neural network. It takes the power grid load pressure level, regional equipment saturation and historical dispatch feedback effectiveness as input features, and outputs the optimal first clustering algorithm radius parameter and the neighborhood density control parameter of the second clustering algorithm.

[0076] This neural network model is a typical multi-layer feedforward neural network, consisting of an input layer, several hidden layers, and an output layer. The input layer receives the current state variables of the system, including: the grid load pressure level, regional equipment saturation, and historical dispatch feedback effectiveness. These input features together constitute the input vector of the neural network, which describes the current dispatch urgency, control resource base, and recent performance of the system.

[0077] The output layer is used to predict the optimal control parameters required for the two clustering models, specifically including:

[0078] The first clustering algorithm (behavioral dimension) outputs the following parameters: Radius parameter: This controls the maximum similarity boundary that determines whether sample points can be assigned to the same cluster. The second clustering algorithm (resource dimension) outputs the following parameters: Neighborhood density control parameters, including: Neighborhood search radius: Represents the maximum nearest neighbor distance of a sample in the resource feature space; Minimum number of neighboring samples: Represents the minimum number of samples required to form a high-density region. The network output vector consists of these three core parameters: the radius of the behavioral cluster, the neighborhood search radius of the resource cluster, and the minimum number of neighboring samples. Depending on the system state, the neural network outputs a set of parameters that best matches the current environment, controlling the execution strategy of the two clustering algorithms.

[0079] The model training phase constructs training samples based on historical scheduling and clustering records. Each set of samples contains three parts: system state input features, including load pressure level, equipment saturation, and historical response effectiveness; the clustering parameter combination used in actual execution, i.e., the radius of behavioral clustering, the neighborhood search radius of resource clustering, and the minimum number of neighborhood samples; and the actual performance indicators of clustering execution and scheduling response, such as user classification accuracy and scheduling response success rate. The training process employs a supervised learning mechanism. Using historical system states and validated optimal parameter combinations as input and output samples, the connection weights in the neural network are continuously adjusted through backpropagation, enabling the model to gradually learn the mapping relationship between input features and optimal parameters. Once the model training error converges to a preset threshold, it can be deployed to the running system. In the actual environment, it receives current state features in real time and outputs optimal parameters to control cluster boundaries or density thresholds, thereby achieving dynamic adaptive clustering decisions.

[0080] In step S4, the power grid load pressure level is calculated based on the ratio between the real-time load and the historical peak load of the target area. It is divided into four levels and the signals are input to the learning model. The saturation of regional equipment is calculated by the ratio of the number of connected regulating devices per unit area or unit load capacity to the theoretical maximum configuration capacity. The historical dispatch feedback effectiveness is obtained by weighting the average response execution rate, average instruction arrival rate and average load change success rate of the user group in the previous multiple response dispatches, forming a comprehensive score of multi-dimensional feedback indicators.

[0081] The power grid load pressure level is calculated based on the ratio between the real-time load of the target area and the historical maximum load peak of that area, reflecting the current urgency of dispatching and the system load pressure. This value is divided into four levels of signals and input into the neural network model, which enhances the model's ability to identify extreme operating conditions, reduces the volatility of real-time input signals, and improves the model's convergence speed and robustness. The level signals can be set according to different preset value ranges.

[0082] Regional equipment saturation is calculated by comparing the number of adjustable devices connected per unit area or unit load capacity with the theoretical maximum configurable number. This indicator objectively reflects the spatial distribution density and access level of controllable resources, and is an important reference factor for assessing whether the current controllable load has a basis for scheduling. In multi-regional and multi-type load scenarios, this indicator helps the model determine whether the resource cluster density is sufficient to form high-quality clusters.

[0083] Historical scheduling feedback effectiveness is calculated by retrieving indicators such as the average response execution rate, average scheduling instruction arrival rate, and average load change success rate of user groups from previous scheduling cycles. After statistical normalization, a comprehensive score is formed, which can be obtained through weighted calculation. The weight corresponds to the degree of influence of the indicator. This score represents the overall execution effect of scheduling behavior over a period of time, and can effectively reflect the behavioral response trend and execution controllability of users to scheduling strategies. It provides historical experience support for the model to judge whether "current clustering should be more cautious".

[0084] The set to be screened needs to be verified for reliability. A response confidence threshold is set for the users in the intersection, i.e. the set to be screened. Only when the user confidence exceeds the preset standard is it considered as an adjustable user resource with high response reliability and selected into the set of high adjustment potential loads.

[0085] Reliability verification refers to obtaining the ratio of the number of successful responses in all scheduled tasks within a set period to the total number of tasks. If the ratio reaches or exceeds the preset response confidence threshold, it is considered to have high response reliability and is included in the high adjustment potential load set; otherwise, it is not included.

[0086] The core principle of reliability verification is as follows: By analyzing the execution of load scheduling tasks by users within a set time period (such as the past month or several scheduling cycles), the ratio between the number of successful responses and the total number of tasks is calculated, and this ratio is compared with a preset response confidence threshold. If the ratio reaches or exceeds the threshold, it indicates that the user has high behavioral consistency and execution reliability in past scheduling executions, and therefore can be identified as a high-response-reliability adjustable user resource, and ultimately included in the high-adjustment-potential load set; otherwise, it will not be included temporarily to ensure the control accuracy of the resource pool.

[0087] For example, for each user in the overlapping user set, the execution records of that user in multiple past actual scheduling tasks are retrieved and evaluated item by item. First, it is confirmed whether the user actually received the scheduling command in each scheduling session. Then, it is determined whether the user responded within a reasonable time after receiving the command and maintained the load adjustment state until the task ended. If the user's response is significantly delayed, the adjustment time is insufficient, or the user exits prematurely in a scheduling session, it is recorded as a response failure; conversely, if the response behavior is stable, the load changes are timely and maintained until the scheduling ends, it is recorded as a response success.

[0088] The preset response confidence threshold can be obtained by dynamically adjusting the threshold model: obtain the following parameters: Λ: power grid load pressure level (value range is 1-4, representing low to extremely high); ∑: regional equipment saturation (normalized to [0,1]); Γ: historical dispatch feedback effectiveness (normalized to [0,1]);

[0089] Formula for dynamically setting response confidence threshold:

[0090] θ=θ0+α1·(Λ-1)+α2·(1-Σ)+α3·(1-Γ);

[0091] θ is the currently calculated response confidence threshold; θ0 is the default base threshold (e.g., 0.70, which can be set according to industry experience); α1, α2, and α3 are constants set through simulation or historical data experience to control the adjustment intensity of each variable on the threshold. The higher the pressure level, the higher the threshold (higher system scheduling risk); the lower the equipment saturation, the higher the threshold (fewer resources require higher confidence); the worse the historical feedback, the higher the threshold (reduce the risk of future failure). Based on the confidence evaluation mechanism constructed from real response history, and combined with the introduction of a dynamic threshold setting strategy driven by the power system operating status, this invention can significantly improve the screening accuracy of reliable response users, establish a highly stable and controllable load adjustment user pool, provide verifiable and quantifiable support for subsequent response calls, and effectively reduce scheduling failure rate and system fluctuation risk.

[0092] Example 2: Building upon Example 1, the high-regulation-potential load set in step S3 can be further subdivided into three subsets—industrial, commercial, and residential—based on user attribute tags. Within each subset, a secondary sorting is performed based on response behavior complexity and equipment diversity to achieve refined, hierarchical management of high-regulation-potential users. This hierarchical result can serve as the basis for power grid scheduling of regulation resources in different scenarios, further reducing the probability of scheduling failures caused by differences in user characteristics and achieving differentiated, highly efficient, and intelligent load regulation.

[0093] Response behavior complexity describes whether user behavior during response scheduling is stable, efficient, and predictable. It can be quantified into the following sub-indicators:

[0094] RBC1: Number of response actions (total number of control actions involved in the response behavior);

[0095] RBC2: Number of response state transitions (number of switching times in the load curve; more indicates complexity);

[0096] RBC3: Response Consistency Index (the degree to which user behavior deviates from the expected curve after scheduling is triggered);

[0097] RBC4: Response Decision Delay (the time interval from receiving a scheduling instruction to the first load change);

[0098] Device diversity describes whether the types of response devices used by the user access system are diverse and whether there are redundant control paths. It can be constructed from the following sub-indicators:

[0099] DDI1: Types and quantities of equipment (e.g., energy storage, electric heat pumps, air conditioners, electric vehicle charging stations, etc.);

[0100] DDI2: Equipment functional distribution breadth (whether each device covers different load periods / purposes, i.e., coverage percentage);

[0101] DDI3: Number of redundant adjustable paths (number of backup adjustable resources in case of failure);

[0102] DDI4: Standardization rate of device access control protocol (the more standardized, the easier to control);

[0103] Euclidean distance 1 between the response behavior complexity vector and the preset standard response vector, and Euclidean distance 2 between the device diversity vector and the preset standard device vector are obtained respectively. Then, a regression model is constructed, and the slope of the regression model is denoted as "response priority level score", which is used to refine the ranking of the subset of users with high adjustment potential.

[0104] The regression model constructed here is a linear regression model. Essentially, it uses the linear correspondence between a set of known sample points to determine the linear mapping between input variables and output scores, ultimately finding a straight line that best suits the distribution trend of these points, and using the slope of this line as the comprehensive response score for the user being evaluated.

[0105] The reason for choosing the linear model is that the input in this problem only contains two Euclidean distances as features, representing the degree of deviation between the user and the standard model. In scenarios with a limited number of dimensions and ranking and discrimination as the main objective, the linear model has sufficient expressive power and computational interpretability, and is also convenient for subsequent deployment and dynamic adjustment in engineering.

[0106] The Euclidean distance represents the distance between a target user's response complexity vector and an idealized standard response vector. This standard vector is typically defined as the user characteristic that performs best in dimensions such as shortest response latency, most stable response action, and most sufficient execution time. The smaller the distance, the closer the user's response behavior is to the ideal response model.

[0107] Euclidean distance 2 represents the distance between a target user's device diversity vector and an idealized standard device structure vector. The standard device vector represents an ideal access structure with diverse devices, complete response coverage, and high control protocol compatibility. A smaller distance indicates more comprehensive and flexible resource allocation for the user. Using these two distances as input variables and known historical priority levels as output variables, a training sample set is constructed. The sample points are fitted using minimum mean square error or other error metrics to train a linear regression model. This model can predict a new user's possible response priority level score based on the two distance indicators. The basic form of this model is:

[0108] Output value = intercept term + independent variable 1 × weight 1 + independent variable 2 × weight 2, where: independent variables 1 and 2 are input feature variables; weights 1 and 2 represent the degree of influence of each variable on the output in the model; the output value is the target score predicted by the model, such as the response priority level score. The model training process can be performed offline in batches or updated periodically to adapt to the dynamic changes in the control scenario and user structure.

[0109] In this scheme, the regression slope reflects the trend strength of the output score change when the two feature variables change. When Euclidean distance 1 and distance 2 change, a larger slope indicates that the score is more sensitive to the degree of deviation of user ability; conversely, a smaller slope indicates a more conservative approach. Therefore, this slope not only has ranking discrimination but can also be used to control the sensitivity and stability of scheduling incentives. In practical applications, this score can serve as a detailed ranking indicator for a set of users with high adjustment potential, enabling the sequential allocation of user resources according to the score gradient, thereby improving the resource scheduling efficiency and success probability of the scheduling system.

[0110] For example, consider a regional power mismatch event during extreme summer heat. In a coastal city experiencing a week of continuous high temperatures (≥38℃), the regional power grid faces peak load. The dispatch center uses load resources clustered into "high-potential user groups" for response control. However, due to some users failing to respond (distorted behavior patterns) and others being unable to execute commands (missing resource information), control fails, leading to emergency power rationing in some areas and causing economic and public outcry.

[0111] Based on the above scenario, a user screening process based on the intersection mechanism of two clusters is proposed:

[0112] First, two independent clustering models are run in parallel: one based on user response behavior characteristics, and the other based on users' adjustment of resource structure and state. After obtaining the clustering results of the two different dimensions, the intersection users are extracted by user identifier matching to ensure that they have both the initiative of behavioral response and the ability of physical adjustment.

[0113] Adaptive clustering parameter adjustment mechanism based on artificial neural networks:

[0114] Using the current operating pressure of the power grid, the saturation of regional control equipment, and historical dispatch feedback data as inputs, an artificial neural network model is trained to predict the optimal clustering parameters (such as behavioral clustering radius and resource density control factor) that are suitable for the current system state.

[0115] User confidence verification mechanism based on historical performance:

[0116] Even if a user is selected into the initial candidate set through bi-clustering, their response performance in past scheduling tasks still needs to be retrospectively analyzed. The response success rate for each user within a set period is calculated and compared with a confidence threshold dynamically calculated based on system status. Only users with a confidence level higher than the threshold are included in the final high-adjustment-potential set.

[0117] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0118] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0119] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0120] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0121] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A load clustering method based on load adjustability potential, characterized in that, Includes the following steps: Step S1: Collect the historical response behavior of various load users and the deployment information of their controllable power resources, and construct behavioral feature vectors and resource feature vectors respectively to represent the user's adjustment capabilities and behavioral patterns in multiple dimensions; Step S2: Based on the behavioral feature vector, a first clustering algorithm with cluster radius as the core of control is used to cluster users in the behavioral dimension; based on the resource feature vector, a second clustering algorithm with density as the core discrimination criterion is used to cluster users in the resource capability dimension. Step S3: Extract the intersection region of the first clustering result and the second clustering result, and identify the user groups identified in the intersection to obtain a set of loads with high adjustment potential, thereby screening out user resources with high response reliability under extreme high load scenarios; Step S4: Based on the current load pressure level in the power grid operation, the saturation level of regional regulation resources, and the feedback effect of the previous round of load response scheduling, the key parameters of the two clustering algorithms are dynamically adjusted using a pre-trained machine learning model to achieve continuous optimization of clustering boundaries and classification accuracy, thereby improving the accuracy of load regulation and the stability of power grid scheduling.

2. The load clustering method based on load adjustability potential according to claim 1, characterized in that, The behavioral feature vector constructed in step S1 includes the following feature dimensions: load response start-up delay, response duration, unit adjustment load, historical response frequency and user response willingness level. All behavioral feature data are used to form a time series matrix through a set sampling period, which serves as the basis for subsequent clustering input.

3. The load clustering method based on load adjustability potential according to claim 2, characterized in that, The resource feature vector constructed in step S1 includes the type of power adjustable device accessed by the load user, response interface compatibility, device availability status, maximum adjustable capacity and availability duration. All resource feature items are input into the second clustering algorithm as a structured parameter matrix. The construction of the resource feature vector also includes the historical failure rate of equipment and maintenance cycle indicators, which are then input into the second clustering algorithm after numerical mapping and standardization.

4. The load clustering method based on load adjustability potential according to claim 3, characterized in that, In step S2, the first clustering algorithm for behavior dimension clustering adopts the K-Medoids algorithm based on radius threshold. By using the preset initial cluster centers and radius boundaries, the behavior feature vectors are classified by similarity to form multiple clusters.

5. The load clustering method based on load adjustability potential according to claim 4, characterized in that, In step S2, the second clustering algorithm used for clustering based on the resource capability dimension adopts the DBSCAN algorithm based on the density aggregation principle. This algorithm is based on the Euclidean distance between users in the resource feature space, and defines the neighborhood search radius and the minimum number of neighborhood samples as control variables. Users with high density distribution characteristics of controllable power resources are formed into high-density clusters; low-density points outside the clusters are regarded as uncontrollable load users and are excluded in the subsequent intersection extraction step.

6. The load clustering method based on load adjustability potential according to claim 5, characterized in that, The intersection extraction operation in step S3 is completed by vector space position matching, that is, the user set output by the first clustering algorithm and the user set output by the second clustering algorithm are mapped in parallel, and an alignment matrix is ​​established with user identifier as the dimension. The user group that belongs to both clusters is selected. This intersection user set has the characteristics of active behavior response and physical availability of resources and devices, and is defined as the set to be selected.

7. The load clustering method based on load adjustability potential according to claim 6, characterized in that, The machine learning model in step S4 is built based on an artificial neural network. It takes the power grid load pressure level, regional equipment saturation and historical dispatch feedback effectiveness as input features, and outputs the optimal first clustering algorithm radius parameter and the neighborhood density control parameter of the second clustering algorithm.

8. The load clustering method based on load adjustability potential according to claim 7, characterized in that, In step S4, the power grid load pressure level is calculated based on the ratio between the real-time load and the historical peak load of the target area. It is divided into four levels and the signals are input to the learning model. The saturation of regional equipment is calculated by the ratio of the number of connected regulating devices per unit area or unit load capacity to the theoretical maximum configuration capacity. The historical dispatch feedback effectiveness is obtained by weighting the average response execution rate, average instruction arrival rate and average load change success rate of the user group in the previous multiple response dispatches, forming a comprehensive score of multi-dimensional feedback indicators.

9. A load clustering method based on load adjustability potential according to claim 8, characterized in that, The set to be screened needs to be verified for reliability. A response confidence threshold is set for the users in the intersection, i.e. the set to be screened. Only when the user confidence exceeds the preset standard is it considered as an adjustable user resource with high response reliability and selected into the set of high adjustment potential loads.

10. A load clustering method based on load adjustability potential according to claim 9, characterized in that, Reliability verification refers to: The ratio of the number of successful responses to all scheduled tasks within a set period to the total number of tasks is obtained. If the ratio reaches or exceeds the preset response confidence threshold, the user is considered to have high response reliability and is included in the high adjustment potential load set; otherwise, the user is not included.

Images