Power consumption behavior clustering analysis method and system based on big data

Through the multi-window sliding slice and depth encoder combined with the Dilicre process Gaussian hybrid dynamic clustering model, the problem that the electricity use behavior clustering method in the existing technology cannot identify the short-term mutation load pattern, and the dynamic, multi-scale and multi-source fusion analysis of the electricity use behavior in industrial parks is achieved, which improves the adaptability and accuracy of the clustering model.

CN120524255AActive Publication Date: 2025-08-22STATE GRID WUWEI POWER SUPPLY CO

Patent Information

Application Number
CN202511014365.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-23
Publication Date
2025-08-22
Estimated Expiration
2045-07-23

AI Technical Summary

Technical Problem

The existing electricity consumption behavior clustering method compresses the dynamic fluctuation information throughout the day to a single static vector, which cannot effectively distinguish the peak and valley event characteristics driven by different physical mechanisms, and is difficult to adapt to changes in production rhythm and seasonal photovoltaic output differences, resulting in the difficulty of reflecting the short-term mutation load mode of multi-source coupling such as electric vehicle fast charging and distributed photovoltaics in industrial parks.

Method used

The active power timing data is decomposed by a multi-window sliding slice mechanism, combined with operation and maintenance logs and meteorological monitoring data for feature fusion, the embedded vector of multi-source features is extracted using a depth encoder, and the number and center of the cluster are automatically determined through the Gaussian hybrid dynamic clustering model based on the Dirichlet process to generate electricity behavior labels.

Benefits of technology

Dynamic, multi-scale and multi-source fusion cluster analysis of the electricity consumption behavior of industrial parks is realized, peak and valley events driven by different physical mechanisms can be accurately identified, adapt to changes in external environment and production rhythm, and improve the generalization ability and application flexibility of the clustering model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120524255A_ABST
    Figure CN120524255A_ABST
Patent Text Reader

Abstract

The invention discloses a power consumption behavior clustering analysis method and system based on big data, and relates to the field of data analysis, and the method comprises the steps: firstly fusing collected active power time sequence data, operation and maintenance logs and meteorological monitoring data, and achieving the construction of sub-sequence-level features through the alignment of a multi-scale sliding window and a time window; then, multi-source fusion feature embedding is extracted by using a pre-trained depth encoder, so that the characterization capability for complex behaviors and mutation events is remarkably improved; further, adaptive determination of the cluster number and the center is realized through a Dirichlet process driven Gaussian hybrid dynamic clustering model, soft distribution is performed on embedded vectors by adopting a probability attribution degree inference method, and different power consumption behavior modes are accurately identified; and finally, according to the clustering labels, typical power consumption behavior labels are automatically distributed to all the parks. Therefore, the refined and intelligent level of energy efficiency management of the industrial park is comprehensively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of data analysis technology, and in particular to a method and system for clustering analysis of electricity consumption behavior based on big data. Background Art

[0002] With the proposal of my country's "dual carbon" goals and the in-depth promotion of the green and low-carbon development strategy, a large number of industrial parks have deployed high-power electric vehicle (EV) fast charging stations (usually with a power scale of 500 kW to 2MW) and distributed photovoltaic power generation systems (with an installed capacity of 1 MW to 5 MW) within the parks. As a result, the park's electricity load has shown short-term and large peak-to-valley fluctuations in the minute-level time series, such as the multi-megawatt "fast charging peak" load caused by EV fast charging, and the "reverse injection" phenomenon of reverse power flow into the power grid caused by the large-scale photovoltaic power generation during the noon period. These two types of characteristics are coupled with each other in time series and have mutational characteristics.

[0003] To accurately support campus energy efficiency management and demand response, unsupervised cluster analysis of massive amounts of electricity consumption time series data is required to identify typical user or campus behavior patterns, thereby identifying typical electricity consumption groups and load characteristics. However, currently mainstream clustering methods primarily employ a strategy for quantizing the overall time series: after partitioning the daily electricity consumption curve into fixed time windows, statistical features such as the mean, variance, peak time, and energy consumption during each time window are extracted and concatenated into a static high-dimensional vector. Alternatively, the original minute-level power value sequence is directly used as a high-dimensional input vector. Subsequently, algorithms such as K-Means and spectral clustering are applied to perform grouping.

[0004] However, this clustering method compresses dynamic fluctuations throughout the day into a single static vector, failing to distinguish between peak and valley event characteristics driven by different physical mechanisms (such as fast charging startup and photovoltaic power generation). This leads to coarse feature fusion, making it difficult for clustering results to effectively reflect key micro-dynamic features such as "fast charging startup transients" and "midday reverse injection." Furthermore, because the clustering objects are fixed as static vectors, the original method struggles to adapt to behavioral drift driven by multiple factors, such as changes in production rhythms, adjustments to operation and maintenance strategies, and seasonal differences in photovoltaic output. It is unable to deeply characterize short-term, sudden load patterns in industrial parks caused by the coupling of multiple sources, such as electric vehicle fast charging and distributed photovoltaics, hindering the implementation of refined energy efficiency management and demand response. Summary of the Invention

[0005] The present application provides a method, system, storage medium, computer program product and electronic device for clustering analysis of electricity consumption behavior based on big data, which is used to at least solve the problems in the current related technologies such as static vectorization processing leading to loss of microscopic dynamic characteristics, unidentifiable physical mechanisms of behavior, and difficulty in adapting to behavioral pattern drift.

[0006] In a first aspect, an embodiment of the present application provides a method for clustering analysis of electricity consumption behavior based on big data, including: collecting active power time series data, operation and maintenance log data and meteorological monitoring data of each industrial park, standardizing the active power time series data to generate a normalized sequence, sliding window slicing the normalized sequence with a preset step size based on multiple preset window lengths to form a subsequence set, and aligning the operation and maintenance log data and meteorological monitoring data to corresponding subsequences according to time windows; inputting the subsequence set and the aligned operation and maintenance log data and meteorological monitoring data into a pre-trained deep encoder to obtain an embedding vector set that integrates multi-source features; inputting the embedding vector set into a pre-trained Gaussian mixture dynamic clustering model based on the Dirichlet process to automatically determine the number of clusters corresponding to the embedding vector set and the cluster centers of each cluster, and assigning each embedding vector to the corresponding cluster according to the cluster centers of each cluster to generate a cluster label for the embedding vector set; determining the typical electricity consumption behavior category to which each industrial park belongs based on the cluster label, and generating corresponding typical electricity consumption behavior labels.

[0007] In the second aspect, the embodiment of the present application provides a clustering analysis system for electricity consumption behavior based on big data, the system comprising: a data acquisition unit for collecting active power time series data, operation and maintenance log data and meteorological monitoring data of each industrial park, performing standardization processing on the active power time series data to generate a normalized sequence, performing sliding window slicing on the normalized sequence with a preset step size based on multiple preset window lengths to form a subsequence set, and aligning the operation and maintenance log data and meteorological monitoring data to corresponding subsequences according to the time window; a feature encoding unit for aligning the subsequence set and the aligned operation and maintenance log data with the meteorological monitoring data; Data is input into a pre-trained deep encoder to obtain an embedding vector set that integrates multi-source features; a clustering analysis unit is used to input the embedding vector set into a pre-trained Gaussian mixture dynamic clustering model based on the Dirichlet process to automatically determine the number of clusters corresponding to the embedding vector set and the cluster centers of each cluster, and assign each embedding vector to the corresponding cluster according to the cluster centers of each cluster to generate a cluster label for the embedding vector set; a behavior labeling unit is used to determine the typical electricity consumption behavior category of each industrial park according to the cluster label, and generate a corresponding typical electricity consumption behavior label.

[0008] According to a third aspect, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the steps of the big data-based electricity consumption behavior clustering analysis method of any embodiment of the present application.

[0009] In a fourth aspect, an embodiment of the present application provides a storage medium on which a computer program is stored, characterized in that when the program is executed by a processor, the steps of the electricity consumption behavior clustering analysis method based on big data of any embodiment of the present application are implemented.

[0010] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the big data-based electricity consumption behavior clustering analysis method of any embodiment of the present application.

[0011] The power consumption behavior clustering analysis method and system based on big data provided by this application can produce at least the following technical effects: (1) Fully retain the dynamic fluctuation characteristics of the industrial park's electricity load at the minute level. The multi-window sliding slicing mechanism is used to decompose long time series data into subsequences covering different time scales, significantly enhancing the sensitivity to short-term sudden events (such as fast charging startup, photovoltaic reverse injection, etc.) and multi-source coupling behavior. Compared with static overall vectorized clustering, the proposed sliding window segmentation and multi-source data fusion method effectively avoids the loss of key information caused by time series feature compression, enabling the model to accurately distinguish peak and valley events driven by different physical mechanisms, and achieve a fine-grained characterization of the electricity consumption pattern of the industrial park.

[0012] (2) By using a deep encoder to fuse the features of each sliding window subsequence with synchronously aligned operation and maintenance logs and meteorological data, the clustering model's adaptability to external environments, production rhythms, and seasonal changes is improved, allowing typical electricity consumption behavior categories to dynamically reflect actual operating conditions and external influencing factors. A Gaussian mixture dynamic clustering model based on the Dirichlet process is used to adaptively identify diverse types of electricity consumption behaviors without pre-specifying the number of clusters. This effectively addresses the drift and evolution of park behavior patterns in different cycles and operating stages, improving the model's generalization and application flexibility.

[0013] Through this technical solution, the bottleneck of traditional static clustering methods has been broken through, and dynamic, multi-scale and multi-source fusion clustering analysis of electricity consumption behavior has been realized. It also provides a data foundation and decision-making basis for park-level energy efficiency management, load forecasting, and the formulation of demand response strategies, significantly improving the ability of refined management and response scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0015] Figure 1 A flowchart of an example of a method for clustering analysis of electricity consumption behavior based on big data according to an embodiment of the present application is shown; Figure 2 An operational flow chart of an example of distribution distance monitoring and concept drift adaptive updating according to an embodiment of the present application is shown; Figure 3 An operational flowchart of an example of training a depth encoder according to an embodiment of the present application is shown; Figure 4 An operational flowchart of an example of training a Gaussian mixture dynamic clustering model based on a Dirichlet process according to an embodiment of the present application is shown; Figure 5 The NMI performance comparison curves of different clustering methods under varying sample sizes are shown; Figure 6 A structural block diagram of an example of a big data-based electricity consumption behavior clustering analysis system according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0016] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0017] It should be noted that in the current research of some scholars on related technologies, a variety of novel clustering methods have been proposed, mainly including the following three directions: density-based clustering methods, clustering methods based on frequency domain or multi-scale subspace transformation, and clustering methods driven by deep learning.

[0018] Density-Based Spatial Clustering of Applications with Noise (DBSCAN) is a commonly used density-based clustering method. It automatically identifies clusters and noise based on sample density and presents multi-level clustering through a hierarchical structure. However, this method is extremely sensitive to the distance threshold (ε) and the minimum number of samples (MinPts), making parameter tuning difficult. Furthermore, campus electricity consumption sequences include both short-duration, minute-long fast-charging bursts and multi-hour photovoltaic fluctuations. This makes it difficult to account for events across multiple timescales using a single, fixed-scale density parameter. This limits the clustering results' ability to detect signals of varying scales.

[0019] In clustering based on frequency domain or multi-scale subspace transformations, the power consumption curve is subjected to Fourier transform or wavelet decomposition to extract spectral energy or wavelet coefficients at different scales, and then clustering is performed on this subspace. This clustering method is advantageous in capturing periodic fluctuations, but it is less responsive to transient, sudden changes, such as non-stationary, pulsed fast-charging startups and photovoltaic reverse injection. Furthermore, because it focuses only on the frequency components of the time series itself, it is difficult to incorporate external operational context, such as production calendars and maintenance logs, into the clustering features, affecting the interpretability and practicality of the results.

[0020] Deep learning-driven clustering methods often use autoencoders or variational autoencoders to automatically learn the mapping of high-dimensional electricity consumption time series to low-dimensional embeddings through neural networks, and combine them with algorithms such as K-Means to complete clustering. Although such methods can extract complex nonlinear features from raw data, they typically use the entire day's electricity consumption curve as a whole tensor input, ignoring explicit attention to local sudden events (such as fast charging start and stop, injection inflection points). In addition, deep models are highly dependent on the preset number of clusters and training distribution. Once the park operation and maintenance strategy is adjusted or there are seasonal differences in photovoltaic output, it is difficult to achieve online adaptive updates of the number of clusters and cluster centers.

[0021] It should be understood that the purpose of the above description of the current related art is only to facilitate the public to better understand the inventive spirit and motivation of this application, and is not to be construed as limiting this application. In addition, the technical solutions described in the above-mentioned current related art are not prior art and may also be undisclosed technical solutions, such as solutions under research or in the laboratory stage.

[0022] In the technical solutions of this application, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved shall comply with the provisions of relevant laws and regulations and shall not violate public order and good morals.

[0023] Faced with the dynamic characteristics of industrial park electricity load with multiple time scales, strong non-stationarity and multi-source coupling, traditional cluster analysis is difficult to take into account both instantaneous mutations and long-term fluctuations, and it is even more difficult to flexibly integrate multivariate spatiotemporal data, resulting in limited recognition accuracy and adaptability of typical electricity consumption behaviors. To this end, in the process of practicing this application, the inventors of this application, "starting from the multi-scale, dynamic and multi-source heterogeneous characteristics of electricity consumption behavior itself", proposed: a new cluster analysis method that integrates multi-window sliding analysis, cross-domain data alignment and deep embedded expression, aiming to achieve detailed analysis and dynamic adaptation of industrial park electricity consumption behavior at different time scales and data sources, and provide a solid data foundation and decision-making support for the scientific formulation of energy efficiency management and demand response strategies.

[0024] Furthermore, Figure 1 A flowchart of an example of a method for clustering analysis of electricity usage behavior based on big data according to an embodiment of the present application is shown.

[0025] The execution entity of the method of the embodiment of the present application can be any controller or processor with computing or processing capabilities. Specifically, it can be implemented by the campus energy efficiency management platform. By implementing a multi-source embedding and adaptive clustering mechanism, it realizes the classification and label generation of typical electricity consumption behaviors, provides a label system with time-series perception and physical consistency for campus-level electricity consumption pattern recognition, and improves the practicality and responsiveness of the data-driven energy efficiency management system.

[0026] In some examples, it can be integrated into an electronic device or terminal through software, hardware, or a combination of software and hardware, and the type of terminal or electronic device can be diverse, such as a mobile phone, tablet computer, or desktop computer, etc.

[0027] like Figure 1 As shown, in step S110, active power time series data, operation and maintenance log data, and meteorological monitoring data of each industrial park are collected, the active power time series data is standardized to generate a normalized sequence, the normalized sequence is sliced ​​by sliding windows with a preset step size based on multiple preset window lengths to form a subsequence set, and the operation and maintenance log data and meteorological monitoring data are aligned to the corresponding subsequences according to the time window.

[0028] It's important to note that, within the context of the "dual carbon" strategy, an increasing number of industrial parks are deploying photovoltaic power generation systems, charging stations, and energy storage equipment, resulting in diverse types of coupled electricity consumption. These behaviors exhibit distinct dynamic fluctuations over time, and load events triggered by different devices (such as electric vehicle fast charging and photovoltaic reverse injection) are highly volatile in localized timescales. Therefore, traditional modeling of overall daily electricity consumption curves is difficult to capture these short-term dynamics.

[0029] In some embodiments, the real-time active power is collected by smart meters at each substation, building, factory and charging pile in the industrial park at a frequency of one minute or five minutes to form a ,in Indicates the Parks in time In addition, the park energy management platform and the structured operation and maintenance events in the dispatching system are collected and operation and maintenance log data is collected, such as equipment maintenance, shift adjustment, temporary dispatch, etc. Each log contains key information such as event type, occurrence time, affected objects, and duration, forming a ,in Indicates the Parks in time Use the local weather station in the park or the authoritative weather service API to collect meteorological monitoring data, such as temperature, humidity, light, wind speed and other environmental variables, to ensure that the data sampling frequency is consistent with the power consumption data, and form ,in Indicates the Parks in time Real-time meteorological monitoring values.

[0030] Then, the power time series of each park and each monitoring object is statistically averaged by day or specified period. and standard deviation , and normalize each sampling point: , formula (1) Where, Indicates the Parks in time Normalized real-time power value.

[0031] In this way, the amplitude differences caused by installed capacity and baseline load between different parks and different monitoring points can be eliminated, the relative change trend can be highlighted, and the subsequent model can focus on the behavior itself rather than the absolute load.

[0032] Subsequently, multi-scale sliding window slicing is performed. Specifically, multiple window lengths can be selected according to business needs. L (such as 30 minutes, 1 hour, 2 hours, etc.), each window uses a sliding step size S (such as 1 minute) to slice the standardized time series: , formula (2) Where, Indicates the Parks at the moment The starting length is The normalized power sequence within the sliding window of . Indicates the Parks at the moment Normalized real-time power value. represents the window length of the sliding window, Indicates the Parks at the moment Normalized real-time power value at the moment.

[0033] A cyclic sliding window is used for all-day and all-season data to capture all potential behavioral fluctuations to the maximum extent possible, including various time series fragment features such as slow load increases, spikes, and sudden drops, so that an original sequence can be cut into a set of overlapping subsequences of multiple scales.

[0034] In addition, for each window, the operation and maintenance logs and meteorological data within the same time window are retrieved and aggregated. The logs match the window in an interval covering manner, recording all event types and occurrence times within the time window; meteorological data are aggregated by window average, extreme value, change rate, etc. Each sliding window subsequence Presented as structured triples ,in is the statistical feature of all event types in the window, This multi-scale structured feature input not only fully captures short-term mutations such as fast charging and load scheduling, but also accounts for medium- and long-term characteristics such as hourly / daily PV fluctuations and production shift adjustments. Furthermore, ensuring that each segment is synchronously associated with external O&M and meteorological context helps subsequent models interpret the causes of clustering.

[0035] It's important to note that, in the context of analyzing electricity usage in industrial parks, while power time series data can reflect dynamic changes in electricity load, these changes are often significantly impacted by park operations (such as maintenance, equipment switching, and load adjustments) and meteorological conditions (such as temperature, radiation, and humidity). By aligning and supplementing the power data with operation and maintenance log data and meteorological monitoring data at the same time window level, we can provide the external intervention and environmental context behind each power subsequence, enabling the model to accurately distinguish between identical behavioral patterns driven by intrinsic load fluctuations and those driven by external events. Consequently, multi-source alignment effectively enhances the multi-dimensional representation of clustering features, enabling the clustering model to better identify differences in electricity usage driven by operations and maintenance or weather, improving detection capabilities and clustering resolution for diverse scenarios (such as load fluctuations under extreme weather conditions and abnormal electricity usage during specific operation and maintenance cycles). This also ensures that the resulting cluster labels and behavioral features are more closely aligned with real-world scenarios.

[0036] In step S120 , the subsequence set and the aligned operation and maintenance log data and meteorological monitoring data are input into a pre-trained deep encoder to obtain an embedding vector set that integrates multi-source features.

[0037] It should be noted that the feature triples after multi-source slicing contain behavioral timing, operation and maintenance, and environmental context, but they are high in dimension, complex in structure, and have strong temporal correlation and multi-source interactivity. It is difficult to extract their internal laws by relying solely on manual rules or shallow mapping. Therefore, it is necessary to use deep encoders to automatically learn compressed, highly discriminative embedding vector expressions from a large amount of multi-source data, so that "similar behaviors" are close to each other in the embedding space and different behaviors are significantly separated, thereby realizing automatic abstract modeling from high-dimensional multi-source raw data to compact behavioral representations.

[0038] The deep encoder architecture can be diverse, including one-dimensional convolutional neural networks (1D-CNNs), gated recurrent units (GRUs), and long short-term memory networks (LSTMs), but this is not a limitation. Furthermore, the network architecture can be designed to cover multi-scale receptive fields, ensuring sensitivity to both local mutations and global trends. Operations and maintenance logs and meteorological features are fused with the primary time series features via an embedding layer or self-attention mechanism to enhance global semantic integration. Ultimately, the encoder outputs high-dimensional embedding vectors of fixed length (e.g., 128 or 256 dimensions). Each embedding vector comprehensively integrates multiple sources of information within the sliding window, including time series, operations, and meteorological information.

[0039] The pre-training mechanism of deep encoders can be diverse. For example, self-supervised prediction (such as power prediction of the next time window, anomaly detection), reconstruction loss (such as the reconstruction error of the autoencoder), or representation optimization based on contrastive learning can be used to ensure that the embedded expression has sufficient generalization and discrimination capabilities for the actual behavior of the park.

[0040] In some examples of the embodiments of the present application, the pre-trained deep encoder includes: multiple parallel one-dimensional convolutional layers, each one-dimensional convolutional layer has a different convolution kernel width, which is used to extract multi-scale temporal features of the input. After the output results of multiple one-dimensional convolutional layers are spliced, at least one residual block containing a hole convolution and a channel attention mechanism is input to fuse and enhance the feature expressions of power data, operation and maintenance log data, and meteorological monitoring data. The output of the residual block is processed by global average pooling and a fully connected layer to obtain an embedding vector that fuses multi-source features.

[0041] In the multi-scale convolution layer, it uses a parallel 1D convolution kernel group: : Convolution layer with kernel width of 3 and number of channels ; : Convolution layer with kernel width of 5 and number of channels ; : Convolution layer with kernel width of 7 and number of channels . Splice the output of each channel group to get the feature map .

[0042] In the residual stacking and channel attention module, several (e.g., 3 layers) residual blocks can be set. Each block contains: 1×1 convolution for dimensionality reduction, ReLU; dilated convolution (expansion rate 2) for context extraction; channel attention (SE module) to adaptively strengthen the operation / meteorological channel; the output of each layer is added to the input to enhance the gradient and information flow. Then, the feature map of the last layer of residual blocks is globally average pooled to obtain a length of After a layer of full connection + LayerNorm, the final embedding vector is output .

[0043] Through the embodiments of the present application, the deep encoder uses multiple parallel one-dimensional convolutional layers with different convolution kernel widths, which can effectively extract key features of the input data at different time scales, capturing multi-granular behavioral information ranging from short-term fast charging and photovoltaic mutations to long-term changes in production rhythm. The multi-scale features are then spliced ​​and input into a residual block containing a dilated convolution and channel attention mechanism. While ensuring the diversity of feature information, it can effectively fuse data features from different sources, making the output embedding vector more representative and discriminative, and strengthening the network's ability to express key events and coupling patterns.

[0044] In step S130, the set of embedded vectors is input into a pre-trained Gaussian mixture dynamic clustering model based on the Dirichlet process, the number of clusters and cluster center parameters obtained by model training are called, and each embedded vector is assigned to the corresponding cluster cluster through the cluster attribution inference method to generate a cluster label for the set of embedded vectors.

[0045] It's important to note that in a high-dimensional embedding space, complex industrial park electricity usage behavior can be automatically grouped through clustering. However, traditional K-Means or static clustering methods require manual specification of the number of clusters and struggle to adapt to dynamic changes in behavioral patterns. Here, we employ the Dirichlet Process Gaussian Mixture Model (DP-GMM) to automatically and adaptively determine the number of clusters and estimate cluster centers within the embedding space. While theoretically allowing for an unlimited number of clusters, in practice, we automatically "activate" a reasonable number of behavioral pattern clusters based on the data, each with a rich representation of its mean and variance.

[0046] Here, the trained clustering model is applied to new data as a "knowledge carrier." The training phase uses historical data and adaptive optimization to determine the optimal number of clusters, center locations, and parameters. The inference phase eliminates the need for re-clustering, leveraging learned patterns to quickly assign and label new data, achieving efficient data segmentation and pattern recognition.

[0047] Furthermore, time-series embedded data from industrial parks often has a complex distribution, with overlaps and anomalies. Therefore, probabilistic attribution inference can be used. Cluster attribution inference methods can be diverse, such as Variational Bayesian (VB), Expectation-Maximization (EM), Maximum A Posterior (MAP), or other equivalent probabilistic inference methods, which are not limited here. Compared to using a rigid distance threshold for assignment, this method comprehensively considers global mixing weights and local feature similarities, making the affiliation between each sample and each cluster more scientific and reasonable. This ensures that label assignments are consistent with the data distribution, even in the face of multimodal, outlier, or ambiguous behavior.

[0048] By feeding a set of embedded vectors into a pre-trained Gaussian mixture dynamic clustering model based on the Dirichlet process, the model fully utilizes the a priori clustering parameters obtained during the pre-training phase, such as the number of clusters and the location of cluster centers. This significantly improves the efficiency of batch clustering and label assignment for new data while ensuring high consistency between label assignments and historical data distributions. Furthermore, combining cluster probability attribution inference with cluster label assignment significantly improves the clustering efficiency and robustness of label discrimination for large-scale multi-source electricity consumption data, enabling flexible adaptation to real-world industrial park scenarios such as complex distributions, anomalous behaviors, and multimodal features.

[0049] In step S140 , the typical electricity usage behavior category of each industrial park is determined based on the clustering label, and a corresponding typical electricity usage behavior label is generated.

[0050] It should be noted that the cluster analysis output in step S130 only generates cluster labels for each sliding window subsequence, but actual business needs require typical power usage behavior categories for each campus as a whole. To this end, it is necessary to statistically summarize the labels of all subsequences within the same campus. Through frequency analysis and business mapping, the micro-clustering results are automatically converted into high-level behavior labels for energy efficiency management and policy optimization.

[0051] In some examples of the present application, for all subsequences within the same industrial park, the frequency of occurrence of each cluster in the industrial park is counted based on the cluster label, and a preset number of clusters with the highest occurrence frequency are designated as the primary clusters for the industrial park. Furthermore, based on a preset mapping relationship between clusters and electricity usage behavior categories, the user behavior category corresponding to the primary cluster is set as the typical electricity usage behavior label for the corresponding industrial park.

[0052] It should be noted that the sliding window subsequence level clustering has grouped the massive minute-level electricity consumption fragments according to the behavior pattern, and each subsequence corresponds to a cluster label. For the same industrial park, the labels of all its subsequences are aggregated to quantify the most common behavior patterns of the park in all fine-grained time windows. The preset number of the most frequent occurrences (e.g. ) clusters, it can be ensured that these “main clusters” can represent the dominant electricity consumption characteristics of the park, while occasionally excluding noise clusters or extreme events.

[0053] For example, for each park , collect the cluster labels of all its sliding window subsequences . Indicates the The cluster label of the first sliding window subsequence in the park, Indicates the The first The cluster labels of the sliding window subsequences, Indicates the The total number of all sliding window subsequences in a park.

[0054] Statistics for each cluster In the park Frequency of occurrence: , formula (3) Where, Indicates the Among the parks, The frequency of occurrence of clusters. Indicates the The sequence index of the sliding window subsequence of the park, Indicates the The first The cluster labels of the sliding window subsequences.

[0055] Sort and select the top Primary Cluster , avoiding occasional extreme clusters dominating the label results.

[0056] It should be noted that clusters themselves are merely numerical identifiers and do not directly reflect business implications. Therefore, after clustering is complete, specialized analysis tools can be introduced. For example, domain experts, large-scale model semantic label summaries, or historical data analysis can be used offline to assign a corresponding "behavior category" description (e.g., "nighttime fast charging spike") to each cluster number. After selecting the primary cluster, table lookup mapping can be used to convert the numerical clustering results into understandable and actionable high-level labels.

[0057] Specifically, a "cluster number-behavior category" mapping table can be established offline to clarify the business behavior corresponding to each cluster (such as "nighttime fast charging peak type", "midday photovoltaic reverse injection type", etc.) based on the typical timing curve shape, operation and maintenance / meteorological background, expert experience, etc. Mapped to corresponding behavior categories , forming a structured tag record , namely the park identification, typical behavior categories and their proportions. Indicates the cluster number, Indicates the The behavioral category labels of clusters, Indicates the Among the parks, The frequency of occurrence of clusters.

[0058] In this embodiment of the present application, the cluster labels of each subsequence are frequency-counted at the park level, and the main cluster with the highest frequency is selected. This automatically extracts the most representative behavioral patterns from the microscopic, massive slicing results, effectively filtering out occasional noise clusters and extreme events, and aggregating the massive amount of fine-grained labels into a small number of high-frequency main clusters, ensuring that the generated typical labels accurately reflect the overall electricity consumption characteristics of the park. Subsequently, through the set cluster-behavior category mapping table, the main clusters are mapped to high-level typical electricity consumption behavior labels, achieving the connection from data-driven clustering output to business-meaning labels.

[0059] Furthermore, the clustering labels of each industrial park can be sent to functional subsystems such as energy efficiency management, time-of-use electricity price strategy, and load forecasting through the system function interface, thereby providing decision-making basis for various intelligent scheduling and abnormal monitoring strategies in different industrial parks.

[0060] Through the embodiments of the present application, by utilizing multi-scale multi-source data slicing, deep embedding learning, dynamic adaptive clustering, and intelligent summarization of cluster labels, fine-grained identification, clustering, and automatic label generation of electricity consumption behaviors in industrial parks are achieved. Thus, through the contextual fusion analysis of multi-source features, fine-grained identification of complex behaviors such as short-term fast charging, long-term photovoltaics, and emergency maintenance in industrial parks is achieved, and the causes behind load changes are clear at a glance. In addition, behavioral labels with high representativeness and business interpretability are generated for each park, realizing the closed-loop operation of various business decision-making systems driven by dynamic labels.

[0061] It should be noted that during the ongoing operation of an industrial park, the external environment (such as weather changes), operation and maintenance strategies (such as load control), or equipment configuration (such as the addition of new fast-charging stations) can cause changes in the distribution of electricity usage patterns. Relying solely on fixed cluster centers makes it difficult for the model to adapt to this "concept drift," thus affecting the accuracy of cluster labels. Therefore, it is necessary to regularly evaluate the similarity between the distribution of embedding vectors in new batches and the distribution of historical cluster centers. When a significant distribution shift occurs, an adaptive model update is triggered.

[0062] Regarding the details of the cluster attribution inference in step S130, in some examples of the embodiments of the present application, the cluster attribution inference method softly assigns the attribution of the embedded vector by comprehensively combining the logarithmic probability expectation of the cluster mixing weight and the logarithmic probability expectation of the Gaussian likelihood of the embedded vector relative to each cluster center, thereby generating a cluster label for the set of embedded vectors.

[0063] It should be noted that in the context of industrial park time-series big data, the embedding vectors generated by the deep encoder already incorporate information from multiple sources (power, operations, and weather). However, actual park behavior often exhibits complex distributions with multimodal, overlapping, and abrupt changes. Using the Gaussian mixture dynamic clustering model, each embedding vector must be appropriately assigned a cluster label, ensuring that similar behaviors are grouped together and heterogeneous behaviors are dispersed. Variational Bayesian inference and log-probability expectation methods can efficiently and precisely accomplish this soft-to-hard assignment process.

[0064] Specifically, the cluster labels of the embedding vector set can be generated by the following operations: For each embedding vector, calculate its The responsibility of each cluster: , formula (4) Where, Indicates the The embedding vector of the fusion multi-source features has an embedding dimension of ; Indicates the The embedding vector for The responsibility of a cluster or cluster center satisfies ; Indicates the The mixing weights of the clusters are derived from the broken stick process construction of the truncated Dirichlet process, satisfying ; For the The mean vector of each cluster represents the location of the cluster center; For the The covariance matrix of a cluster, describing the distribution shape of the cluster; Indicates proportional to the sign. represents the Gaussian probability density function.

[0065] In variational Bayesian inference, for each embedding vector ( is a vector index), to mix the weights , Gaussian parameters As a priori, calculate its The responsibility of a cluster , that is, the probability weight of belonging to each cluster.

[0066] In the responsibility calculation formula of formula (4), the first term is the mixed weight The expectation of is used to measure the popularity of the cluster; the second term is the expectation of the Gaussian distribution likelihood, which measures With the cluster center The distance (similarity) between the two clusters is calculated. By combining the expected logarithmic probability of the mixed weights of each cluster with the expected logarithmic probability of the Gaussian likelihood of the embedding vector relative to each cluster center, a soft assignment of the attribution degree to each embedding vector is effectively achieved. This attribution degree not only fully reflects the degree of probabilistic match between the embedding vector and each cluster center, but also incorporates the prior weight information of the clustering model. Thus, the assignment of responsibility fully reflects the match between each embedding vector and each cluster, as well as the importance of the cluster itself, improving the robustness of label assignment in multimodal and multi-distribution clustering scenarios.

[0067] is the expectation of the weight of the folded stick process, which represents the random variable The logarithmic expectation under the variational distribution is used to evaluate the credibility of cluster weights: , formula (5) Where, Indicates the first The two posterior parameters of the Beta distribution control The shape of the distribution and the probability of generating new clusters; represents the Digamma function, Indicates the number of nodes from cluster 1 to cluster The cumulative cluster index of the cluster. and Indicates the The two posterior parameters of the Beta distribution.

[0068] In the above formula (5), the logarithmic probability expectation of the cluster mixing weight is expressed analytically, specifically using the variational inference based on the Dirichlet process, where the mixing weight Obey the Beta distribution and use the Digamma function to calculate the two posterior parameters of the Beta distribution This method not only quantifies the prior importance of each cluster in the overall distribution, but also effectively reflects the dynamic probability mechanism of new cluster generation by accumulating the probability of previous clusters.

[0069] is the expectation of the log-likelihood of the Gaussian distribution: , formula (6) Where, Indicates and The multivariate Gaussian probability density function described; represents the vector transpose operation, For the The precision matrix of clusters is defined as . represents the number of embedding dimensions of the embedding vector, Indicates the The expected value of the logarithm of the covariance matrix of the clusters.

[0070] In Equation (6), the expectation of the log-likelihood of the Gaussian distribution is quantitatively expressed, incorporating the distance between each embedding vector and the cluster center, the covariance structure of the clusters, and the precision matrix into the degree of attribution calculation. This approach not only fully considers the dimensionality of the embedding space, the distribution shape of each cluster, and the Euclidean distance between the sample and the center, but also uses the expectation operator to model parameter uncertainty. This refined probabilistic assessment method effectively improves the discriminative power and stability of cluster attribution inference, enabling the model to more accurately distinguish different electricity consumption patterns in complex industrial park multi-source behavioral data.

[0071] Each embedding vector Assign to The largest cluster : , formula (7) Where, Is a mathematical operator, indicating that among all clusters, the degree of belonging is selected Largest cluster index .

[0072] Until each embedding vector in the set of embedding vectors is assigned a unique cluster label.

[0073] In multimodal clustering scenarios, each vector sample is assigned a certain probability across all clusters, but each embedded vector needs to be assigned a unique, deterministic cluster label. Therefore, by maximizing the responsibility principle, each vector sample is assigned to the cluster to which it most likely belongs, achieving the transition from soft probability to hard labeling.

[0074] Through the embodiments of the present application, in the face of the problems of multimodality, high distribution overlap and behavioral mutation in the big data cluster analysis scenario of electricity consumption behavior in industrial parks, a Gaussian mixture model responsibility allocation method based on variational Bayesian inference is adopted. It can combine the mixed weight of each cluster cluster and the multidimensional similarity of the embedding vector and each center, and through the rigorous mathematical modeling of the expected logarithmic probability, carefully characterize the probability distribution of each embedded vector sample belonging to each cluster. Furthermore, by maximizing the principle of responsibility, the soft allocation is converted into a unique cluster label, which realizes the accurate discrimination of mainstream, marginal and abnormal electricity consumption patterns. As a result, the accuracy and robustness of pattern discovery and label assignment of electricity consumption behavior in complex industrial parks can be significantly improved.

[0075] It should be noted that in the cluster analysis scenario of electricity consumption behavior in industrial parks, data distribution will change dynamically with the change of seasons, changes in production rhythm, equipment upgrades or adjustments to operation and maintenance strategies, that is, there is a "concept drift" phenomenon. If the clustering model always relies on the cluster centers obtained by training historical data, it will be difficult to capture the emerging behavior patterns, resulting in the gradual disconnection of cluster labels and actual electricity consumption behavior. In view of this, Figure 2 An operational flowchart of an example of distribution distance monitoring and concept drift adaptive updating according to an embodiment of the present application is shown.

[0076] like Figure 2 As shown, in step S210, the distribution distance between the embedding vector set and the historical cluster center is calculated.

[0077] Here, the historical cluster centers are cluster centers automatically determined and saved for the embedding vector sets of the previous batches by using a Gaussian mixture dynamic clustering model based on the Dirichlet process.

[0078] In some embodiments, the global difference between two groups of distributions is measured by distribution distance. The distance measurement indicators can be diverse, such as JS divergence, relative entropy, maximum mean difference (MMD), etc. The larger the distance, the more significant the separation of the new behavior distribution from the original model, and the model structure or parameters need to be updated.

[0079] It should be noted that in the cluster analysis of electricity consumption behavior in industrial parks, the embedded vector sets and historical cluster centers are often distributed in high-dimensional space, and these distribution structures may be multimodal, heteroscedastic, or even have nonlinear distribution drift. Therefore, in some examples of the embodiments of the present application, the distribution distance uses Wasserstein distance. Wasserstein distance has the ability to measure the "overall shape difference" of two probability distributions. It can not only reflect the changes in the mean and variance of the distribution, but also capture the main direction of the distribution, the degree of expansion, and the non-overlapping areas between the distributions. Therefore, using Wasserstein distance as a distribution distance metric can more comprehensively and intuitively measure the true difference between the distributions of new and old data, which is better than distance metrics based only on first-order and second-order statistics such as mean or variance.

[0080] For example, let the new batch embedding set be , the historical cluster center set is , then we can calculate: , formula (8) Where, represents the distribution distance, represents the new embedding vector, Represents the historical cluster center.

[0081] In step S221 , if the distribution distance does not exceed the preset distance threshold, only the historical cluster center position is updated.

[0082] Here, when the distribution of the new batch data is highly consistent with the historical cluster centers, it means that the original clustering model is still valid. Only a small position correction of the cluster centers is needed to adapt to slight behavioral fluctuations and avoid unnecessary retraining of the entire model.

[0083] Specifically, for each cluster , using the mean of the embedding vectors assigned to the cluster by the new batch, the historical center Perform weighted update: , formula (9) in, Indicates the The updated cluster centers of the clusters, For the new batch Cluster mean, is the updated weight coefficient (such as 0.9~0.99).

[0084] In addition, second-order statistics such as covariance and sample size can be updated synchronously to describe the shape of cluster distributions. This allows for rapid response to small data fluctuations, ensuring cluster label continuity without requiring full model retraining and reducing system resource consumption.

[0085] In step S223, if the distribution distance exceeds the preset distance threshold, it is determined that the Gaussian mixture dynamic clustering model based on the Dirichlet process has a concept drift phenomenon, and the Gaussian mixture dynamic clustering model based on the Dirichlet process is fine-tuned based on the embedding vector set.

[0086] It should be noted that if the distribution of new data deviates significantly from the historical cluster center, it often means that new behavior patterns, equipment upgrades or extreme operation and maintenance events (such as energy-saving strategy adjustments) have emerged, making the original clustering structure no longer applicable, resulting in "concept drift". Instead, model fine-tuning or retraining should be triggered to dynamically adapt to new scenarios and ensure that cluster labels remain valid.

[0087] Specifically, when The fine-tuning process is entered at this time, and the original DP-GMM model can be quickly iterated by using incremental EM algorithm, partial parameter unlocking training, etc. Represents the critical threshold for concept drift. Furthermore, while training data allocation primarily relies on embeddings from new batches, it can also incorporate representative historical samples to prevent the model from overfitting to anomalies or short-term events. This effectively addresses significant behavioral changes and new scenarios, allowing for timely adaptation to shifts in electricity usage patterns caused by emergencies and major policy adjustments.

[0088] The embodiments of this application monitor the distribution distance between the new batch embedding vectors and the historical cluster centers, and adopt a distribution distance monitoring and concept drift adaptive update mechanism. This allows for rapid incremental correction of cluster centers when the behavior distribution is stable, maintaining model efficiency and cluster label continuity. It also allows for timely triggering of model fine-tuning or retraining when significant changes in the behavior distribution occur, ensuring that cluster labels accurately and synchronously reflect the electricity usage patterns of the industrial park.

[0089] Figure 3 An operational flowchart of an example of training a depth encoder according to an embodiment of the present application is shown.

[0090] like Figure 3 As shown, in step S310, a candidate subsequence set is sampled from the historical subsequence set, and a representative shape subsequence is determined from the candidate subsequence set based on an entropy gain method.

[0091] Here, the shape subsequence refers to the power value segment captured within a sliding window in the power consumption time series curve. Its essence is the shape (or "contour") of the power consumption over time within a certain period of time. For example, a segment that rises sharply and then stabilizes may represent a fast charging startup behavior pattern, and a sudden drop after a midday peak may correspond to a sudden drop in photovoltaic output behavior pattern, and so on.

[0092] It should be noted that electricity consumption in industrial parks exhibits a rich variety of temporal patterns. However, different types of fluctuations, mutations, or periodic changes often occur at disparate rates within a large sample pool, making direct sampling difficult to ensure that the training set covers and represents all types of behavior. Using entropy gain to select representative subsequences with diverse shapes and high pattern discrimination from a set of candidate subsequences helps strengthen the model's ability to identify abnormal behaviors and extreme patterns, improving the encoder's generalization and robustness.

[0093] More specifically, after slicing different time windows, the standardized power subsequences of all parks and all time periods are used as the basic pool to form a historical subsequence set. For each candidate subsequence , using Euclidean distance as the metric, calculate its The optimal alignment distance of all historical subsequences in Divide the history set into two categories and calculate Information entropy And the group entropy, the entropy gain method is expressed by the following formula: , formula (10) Where, Candidate subsequence The entropy gain value, is the set of historical subsequences, is the information entropy function, Candidate subsequence The optimal partitioning threshold of the Euclidean distance with other subsequences in the historical subsequence set; Representation and candidate subsequence The Euclidean distance is not greater than the optimal partition threshold The set of historical subsequences of Representation and candidate subsequence The Euclidean distance is greater than the optimal partition threshold The set of historical subsequences of Indicates the number of elements in the collection.

[0094] Specifically, The larger the value is, the more significant the improvement of the overall information diversity by the subsequence. Several subsequences with the largest entropy gain can be selected as representative shape subsequences to enhance the diversity and discrimination of training samples, significantly improve the encoder's ability to represent mainstream and abnormal behaviors, thereby covering rich behaviors, maximizing diversity with fewer samples, and improving training efficiency.

[0095] In step S320 , positive and negative sample pairs are constructed from the historical subsequence set based on the representative shape subsequence, wherein the subsequence pairs corresponding to the same shape subsequence are used as positive sample pairs, and the subsequence pairs not containing the same shape subsequence are used as negative sample pairs.

[0096] Deep contrastive learning relies on a proper division of positive and negative pairs: positive pairs are required to have highly similar shapes and behavioral patterns, while negative pairs should come from significantly different patterns. By using representative shape subsequences, the subsequences with the smallest Euclidean distance are classified as positive pairs with the same shape, while those without the same shape are classified as negative pairs. This maximizes differentiation and enables the encoder to learn essential behavioral characteristics.

[0097] In step S330 , based on the positive and negative sample pairs, the depth encoder is trained using a contrastive loss function with a temperature adjustment factor to determine network parameters of the pre-trained depth encoder.

[0098] Here, a contrastive loss function is used to push positive pairs closer together in the embedding space and negative pairs farther apart. Furthermore, a temperature adjustment factor is used to adjust the sharpness of the distribution and the range of gradient flow. This helps the model form a compact yet dispersed inter-class and intra-class structure in high-dimensional space, making it more adaptable to multimodal and highly dynamic industrial scenarios.

[0099] For example, for each pair of positive samples ( ) and negative samples ( ), define the contrast loss function with temperature adjustment factor : , formula (11) Where, Represents a positive sample With the current sample The cosine similarity between Represents negative samples With the current sample The cosine similarity between is the temperature hyperparameter. Indicates that the current sample The positive sample feature vectors that belong to the same representative shape subsequence, Indicates that the current sample Negative sample feature vectors that do not belong to the same representative shape subsequence.

[0100] Input positive and negative samples in each batch, accumulate the loss for all sample pairs and backpropagate, and optimize the encoder parameters in batches. Temperature hyperparameter Determined by experience or cross-validation (e.g., 0.05–0.2), this value controls the sample's sensitivity to distance distribution. Through contrastive learning optimization, this allows for tighter clustering of segments with the same pattern, clearer separation of different patterns, and improved mini-batch training efficiency.

[0101] Through the embodiments of the present application, the entropy gain is used to select representative subsequences that can maximize information diversity, combined with the automatic construction of positive and negative sample pairs based on Euclidean distance, and then the contrast loss function with a temperature adjustment factor is used to train the deep encoder. This can significantly improve the model's ability to distinguish different electricity consumption behavior patterns, and can generate multi-source time series embedding vectors that are highly discriminative, structured in distribution, and sensitive to abnormal behavior.

[0102] Figure 4 An operational flowchart of an example of training a Gaussian mixture dynamic clustering model based on a Dirichlet process according to an embodiment of the present application is shown.

[0103] like Figure 4 As shown, in step S410, a pre-trained deep encoder is used to generate a set of historical embedding vectors corresponding to a set of historical subsequences, and the cluster centers are initialized based on the historical embedding vector set through K-Means++ algorithm analysis.

[0104] It should be noted that if you directly perform dynamic clustering on high-dimensional embedding vectors, you first need to initialize the cluster centers with high quality. Otherwise, it is very easy to fall into local optimality or produce label drift. K-Means++ is an initialization method that can maximize the distance between samples and reasonably distribute the initial center points. It can effectively avoid performance fluctuations caused by the initial randomness of clustering.

[0105] In some embodiments, an embedding vector is randomly selected from the set of historical embedding vectors as the first cluster center. Then, for each embedding vector in the set of historical embedding vectors that has not yet been selected as a cluster center, the probability of the embedding vector being selected as the next cluster center is determined according to the proportion of the square of the minimum distance between the embedding vector and the selected cluster center in all the embedding vectors that have not yet been selected, until a certain number of embedding vectors is selected. M Cluster centers, M is the preset number of cluster centers.

[0106] The core of the K-Means++ algorithm is that subsequent centers are no longer randomly selected, but rather tend to be located farther from existing centers. By calculating the square of the minimum distance between each unselected embedding vector and the selected center, the new centers are effectively forced to cover scattered, rare, or extreme patterns in the data space, improving the spatial representativeness and distribution diversity of the initialization, ensuring that the initial centers are highly dispersed and have global coverage.

[0107] In step S420, the studentt The distribution calculates the initial soft attribution of each historical embedding vector to the cluster center.

[0108] It should be noted that in high-dimensional space, the "affiliation relationship" between data points and cluster centers should not be divided by hard binary values, but should adopt a probabilistic and smooth soft attribution method, which not only reflects the distance relationship but also is compatible with complex distributions such as overlap and outliers. t The distribution kernel (fat-tailed distribution) is naturally suitable for high-dimensional clustering scenarios. It can adaptively adjust the initial degree of belonging (similarity weight) of samples to each cluster to avoid the adverse effects of outliers or noise on the initial clustering results.

[0109] , formula (12) Where, For the History embedding vector For the first Cluster centers The initial soft attribution of Indicates the The historical embedding vectors generated by the deep encoder, the total number of historical embedding vectors contained in the historical embedding vector set is ; For the The location vectors of cluster centers, is the number of cluster centers; For students t The degrees of freedom hyperparameter of the distribution is used to adjust the fat-tail characteristics of the distribution; is the Euclidean distance operator, represents the index of any cluster center, For the The location vectors of the cluster centers.

[0110] here, The larger the value of is, the closer it is to the normal distribution, and the smaller it is, the more tolerant it is to outliers. In addition, you can also set each For all Center All calculations are performed to form a soft attribution matrix.

[0111] In step S430, the corresponding target soft attribution degree is calculated according to the initial soft attribution degree. is the initial soft attribution It is obtained by normalizing the square of .

[0112] Here, the square of the initial soft attribution degree is used to improve the discrimination of the attribution degree distribution (i.e., enhancing the confidence of high attribution degrees and suppressing the influence of low attribution degrees), and normalization is performed to ensure that the sum of the target soft attribution degrees of all clusters is 1. This target soft attribution degree not only numerically strengthens high-attribution degree clusters and suppresses low-attribution degree clusters, but also, after normalization, can describe the characteristics of the global probability distribution, enabling guidance based on this global probability distribution.

[0113] In step S440 , the cluster center position is optimized and updated by minimizing the KL divergence loss function between the initial soft attribution degree and the target soft attribution degree.

[0114] The target soft attribution score, obtained by squaring and normalizing the initial soft attribution score, essentially constructs a more concentrated and discriminative ideal probability distribution, which serves as the "learning target" during model optimization. To enable the model to adaptively aggregate similar behaviors and separate divergent ones, the initial soft attribution score (based on distance) must be as close as possible to the target soft attribution score (based on the global probability distribution).

[0115] Here, the initial soft attribution is considered as the model’s current predicted distribution of attribution for each embedding vector, while the target soft attribution serves as the “teacher distribution” to which the model is expected to converge.

[0116] Use KL divergence loss as the optimization objective: , formula (13) Where, Represents the KL divergence value; Indicates the The embedding vector corresponds to the The probability of belonging to the cluster center; Represents the total number of historical embedding vectors contained in the historical embedding vector set, is the number of cluster centers; is the index of the embedding vector, is the index of the cluster center.

[0117] Adjust the cluster center through optimization methods such as gradient descent , and iterate continuously until convergence.

[0118] Using the KL divergence loss function to measure the distance between the model prediction distribution and the target distribution can effectively guide the optimization direction of the model cluster center, so that the actual output distribution gradually approaches a clearer and more discriminative target distribution, thereby improving the accuracy and robustness of cluster label assignment.

[0119] According to the target attribution weight, the vectors of all samples are weighted averaged and the new position of each cluster center is updated: , formula (14) Where, Indicates the A history embedding vector generated by the deep encoder, Represents the total number of historical embedding vectors contained in the historical embedding vector set, is the index of the embedding vector; Indicates the The historical embedding vectors are The target soft belonging degree of cluster centers, Indicates the The updated positions of the cluster centers.

[0120] Through multiple rounds of iteration, the cluster centers converge from the initial values ​​to the actual data density areas. As a result, the attribution degree and cluster centers are synchronously and adaptively updated, improving clustering quality and discrimination. This also avoids the misclassification of transitional samples by hard-label clustering and improves adaptability to abnormal behavior and pattern drift.

[0121] In step S450, the Dirichlet process is used to recursively assign a cluster number to each embedding vector, and the number of non-empty clusters is used as the number of clusters of the clustering model.

[0122] Specifically, the Dirichlet process is represented by the Chinese Restaurant Process (CRP) as follows: , formula (15) Where, Indicates the The cluster labels of the embedding vectors, Indicates the The index of the cluster to which the embedding vector is finally assigned, and the corresponding cluster has the highest probability of assignment; Assign probability to clusters, indicating the The embedding vector is assigned to the The probability of a cluster, conditional on the previous The cluster assignment of the embedding vectors is known; Indicates the front The embedding vectors have been assigned to The number of samples in each cluster; represents the concentration hyperparameter of the Dirichlet process, which is used to adjust the probability of new clusters appearing.

[0123] It's important to note that electricity usage patterns in industrial parks are not fixed, and the number of clusters should be able to automatically grow or shrink based on the data scenario. The Dirichlet process (DP), a typical method for nonparametric Bayesian clustering, allows for a data-driven determination of the final number of clusters. The Chinese restaurant process implements this: for each incoming sample, the number and composition of clusters is adaptively characterized based on the popularity or innovation probability of existing clusters. This eliminates the need to pre-set the number of clusters, automatically discovers new patterns in electricity usage, and dynamically captures changes in patterns due to seasonal changes, policy adjustments, and other factors.

[0124] In step S460, a preset distance threshold is determined based on the distribution distance between the embedding vector set of the historical batch and the cluster center. θ .

[0125] Specifically, the Wasserstein distance between the historical embedding vector set of each historical batch and the cluster center is calculated, and the preset distance threshold is updated based on the 95% quantile of the distance distribution of all historical batches. θ .

[0126] It should be noted that the dramatic temporal variations in distributed photovoltaic output and EV fast-charging consumption can cause behavioral patterns in industrial scenarios to change significantly over time, so clustering models cannot be "set in stone." By calculating the Wasserstein distance between each batch's embedding distribution and the cluster center distribution and dynamically updating the distribution distance threshold, we can intelligently determine whether the current clustering model is invalid or needs updating, enabling online self-maintenance and self-evolution of the model.

[0127] In addition, if the distance of the new batch exceeds the threshold, model retraining or fine-tuning (such as resetting the center, adding new clusters, etc.) is automatically triggered; otherwise, only a small update of the location parameters of the cluster center is performed.

[0128] The present invention employs a Gaussian mixture dynamic clustering model based on the Dirichlet process. Instead of presetting the number of clusters, the model adaptively expands or merges clusters based on the actual structure of the data through a Chinese restaurant process. The use of a non-parametric Bayesian strategy enables the model to flexibly capture emerging electricity usage patterns and automatically merge them when they disappear. This model continuously reflects concept drift or behavioral mutations in the real business of a real industrial park, enhancing the clustering model's adaptive and evolutionary capabilities.

[0129] In order to further prove the superiority of this technical solution, the comparative experimental verification part of this technical solution will now be expanded.

[0130] Specifically, to verify the effectiveness and superiority of the proposed "Electricity Consumption Behavior Cluster Analysis Method Based on Multi-Source Deep Embedding and Dynamic Probabilistic Clustering" in the industrial park scenario, we used high-frequency power consumption data, operation and maintenance logs, and meteorological monitoring data from real industrial parks in a certain province from January 2022 to December 2023. Forty industrial parks equipped with EV fast-charging stations and distributed photovoltaic systems were selected as test objects. Power data was sampled with a one-minute period, and supporting data was aligned to a one-hour benchmark using a sliding window method.

[0131] Method A (this paper's solution): Multi-source deep encoder + Dirichlet process Gaussian mixture dynamic clustering (DP-GMM), with soft attribution KL fine-tuning; Method B (whole vector K-Means): After the traditional aligned time series is vectorized as a whole, K-Means clustering is used; Method C (Multi-scale Wavelet K-Means): Wavelet transform features + K-Means clustering; Method D (Autoencoder + Spectral Clustering): Deep Autoencoder Embedding + Spectral Clustering; Method E (DBSCAN density clustering): density-based unsupervised clustering with parameters optimized by grid search.

[0132] All methods used the same data slicing and normalization process, and the number of clusters was set to automatic (DP-GMM, DBSCAN) or best practice (the rest of the methods).

[0133] In order to comprehensively evaluate the clustering effect, this paper selects the following indicators: NMI (Normalized Mutual Information): measures the consistency between cluster labels and expert manual labels; Silhouette Coefficient: measures the internal compactness of the sample and the separation between clusters; Cluster Purity: The proportion of main class samples in the main cluster; Runtime: The time it takes for the clustering algorithm to run on data of the same size, in seconds; Behavior capture rate: The recognition coverage of key power behavior patterns such as "fast charging transients" and "reverse injection".

[0134] Table 1. Clustering performance comparison table

[0135] As shown in Table 1, the method provided in this embodiment (i.e., Method A) significantly outperforms the comparison methods across all mainstream clustering evaluation metrics. In particular, it achieves significant improvements over traditional methods such as K-Means and density clustering in terms of NMI, cluster purity, and key behavior capture rate. It can more accurately identify complex power usage patterns within a campus, such as sudden and overlapping patterns. This is attributed to the fusion of multi-source contextual features, the high expressiveness of the deep encoder, and the dynamic cluster number adaptation mechanism. Furthermore, while Method A's runtime is slightly longer than that of fast methods like K-Means, it outperforms traditional spectral and density clustering, and offers greater scalability and business interpretability.

[0136] Figure 5 The NMI performance comparison curves of different clustering methods under varying sample sizes are shown. Figure 5 The horizontal axis of the coordinate system in the figure represents the sample size (individuals) of the electricity consumption behavior clustering task, and the vertical axis represents the NMI index, which measures the consistency between the clustering results and manual labels. The proposed multi-source deep embedding + dynamic probabilistic clustering (i.e., method A) is compared with K-Means ensemble (method B), wavelet K-Means (method C), autoencoder + spectral clustering (method D), and DBSCAN density clustering (method E). Figure 5 It can be seen intuitively that the proposed method A achieves the highest NMI performance across the entire sample size range, maintaining excellent clustering consistency and resolution as the data size increases, significantly outperforming other methods. Traditional methods perform reasonably well with small samples, but as the sample size increases and the scenario complexity increases, the NMI improvement is limited and even plateaus, demonstrating insufficient generalization capabilities. This validates the clustering advantages of this technical solution in complex industrial electricity data scenarios.

[0137] Furthermore, the distribution of each cluster label across typical electricity usage behaviors (such as EV fast charging, PV reverse injection, and seasonal load drift) is further analyzed. Method A accurately classifies key behavioral events into distinct clusters, reducing the likelihood of behavioral confusion or cluster label drift and enabling named matching of primary behavioral patterns after label mapping.

[0138] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of combined actions, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application. In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0139] Figure 6 A structural block diagram of an example of a big data-based electricity consumption behavior clustering analysis system according to an embodiment of the present application is shown.

[0140] like Figure 6 As shown, the electricity consumption behavior clustering analysis system 600 based on big data includes a data acquisition unit 610 , a feature encoding unit 620 , a clustering analysis unit 630 and a behavior marking unit 640 .

[0141] The data acquisition unit 610 is used to collect active power time series data, operation and maintenance log data and meteorological monitoring data of each industrial park, standardize the active power time series data to generate a normalized sequence, slice the normalized sequence with a preset step size based on multiple preset window lengths to form a subsequence set, and align the operation and maintenance log data and meteorological monitoring data to corresponding subsequences according to the time window.

[0142] The feature encoding unit 620 is used to input the subsequence set and the aligned operation and maintenance log data and meteorological monitoring data into a pre-trained deep encoder to obtain an embedding vector set that integrates multi-source features.

[0143] The cluster analysis unit 630 is used to input the embedding vector set into a pre-trained Gaussian mixture dynamic clustering model based on the Dirichlet process to automatically determine the number of clusters corresponding to the embedding vector set and the cluster centers of each cluster, and assign each embedding vector to a corresponding cluster according to the cluster centers of each cluster to generate a cluster label for the embedding vector set.

[0144] The behavior labeling unit 640 is used to determine the typical electricity usage behavior category of each industrial park according to the clustering label, and generate a corresponding typical electricity usage behavior label.

[0145] In some embodiments, an embodiment of the present application provides a non-volatile computer-readable storage medium, in which one or more programs including execution instructions are stored. The execution instructions can be read and executed by electronic devices (including but not limited to computers, servers, or network devices, etc.) to execute the steps of any of the above-mentioned methods of electricity consumption behavior clustering analysis based on big data in the present application.

[0146] In some embodiments, the embodiments of the present application also provide a computer program product, which includes a computer program stored on a non-volatile computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer executes any step of the above-mentioned electricity consumption behavior clustering analysis method based on big data.

[0147] In some embodiments, an embodiment of the present application also provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the steps of the electricity consumption behavior clustering analysis method based on big data.

[0148] The above-mentioned product can execute the method provided in the embodiment of this application, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the method provided in the embodiment of this application.

[0149] The electronic devices of the embodiments of the present application exist in various forms, including but not limited to: (1) Mobile communication devices: These devices are characterized by their mobile communication capabilities and their primary purpose is to provide voice and data communications. These terminals include smartphones, multimedia phones, feature phones, and low-end phones.

[0150] (2) Ultra-mobile personal computer devices: These devices fall under the category of personal computers and have computing and processing capabilities, and generally also have mobile Internet access. These terminals include PDAs, MIDs, and UMPCs.

[0151] (3) Portable entertainment devices: These devices can display and play multimedia content. They include audio and video players, handheld game consoles, e-books, smart toys, and portable car navigation devices.

[0152] (4) Other onboard electronic devices with data interaction functions, such as onboard computer devices installed in vehicles.

[0153] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0154] Through the description of the above embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a general hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the relevant technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A cluster analysis method for electricity consumption behavior based on big data, characterized in that: The method comprises: Active power time series data, operation and maintenance log data, and meteorological monitoring data are collected from each industrial park. The active power time series data is normalized to generate a normalized sequence. The normalized sequence is sliced ​​using sliding windows with a preset step size based on multiple preset window lengths to form a set of subsequences. The operation and maintenance log data and meteorological monitoring data are aligned to corresponding subsequences according to the time window. Inputting the subsequence set and the aligned operation and maintenance log data and meteorological monitoring data into a pre-trained deep encoder to obtain an embedding vector set that integrates multi-source features; Inputting the embedding vector set into a pre-trained Gaussian mixture dynamic clustering model based on the Dirichlet process, calling the cluster number and cluster center parameters obtained by model training, and assigning each embedding vector to a corresponding cluster cluster through a cluster attribution inference method to generate a cluster label for the embedding vector set; The typical electricity usage behavior category of each industrial park is determined according to the clustering labels, and corresponding typical electricity usage behavior labels are generated.

2. The method according to claim 1, characterized in that After assigning each embedding vector to a corresponding cluster according to the cluster center of each cluster to generate a cluster label for the set of embedding vectors, the method further includes: Calculating a distribution distance between the embedding vector set and a historical cluster center; the historical cluster center is a cluster center automatically determined and saved by the Gaussian mixture dynamic clustering model based on the Dirichlet process for a previous batch of embedding vector sets; If the distribution distance does not exceed the preset distance threshold, only updating the historical cluster center position; and If the distribution distance exceeds a preset distance threshold, it is determined that concept drift occurs in the Gaussian mixture dynamic clustering model based on the Dirichlet process, and fine-tuning training is performed on the Gaussian mixture dynamic clustering model based on the Dirichlet process based on the embedding vector set.

3. The method according to claim 2, characterized in that The distribution distance adopts Wasserstein distance.

4. The method according to claim 3, characterized in that The training of the deep encoder includes: Sampling a candidate subsequence set from the historical subsequence set, and determining a representative shape subsequence from the candidate subsequence set based on an entropy gain method; constructing positive and negative sample pairs from the historical subsequence set based on the representative shape subsequence, wherein the subsequence pairs corresponding to the same shape subsequence are used as positive sample pairs, and the subsequence pairs not containing the same shape subsequence are used as negative sample pairs; Based on the positive and negative sample pairs, the depth encoder is trained using a contrast loss function with a temperature adjustment factor to determine network parameters of the pre-trained depth encoder; The entropy gain method is expressed by the following formula: , Where, Candidate subsequence The entropy gain value, is the set of historical subsequences, is the information entropy function, Candidate subsequence an optimal partitioning threshold of the Euclidean distance with other subsequences in the historical subsequence set; Representation and candidate subsequence The Euclidean distance is not greater than the optimal partition threshold The set of historical subsequences of Representation and candidate subsequence The Euclidean distance is greater than the optimal partition threshold The set of historical subsequences of Indicates the number of elements in the collection.

5. The method according to claim 4, characterized in that The training of the Gaussian mixture dynamic clustering model based on the Dirichlet process includes: A pre-trained deep encoder is used to generate a set of historical embedding vectors corresponding to a set of historical subsequences, and a K-Means++ algorithm is used to analyze and initialize cluster centers based on the set of historical embedding vectors. Employ students t The distribution calculates the initial soft attribution of each historical embedding vector to the cluster center: , Where, For the History embedding vector For the first Cluster centers The initial soft attribution of Indicates the The historical embedding vectors generated by the deep encoder, the total number of historical embedding vectors contained in the historical embedding vector set is ; For the The location vectors of cluster centers, is the number of cluster centers; For students t The degrees of freedom hyperparameter of the distribution is used to adjust the fat-tail characteristics of the distribution; is the Euclidean distance operator, represents the index of any cluster center, For the The location vectors of cluster centers; The corresponding target soft attribution degree is calculated according to the initial soft attribution degree. is the initial soft attribution The square of is normalized to obtain; The cluster center position is optimized and updated by minimizing the KL divergence loss function between the initial soft attribution degree and the target soft attribution degree: , Where, Indicates the A history embedding vector generated by the deep encoder, Represents the total number of historical embedding vectors contained in the historical embedding vector set, is the index of the embedding vector; Indicates the The historical embedding vectors are The target soft belonging degree of cluster centers, Indicates the The updated positions of the cluster centers; The Dirichlet process is used to recursively assign a cluster number to each embedding vector, and the number of non-empty clusters is used as the number of clusters in the clustering model. The Dirichlet process is represented by the following Chinese restaurant process: , Where, Indicates the The cluster labels of the embedding vectors, Indicates the The index of the cluster to which the embedding vector is finally assigned, and the corresponding cluster has the highest probability of assignment; Assign probability to clusters, indicating the The embedding vector is assigned to the The probability of a cluster, conditional on the previous The cluster assignment of the embedding vectors is known; Indicates the front The embedding vectors have been assigned to The number of samples in each cluster; represents the concentration hyperparameter of the Dirichlet process, which is used to adjust the probability of new clusters appearing; Determine the preset distance threshold based on the distribution distance between the embedding vector set of the historical batch and the cluster center θ , specifically including: Calculate the Wasserstein distance between the historical embedding vector set of each historical batch and the cluster center, and update the preset distance threshold based on the 95% quantile of the distance distribution of all historical batches θ .

6. The method according to claim 5, characterized in that The initializing cluster centers by analyzing the historical embedding vector set through the K-Means++ algorithm includes: Randomly select an embedding vector from the historical embedding vector set as the first cluster center; For each embedding vector in the historical embedding vector set that has not been selected as a cluster center, the probability of the embedding vector being selected as the next cluster center is determined according to the proportion of the square of the minimum distance between the embedding vector and the selected cluster center in all the embedding vectors that have not been selected, until the selection is obtained. M Cluster centers, M is the preset number of cluster centers.

7. The method according to claim 1, characterized in that The pre-trained deep encoder includes: Multiple parallel one-dimensional convolutional layers, each with a different convolution kernel width, are used to extract multi-scale temporal features of the input; After the output results of the multiple one-dimensional convolutional layers are spliced, at least one residual block including a dilated convolution and a channel attention mechanism is input to fuse and enhance the feature expressions of the power data, operation and maintenance log data, and meteorological monitoring data; The output of the residual block is processed by global average pooling and a fully connected layer to obtain an embedding vector that fuses multi-source features.

8. The method according to claim 1, characterized in that The cluster attribution inference method softly assigns the attribution of the embedding vector by combining the logarithmic probability expectation of the cluster mixing weight and the logarithmic probability expectation of the Gaussian likelihood of the embedding vector relative to each cluster center, thereby generating the cluster label of the embedding vector set: For each embedding vector, calculate its The responsibility of each cluster: , Where, Indicates the The embedding vector of the fusion multi-source features has an embedding dimension of ; Indicates the The embedding vector for The responsibility of a cluster or cluster center satisfies ; Indicates the The mixing weights of the clusters are derived from the broken stick process construction of the truncated Dirichlet process, satisfying ; For the The mean vector of each cluster represents the location of the cluster center; For the The covariance matrix of a cluster, describing the distribution shape of the cluster; Indicates proportional to the sign; represents the Gaussian probability density function; is the expectation of the weight of the folded stick process, which represents the random variable The logarithmic expectation under the variational distribution is used to evaluate the credibility of cluster weights: , Where, Indicates the first The two posterior parameters of the Beta distribution control The shape of the distribution and the probability of generating new clusters; represents the Digamma function, Indicates the number of nodes from cluster 1 to cluster The cumulative cluster index of the cluster, and Indicates the Two posterior parameters of the Beta distribution; is the expectation of the log-likelihood of the Gaussian distribution: , Where, Indicates the mean and covariance The multivariate Gaussian probability density function described; represents the vector transpose operation, For the The precision matrix of clusters is defined as ; represents the number of embedding dimensions of the embedding vector, Indicates the The expected value of the logarithmic determinant of the covariance matrix of the clusters; Each embedding vector Assign to The largest cluster : , Where, Indicates that among all clusters, the degree of belonging is selected Largest cluster index ; Until each embedding vector in the embedding vector set is assigned a unique cluster label.

9. The method according to claim 1, characterized in that Determining the typical electricity usage behavior category of each industrial park based on the cluster labels and generating corresponding typical electricity usage behavior labels includes: For all subsequences in the same industrial park, the frequency of occurrence of each cluster in the industrial park is counted according to the cluster label, and the preset number of clusters with the highest occurrence frequency are regarded as the main clusters of the industrial park; According to the preset mapping relationship between clusters and electricity consumption behavior categories, the user behavior category corresponding to the main cluster is set as a typical electricity consumption behavior label of the corresponding industrial park.

10. A power consumption behavior clustering analysis system based on big data, characterized in that: The system comprises: a data acquisition unit, configured to collect active power time series data, operation and maintenance log data, and meteorological monitoring data from each industrial park, perform normalization processing on the active power time series data to generate a normalized sequence, perform sliding window slicing on the normalized sequence based on multiple preset window lengths and with a preset step size to form a set of subsequences, and align the operation and maintenance log data and meteorological monitoring data to corresponding subsequences according to time windows; A feature encoding unit, configured to input the subsequence set and the aligned operation and maintenance log data and meteorological monitoring data into a pre-trained deep encoder to obtain an embedding vector set that fuses multi-source features; a cluster analysis unit, configured to input the embedding vector set into a pre-trained Gaussian mixture dynamic clustering model based on a Dirichlet process to automatically determine the number of clusters corresponding to the embedding vector set and the cluster centers of each cluster, and assign each embedding vector to a corresponding cluster according to the cluster centers of each cluster to generate a cluster label for the embedding vector set; The behavior labeling unit is used to determine the typical electricity consumption behavior category of each industrial park according to the cluster label and generate a corresponding typical electricity consumption behavior label.

Citation Information

Patent Citations

  • Clustering analysis method and device for power consumption behaviors of power customers

    CN113450142A

  • Method and device for training Gaussian mixture model in memory Dirichlet process

    CN117576497A

  • Intelligent water conservancy integrated big data analysis method based on clustering algorithm

    CN120124999A

  • County-wide photovoltaic prediction method based on cluster division and data enhancement

    WO2023201552A1

Cited By

  • Industrial Internet of Things time series data anomaly detection method based on concept drift identification

    CN120873698A

  • Nerve patient rehabilitation training method based on task training and brain-computer interaction

    CN121034540A

  • Hot spot area intelligent identification method based on geographic information and user power consumption data

    CN121051689A

  • Artificial intelligence multi-group pulmonary nodule auxiliary diagnosis system

    CN121075617A

  • Load clustering method and system based on multiple types of parks

    CN121167345A