Load collaborative management method, system and equipment based on multi-dimensional perception

Abstract

The invention discloses a load collaborative management method, system and device based on multi-dimensional perception, and the method comprises the steps: obtaining multi-dimensional data which can affect a load through a deployed edge node; extracting features of the multi-dimensional data in different dimensions to form an initial load feature vector; inputting the initial load feature vector into a pre-trained model to obtain a load feature vector output by the model; clustering the load nodes based on the load feature vectors, and correspondingly forming group-level units by clustering groups obtained by clustering; and selecting a representative node of each group-level unit, receiving a scheduling instruction through the representative nodes, and coordinating the allocation of resources in the group-level units. Due to the fact that the multi-dimensional data including the environment, the user behavior, the equipment state and the like influencing the load are collected, the high-quality data are extracted from the multi-dimensional data, the load feature vector is formed and used for determining the load management decision, and the accuracy of the load management decision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system load management, and particularly relates to a load collaborative management method, system and equipment based on multi-dimensional perception. BACKGROUND

[0002] With the deepening of the construction of new power systems and the continuous improvement of renewable energy penetration, the demand for flexible load-side regulation and management capabilities of power systems is increasingly urgent.

[0003] In order to achieve the purpose of reasonable load management, it is necessary to fully perceive the relevant data of the load. At present, the load management system mainly relies on single-dimensional power load data (such as historical load values) to make load management decisions. However, the actual operation of the load is often affected by other factors, so the load management scheme made based on single-dimensional power load data has the problem of inaccurate decision-making. SUMMARY

[0004] The present application aims to provide a load collaborative management method, system and equipment based on multi-dimensional perception, which makes load prediction through multi-dimensional data obtained, and then makes scheduling instructions for load management, solving the problem of inaccurate decision-making of the load management scheme made based on single-dimensional power load data.

[0005] The present application is implemented by the following technical solutions:

[0006] The first aspect of the present application provides a load collaborative management method based on multi-dimensional perception, comprising:

[0007] Obtaining multi-dimensional data capable of affecting the load through the deployed edge node; the multi-dimensional data at least includes power load data, environmental parameter data, user behavior data and device state data;

[0008] Extracting the features of the multi-dimensional data in different dimensions to form an initial load feature vector;

[0009] Inputting the initial load feature vector into a pre-trained model to obtain a load feature vector output by the model; the load feature vector is a vector for feature screening and anomaly detection of the initial load feature vector;

[0010] Based on the load feature vector, clustering the load nodes, and corresponding each cluster group to form a group-level unit;

[0011] Selecting a representative node of each group-level unit, receiving a scheduling instruction through the representative node, and coordinating the allocation of resources in the group-level unit; the scheduling instruction is determined based on the predicted load of the load feature vector.

[0012] In an implementable manner, the method further comprises:

[0013] Based on the preset optimization dimension, a resource allocation optimization target is determined;

[0014] Based on the resource allocation optimization target, an allocation strategy is determined, and the scheduling instruction is formed;

[0015] The allocation strategy is deployed based on a preset component drag-and-drop interface and a strategy template library.

[0016] In an implementable manner, the method further comprises:

[0017] An evaluation system of multi-dimensional indexes is established to evaluate the determined allocation strategy and determine an optimal strategy;

[0018] Based on a preset reinforcement learning algorithm, parameters of the optimal strategy are learned, and the parameters are used as parameter adjustment basis of the load prediction model.

[0019] In an implementable manner, the extracting of the features of the multi-dimensional data in different dimensions to form an initial load feature vector comprises:

[0020] Through a long short-term memory network, time sequence features of the load are extracted from the multi-dimensional data;

[0021] By using a graph convolution network, topological features of the load are extracted from the multi-dimensional data;

[0022] By using a clustering algorithm, behavior features of the user are extracted from the multi-dimensional data;

[0023] The time sequence features, the topological features and the behavior features are weighted and fused to form an initial load feature vector.

[0024] In an implementable manner, the inputting of the initial load feature vector into a pre-trained model to obtain a load feature vector output by the model comprises:

[0025] The initial load feature vector is input into a pre-trained feature extraction model to obtain an intermediate feature vector containing important features output by the feature extraction model;

[0026] The intermediate feature vector is input into a pre-trained anomaly detection model to obtain a load feature vector after repair of abnormal data output by the anomaly detection model.

[0027] In an implementable manner, the method further comprises:

[0028] The clusters obtained by clustering are monitored from within-cluster response variance, response deviation and communication quality;

[0029] When any monitoring item reaches a corresponding condition, and the function value of the re-clustering decision function reaches a preset threshold, a re-clustering operation is triggered.

[0030] In an available implementation, the selecting of the representative node of each group-level unit specifically includes:

[0031] An evaluation system is designed, which takes response stability, communication capability, computing capability and reliability as evaluation indexes, to evaluate each load node in the group-level unit.

[0032] The load node with the maximum weighted value of each evaluation index is selected as the representative node of the group-level unit.

[0033] The second aspect of the application provides a load collaborative management system based on multi-dimensional perception, comprising:

[0034] A data acquisition unit is configured to acquire multi-dimensional data capable of affecting the load through the deployed edge node; the multi-dimensional data at least includes power load data, environmental parameter data, user behavior data and device state data.

[0035] A feature extraction unit is configured to extract features of the multi-dimensional data in different dimensions to form an initial load feature vector.

[0036] A feature processing unit is configured to input the initial load feature vector into a pre-trained model to obtain a load feature vector output by the model; the load feature vector is a vector for feature screening and anomaly detection on the initial load feature vector.

[0037] A clustering unit is configured to cluster load nodes based on the load feature vector, and form each group-level unit corresponding to each clustering group obtained by clustering.

[0038] A resource allocation unit is configured to select a representative node of each group-level unit, receive a scheduling instruction through the representative node, and coordinate allocation of resources in the group-level unit; the scheduling instruction is determined based on the predicted load of the load feature vector.

[0039] The third aspect of the application provides an electronic device, comprising: a processor, a memory and a program or instruction stored on the memory and executable on the processor, wherein the program or instruction is executed by the processor to implement the steps of the above method.

[0040] The fourth aspect of the application provides a storage medium, comprising: a program or instruction stored on the storage medium, wherein the program or instruction is executed by a processor to implement the steps of the above method.

[0041] Compared with the prior art, the application has the following advantages and beneficial effects:

[0042] The embodiment of the present application obtains multi-dimensional data capable of affecting the load, and performs important feature screening and abnormality detection on the multi-dimensional data to obtain a high-quality load feature vector. The load prediction is performed based on the obtained load feature vector, and then the scheduling instruction corresponding to the load management decision is determined. Since the multi-dimensional data including the environment, user behavior and device state affecting the load is collected, and the high-quality data is extracted from the multi-dimensional data to form the load feature vector, the load management decision is determined, and the accuracy of the load management decision is improved. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor. In the drawings:

[0044] Figure 1 A flowchart of a load management method provided by the embodiment of the present application;

[0045] Figure 2 A structural schematic diagram of a load management system provided by the embodiment of the present application;

[0046] Figure 3 A structural schematic diagram of a computing device provided by the embodiment of the present application. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of the present application more clear, the following will further describe the present application in detail with reference to the embodiments and drawings. The exemplary embodiments of the present application and their descriptions are only used to explain the present application, and should not be regarded as a limitation on the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0048] Those skilled in the art can know that, with the development of technology and the appearance of new scenarios, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

[0049] The terms "include" and "have" and any variations thereof in the specification and claims of the present application and the above drawings are intended to cover the non-exclusive inclusion, so that the processes, methods, systems, products or devices containing a series of units do not have to be limited to those units, but can include other units not clearly listed or inherent to these processes, methods, products or devices.

[0050] Embodiment one

[0051] Embodiment 1 of this application provides a load collaborative management method based on multi-dimensional perception, which solves the problem of inaccurate decision-making in load management schemes based on single-dimensional power load data.

[0052] The subject executing this method can be any computing device capable of implementing the method, such as a server, mobile phone, personal computer, smart wearable device, smart robot, etc.

[0053] Furthermore, the embodiments of this application do not limit the execution order of different steps. When using the method provided in the embodiments of this application, the execution order of different steps can be adjusted according to actual needs.

[0054] For ease of description, the following uses a load collaborative management device based on multi-dimensional perception as the execution subject of this method to provide a detailed description of the method provided in this application embodiment.

[0055] like Figure 1 The diagram shown is a flowchart illustrating a load coordination management method based on multi-dimensional perception, provided in an embodiment of this application. The method includes the following steps 11 to 15:

[0056] Step 11: Obtain multi-dimensional data that can affect the load through the deployed edge nodes.

[0057] Deploy edge intelligent gateways in areas with abundant resources. As edge nodes, integrate multiple communication interfaces to collect multi-dimensional data in a unified manner and form dataset D.

[0058] The edge intelligent gateway G(i) integrates a communication interface I, which can be I = {Ethernet, WiFi, 4G / 5G, LoRa, Zigbee}. Ethernet represents the Ethernet communication interface, WiFi represents the wireless local area network communication interface, 4G / 5G represents the cellular mobile network communication interface, LoRa represents the low-power wide area network communication interface, and Zigbee represents the Zigbee wireless communication interface.

[0059] The collected multidimensional data includes at least power load data. Environmental parameter data User behavior data and device status data .

[0060] The edge gateway integrates an ARM64 processor and a GPU (Graphics Processing Unit) acceleration module, and uses a timestamp synchronization function. and data format normalization function These are used to unify the time base for multidimensional data, and to standardize the format and range of data, respectively.

[0061] At the edge node side, PTP (Network Time Protocol) / NTP (Precision Time Protocol) time synchronization is performed (drift <1ms). Missing / outlier points are linearly or spline interpolated and assigned quality labels q∈{0,1}. The units are unified to kW / ℃ / % / dBm. A unified time axis T and sliding window W=[t−15min,t] are constructed to ensure semantic synchronization and comparability of cross-source data.

[0062] Step 12: Extract the features of the multidimensional data in different dimensions to form an initial load feature vector.

[0063] Feature extraction is performed on multidimensional data from the time, space, and behavioral dimensions, specifically including:

[0064] Using LSTM Long Short-Term Memory Network , This represents the load value at time t. Representing the hidden state vector of the previous time step, the temporal features of the load are extracted from the multidimensional data to identify periodic patterns. , These represent daily, weekly, and seasonal periods, respectively. A bidirectional LSTM is used to extract the hidden states of the time series. and with attention Aggregation .

[0065] Using graph convolutional networks The topological features of the load are extracted from the multidimensional data, where A is the adjacency matrix and X is the node feature. A two-layer GCN is constructed based on the topology G(A,X) to extract spatial correlation features F_S, and node aggregation is performed using average pooling.

[0066] K-means clustering algorithm The process involves analyzing user patterns, where U is the feature matrix and k is the cluster number, to extract user behavioral features from the multidimensional data. Session segmentation is performed on the U behaviors, and K-means clustering (k=6) is conducted to obtain behavioral clusters, forming one-hot vectors. This is used to express the user / device's operating mode and scene sequence characteristics.

[0067] The implementation process for feature fusion can be: through an attention mechanism. The time-series features, topological features, and behavioral features are weighted and fused to form an initial load feature vector. where i ∈ {T, S, B}, Indicates weight, Indicates characteristics.

[0068] Another approach to feature fusion is to perform multimodal fusion using linear projection and LayerNorm to obtain... .

[0069] Dimensions with poor quality (q=0) are weighted down by 0.5 to enhance robustness in noisy and missing test scenarios; pruned and quantized models are deployed on the edge lightweight inference engine to ensure inference latency. <5ms, configure historical / prediction window =15min =1min, feature update frequency =1 / 60Hz.

[0070] Step 13: Input the initial load feature vector into the pre-trained model to obtain the load feature vector output by the model; the load feature vector is a vector for feature filtering and anomaly detection of the initial load feature vector.

[0071] The initial load feature vector is input into a pre-trained feature extraction model to obtain an intermediate feature vector containing important features output by the feature extraction model, specifically including:

[0072] A lightweight deep learning inference engine is deployed at edge nodes, and model pruning and quantization compression techniques are used to optimize the model, achieving millisecond-level real-time feature extraction. A sliding window mechanism is established, with a 15-minute history window and a 1-minute prediction window, to achieve real-time refresh of feature vectors.

[0073] By integrating incremental learning algorithms, the model parameters of the feature extraction model are updated online to adapt to environmental changes. For the trained feature extraction model, a feature importance evaluation mechanism is deployed to screen important features (or high-contribution features) and intermediate feature vectors, reducing redundant calculations and thus optimizing the efficiency of subsequent calculations.

[0074] The intermediate feature vector is input into a pre-trained anomaly detection model to obtain the load feature vector output by the anomaly detection model after repairing the anomaly data, specifically including:

[0075] Anomaly detection is achieved through a pre-trained anomaly detection model (isolated forest algorithm) using intermediate feature vectors. Anomaly types are accurately identified through a three-layer detection strategy, and data quality is ensured through a multi-strategy repair mechanism.

[0076] An unsupervised anomaly detection model based on the Isolation Forest algorithm was constructed. Abnormal scores ,in For path length, ) is a standardization constant. This represents the average path length of data point x across all isolated trees.

[0077] Three-layer detection strategy Basic anomaly detection threshold Statistical anomaly detection Chi-square test Spatiotemporal correlation detection Correlation coefficient .

[0078] One implementation of the three-layer detection strategy is as follows: Based on the anomaly score s(x,n) output by the isolated forest, a basic threshold θ1 is set (e.g., θ1=0.8). When s(x,n)≥θ1, it is marked as "suspected anomaly" and proceeds to the next layer of detection; otherwise, it is judged as normal data. For "suspected anomaly" data, a chi-square test is used to verify whether it deviates from the statistical distribution. Assuming that normal data follows a certain distribution (e.g., Gaussian distribution), the deviation of the suspected anomaly data from this distribution is calculated. Set threshold ;when At that time, it was confirmed as a "statistical anomaly". For the "statistically anomaly" data, its correlation with spatiotemporal correlation data was analyzed. (Same as the load of adjacent equipment in the same area, historical data of the same period), when When confirmed as a genuine anomaly, among which, To analyze abnormal data and historical data from the same period, To set a threshold.

[0079] Establish a set of repair strategies Multi-strategy repair mechanisms, including similarity interpolation Prediction and estimation Expert Rules Repair and ensure that the data quality meets processing requirements.

[0080] For short-term single-point anomalies (such as missing / incorrect load data for a certain 10 minutes), and there are reference data with similar characteristics (such as the same equipment's data from the same period 3 days ago, or the same type of equipment's data from the same period), the similarity between the anomaly and the reference data (such as Euclidean distance or cosine similarity) can be calculated, and the K most similar data can be selected for weighted interpolation.

[0081] For continuous anomalies across multiple time periods (such as sensor malfunction causing data anomalies within 1 hour) and where the data exhibits strong temporal regularity, a prediction model (such as LSTM or ARIMA) can be trained based on historical normal data to predict reasonable values ​​for the abnormal time periods.

[0082] For special anomalies (such as sudden changes caused by equipment start-up and shutdown, or special load patterns during holidays), the model may have difficulty capturing the patterns. However, it is possible to pre-set repair rules based on domain knowledge (such as "the load should rise rapidly from 0 to 80%-120% of the rated power when the equipment starts up" and "the average load on holidays is 60% of that on weekdays") and correct the abnormal values ​​according to the rules.

[0083] The above describes the acquisition of multidimensional data and the preprocessing of the acquired multidimensional data. Addressing the issues of high latency in single-dimensional data perception and centralized processing in existing technologies, an innovative edge intelligent fusion perception technology for multi-source heterogeneous data is proposed. By deploying edge computing nodes, real-time acquisition and millisecond-level processing of multidimensional data such as power load, environmental parameters, user behavior, and equipment status are achieved. A deep fusion feature model of time series, spatial topology, and behavioral patterns is constructed, and an intelligent anomaly detection algorithm based on isolated forests is adopted, significantly improving the accuracy of data perception and the real-time performance of processing.

[0084] The following is a specific example of applying the above implementation process:

[0085] Specific application scenarios include: targeting single or multi-campus data center clusters where the load consists of IT equipment and cooling systems. Peak shaving and demand response must be achieved without affecting the business SLA (Service Level Agreement) and the boundaries of the data center environment (temperature, humidity, airflow). Typical facilities include CRAC (Computer Room Air Conditioner), Chillers, cooling towers, water pumps, PDUs (Power Distribution Units), UPS (Uninterruptible Power Supply), server clusters, and operation and maintenance systems. The scenario has the following characteristics: (1) The load peak is obvious and related to weather and work batches; (2) There is significant coupling between IT (Information Technology, which refers to information technology-related equipment such as servers and workstations) and cooling, and the migration of heat load will cause the electrical load and temperature control to be linked; (3) The operation constraints are strict, and SLA, redundancy and alarm thresholds limit the adjustable range; (4) Events are frequent, and operation and maintenance operations, task migration and expansion and contraction will cause short-term load disturbances.

[0086] The multidimensional data acquisition and processing process includes:

[0087] Edge multi-source acquisition and preprocessing: Deploy edge smart gateways on the interface sides of BMS (Building Management System), CRAC / Chiller, cooling towers and water pumps, PDU / UPS, server monitoring and weather stations, etc., to uniformly access multi-source data, complete time synchronization, unit standardization and missing / anomaly handling, generate quality tags and establish a unified time axis and sliding window (15 minutes of history, 1 minute of prediction), to provide consistent and comparable input for subsequent fusion.

[0088] Spatiotemporal behavioral feature fusion modeling: The time dimension extracts the change patterns of IT load rate, return air temperature, chilled water supply and return temperature, and data center power; the spatial dimension identifies thermal and electrical coupling relationships based on the data center-rack-loop topology; the behavioral dimension identifies event patterns such as operation and maintenance operations, batch processing jobs, and load migration; feature fusion and standardization are completed within the edge lightweight engine to ensure millisecond-level updates and stable output for prediction, clustering, and control.

[0089] Anomaly detection and data quality assurance: Anomalies are identified using rules and statistical models, including those for exceeding limits, breaking points, and sudden changes. Anomalies are then categorized as minor / moderate / severe and handled accordingly: imputation, resampling and labeling, channel isolation, and rollback. With a target data quality score of at least 85%, completeness, accuracy, consistency, and timeliness are continuously evaluated to ensure reliable input.

[0090] Taking a hyperscale data center as an example, the latency of a single edge inference is less than 5 milliseconds, and features are updated stably on a minute-by-minute basis; data consistency is no less than 95%, and the quality score is no less than 85%. Without violating SLA and environmental boundaries, the fused features can stably drive subsequent clustering and collaborative optimization, support peak reduction and valley replenishment, and ensure operational reliability and economy.

[0091] Step 14: Based on the load feature vector, cluster the load nodes and form each cluster group into a corresponding group-level unit.

[0092] Based on the load feature vector, the incremental DBSCAN (Density-Based Spatial Clustering of Applications with Noise) online clustering algorithm is used to cluster the load nodes.

[0093] DBSCAN is an unsupervised clustering algorithm that groups densely connected points into the same cluster by defining "core points, boundary points, and noise points." Core points are points with a number of points in their ε-neighborhood greater than or equal to MinPts (the core skeleton of the cluster). Boundary points are points with a number of points in their ε-neighborhood less than MinPts but falling within the ε-neighborhood of a core point (dependent on the core point and belonging to the cluster). Noise points are neither core points nor boundary points (isolated points, which can be considered outliers).

[0094] ε represents the radius of the neighborhood, and MinPts represents the minimum number of points.

[0095] An adaptive density threshold adjustment module is designed to dynamically adjust the ε parameter (range [0.1, 0.5]) based on the real-time data distribution, avoiding clustering bias caused by parameter fixation.

[0096] A dynamic boundary point identification strategy is introduced, which accurately classifies core points, boundary points, and noise points by calculating local density reachability and neighborhood stability.

[0097] Deploy an incremental update engine. When new data points are added, avoid fully re-clustering the data; instead, recalculate only the local clustering structure, reducing update time from 30 minutes to 2 minutes. Set a 24-hour sliding window to maintain historical information, balancing clustering stability and adaptability.

[0098] One approach to local clustering is to partition historical data according to temporal or spatial features, build a KD (k-dimensional tree, a spatial indexing structure for high-dimensional data) tree index, and record the core point set, boundary point affiliation, and noise point label for each partition, as well as the density statistics within the partition. When a new data point is added, for the new data point, the historical points in its neighborhood are quickly queried through the KD tree. If the neighborhood contains a core point, it is determined that the new data point may affect the local cluster to which the core point belongs; otherwise, it is only necessary to determine whether the new data point is a noise point. If there is a core point in the neighborhood of the new data point, only the local cluster to which the core point belongs needs to be recalculated, and the core point / boundary point status needs to be updated.

[0099] This embodiment also sets the triggering conditions for re-clustering: monitoring the intra-cluster response variance, response bias, and communication quality of each cluster obtained by clustering; when any monitoring item reaches the corresponding condition and the function value of the re-clustering decision function reaches a preset threshold, the re-clustering operation is triggered.

[0100] Establish a three-tier triggering system In-group response variance Monitoring trigger conditions Response deviation Evaluate Communication quality monitor or .

[0101] In-group response variance Monitoring: Individuals within the same cluster (such as users with the same behavior pattern or devices of the same type) should exhibit consistent responses to the same trigger signals (such as electricity price adjustments or dispatch instructions). Excessive response variance indicates significant differences among individuals within the cluster, suggesting a failure of the clustering structure. The mean response represents the average response value of all individuals (such as users with the same behavior pattern or devices of the same type) within a cluster to a specific trigger signal (such as electricity price adjustment or dispatch command), reflecting the overall response level of the cluster. This represents the standard deviation of the response, characterizing the dispersion (range of fluctuation) between the individual response values ​​and the average response value within a cluster. A smaller value indicates more consistent individual responses, while a larger value indicates more significant individual differences. (Formula) By correlating the standard deviation with the mean using a coefficient of 0.15, it is shown that under normal circumstances, the fluctuation range of the response should be controlled within 15% of the mean; 15% is just an example.

[0102] Response deviation Assessment: The deviation between an individual's actual response and the ideal response predicted by the clustering model. If the deviation is too large, it indicates that the individual has deviated from the original clustering pattern and needs to be re-clustered. The response deviation is a dimensionless percentage indicator that represents the relative deviation between the actual response value and the command value. The larger the value, the more serious the deviation between the actual execution result and the command requirements. It represents the actual response power, which characterizes the actual power adjustment value generated by the equipment or load after receiving the control command; This indicates the power requirement of the instruction, representing the expected power adjustment instruction value issued by the system or dispatch center to the equipment / load.

[0103] Communication quality Monitoring: Clustering relies on high-quality real-time data (such as individual response data and status data). If the communication quality is substandard, or if the data is lost or incorrect, the clustering results will be distorted, and re-clustering needs to be triggered to correct them. Indicates transmission delay. This indicates the packet loss rate.

[0104] Design a re-aggregation decision function ,in These are the weighting coefficients.

[0105] To avoid frequent fluctuations caused by re-aggregation occurring immediately upon triggering of the aforementioned monitoring items, the re-aggregation decision function uses a weighted fusion of re-aggregation costs. Reunification Benefits and cluster stability The decision on whether to perform re-aggregation balances cluster stability and adaptability.

[0106] Among them, re-aggregation cost The normalized index (range [0,1]) quantifies the various costs generated during the re-aggregation process. The higher the cost, the closer the quantified value is to 1, and the smaller its positive contribution to the decision value (or it can be adjusted through negative weights). Re-aggregation benefits. The normalized index (range [0,1]) represents the various benefits derived from quantification after re-aggregation. Higher benefits are associated with a value closer to 1, indicating a greater positive contribution to the decision value. Cluster stability. The normalized index (range [0,1]) quantifies the stability of the current cluster structure. The higher the stability, the closer the quantified value is to 1, indicating that the current cluster does not need to be adjusted and the stronger the support for "non-reclustering".

[0107] Establish adaptive learning algorithm With a learning rate γ=0.01, parameter self-optimization is achieved. This is accomplished by real-time optimization of the trigger threshold parameter. This allows it to dynamically match the dynamic changes in the system. Among them, This represents the trigger threshold parameter at the current moment. This represents the updated threshold parameter at the next time step. This represents the gradient of the objective function J with respect to the threshold parameter. The adaptive learning algorithm dynamically adjusts the re-aggregation trigger threshold θtrigger to ensure accurate triggering when re-aggregation should occur and avoids false triggering when it shouldn't, thus balancing the adaptability and stability of clustering.

[0108] Step 15: Select a representative node for each group-level unit, receive scheduling instructions through the representative node, and coordinate the allocation of resources in the group-level unit; the scheduling instructions are determined by the load predicted based on the load feature vector.

[0109] The selection of representative nodes for each group-level unit specifically includes: designing an evaluation system that includes response stability, communication capability, computing capability, and reliability as evaluation indicators, and evaluating each load node within the group-level unit; and selecting the load node with the largest weighted value of each evaluation indicator as the representative node of the group-level unit.

[0110] The design incorporates a four-dimensional evaluation index system: response stability (historical deviation <5%), communication capability (latency <2 seconds, signal strength >−70dBm), computing power (clock frequency >1GHz, memory >2GB), and reliability (online duration >95%, failure rate <1%). AHP is used to determine weights, and TOPSIS is used to calculate the comprehensive score. The optimal node is then selected as the representative node.

[0111] For each node in the cluster obtained by clustering, calculate its response stability, communication capability, computing capability, and reliability. Select the node with the highest weighted value in the cluster as the representative node.

[0112] This embodiment also includes: constructing a three-tiered control architecture of "dispatch center - group representative - terminal device", clearly defining the responsibilities of each level. The dispatch center is responsible for issuing power adjustment commands and real-time price signals; the group representative node is responsible for receiving commands from higher levels and coordinating the optimal allocation of resources within the group; and the terminal device is responsible for executing specific response actions and providing timely feedback on its operating status. An implementation process for the intra-group terminal bidding mechanism is designed. Each terminal device submits standardized bid information, including response capacity, response price, and response time, to the group representative node based on its own response cost assessment, current operating status analysis, and user comfort preference settings. The group representative node runs an economic scheduling optimization algorithm, comprehensively considering multiple factors such as terminal response cost, response speed, and device reliability, and determines the response combination scheme with the lowest cost and best effect through optimization calculation, and issues execution commands to the winning terminal. A lightweight intra-group communication system based on the MQTT protocol is established, supporting three flexible communication modes: groupcast, multicast, and on-demand. Command confirmation and status feedback mechanisms are configured to ensure the timeliness, reliability, and full traceability of control command transmission and response effects.

[0113] To address the limitations of existing offline and static clustering technologies, an adaptive dynamic clustering technique based on an improved incremental DBSCAN is proposed. This technique achieves real-time clustering updates in the high-dimensional feature space through an online learning mechanism, designs a dynamic re-clustering mechanism with multi-dimensional triggering conditions such as intra-cluster variance threshold and response bias, and establishes a two-layer control structure combining cluster-level representative selection and terminal bidding to achieve refined clustering control and high-precision response of load resources.

[0114] In one feasible implementation, the method of this embodiment further includes constructing a microservice-based resource orchestration and spatiotemporal multi-layer collaborative optimization strategy. Addressing the issues of tight coupling and single optimization levels in existing system architectures, a microservice-based resource orchestration and spatiotemporal multi-layer collaborative optimization technology is proposed. A Resource Capability Description Language (RDL) is designed to standardize the encapsulation of heterogeneous load resources; a drag-and-drop visual orchestrator is developed to support rapid strategy combination and second-level deployment; a multi-dimensional collaborative optimization framework is constructed, encompassing second-level, minute-level, hour-level time layers and device-level, group-level, region-level, and system-level spatial layers, integrating a prediction-feedback closed-loop mechanism to achieve adaptive adjustment of strategy parameters.

[0115] In this implementation, the first step is to build a Resource Description Language (RDL) and a standardized microservice encapsulation system.

[0116] A resource capability description language based on JSON Schema is designed to uniformly define the basic attributes (equipment type, rated power, adjustment range), dynamic characteristics (response time, adjustment accuracy, available time period), constraints (start-stop time, comfort constraints, protection constraints), and interface specifications (data interface, control interface, status interface) of load resources. Docker container technology is used to standardize the encapsulation of heterogeneous load resources. Each resource type (each function of the system corresponds to one resource type) corresponds to an independent microservice container, integrating four core modules: resource simulator, control logic, status monitoring, and communication interface. A Consul-based microservice registration and discovery mechanism is established to achieve dynamic registration, health checks, fault recovery, and load balancing, supporting hot-swapping and elastic scaling. An API gateway is deployed to uniformly manage microservice interfaces, providing access entry points, authentication, rate limiting control, and security protection.

[0117] Develop a drag-and-drop visual programmable programmer and a rapid strategy deployment platform. Specifically, this could involve: developing a drag-and-drop visual orchestration interface based on React + D3.js to provide a WYSIWYG strategy design experience; designing a rich component library including load resource components (air conditioning, energy storage, photovoltaics, electric vehicles, etc.), logic control components (conditional judgment, loop control, data processing), connection components (data flow, control flow, feedback flow), and constraint components (time constraints, capacity constraints, user constraints); implementing a strategy template library management system with pre-built templates for common strategies such as peak shaving and valley filling, emergency response, and economic optimization, supporting import / export, version control, and permission management; and designing a strategy compilation engine to convert visual orchestration into Resource Instruction Language (RSL), enabling second-level strategy distribution and hot-update deployment via the gRPC protocol.

[0118] Constructing a spatiotemporal multidimensional collaborative optimization framework includes: determining resource allocation optimization objectives based on preset optimization dimensions; determining allocation strategies based on the resource allocation optimization objectives and forming the scheduling instructions; the allocation strategies are deployed based on a preset component drag-and-drop interface and strategy template library.

[0119] A spatiotemporal four-dimensional collaborative optimization framework is constructed. In the time dimension, three layers are designed: second-level real-time optimization (response to sudden events), minute-level rolling optimization (adaptation to load changes), and hour-level predictive optimization (daily scheduling plan). In the spatial dimension, four layers of optimization are established: equipment level, group level, region level, and system level.

[0120] The algorithm is designed as a hierarchical structure, with the upper layer determining the target allocation and the lower layer solving the strategy. Lagrange dual decomposition is used to handle inter-layer coupling constraints, and global optimum is achieved through iterative coordination. The scheduling instructions corresponding to the solved strategy are sent to representative nodes; the strategy is deployed via a pre-defined component drag-and-drop interface and strategy template library.

[0121] An adaptive adjustment mechanism for the objective function is established to dynamically adjust the weights of economy, reliability, and comfort based on real-time electricity prices, operating status, and user preferences, thereby achieving a dynamic balance among multiple objectives.

[0122] In one feasible implementation, the method of this embodiment further includes: establishing a multi-dimensional indicator evaluation system to evaluate the determined allocation strategy and determine the optimal strategy; learning the parameters of the optimal strategy based on a preset reinforcement learning algorithm, and using the parameters as the basis for adjusting the parameters of the load forecasting model.

[0123] An integrated multi-timescale prediction system is used to predict allocation strategies: short-term prediction (15 minutes to 2 hours) uses an LSTM network, medium-term prediction (1 day to 1 week) uses a support vector machine, and long-term prediction (1 month to 1 year) uses a Kalman filter.

[0124] Establish a strategy effectiveness evaluation system to monitor four dimensions of indicators in real time: response accuracy, execution speed, user satisfaction, and economic benefits. The overall score should be ≥85%.

[0125] We design a parameter adaptation mechanism based on reinforcement learning, modeling parameter adjustment as a Markov decision process. We employ the Q-learning algorithm to learn the optimal parameter policy, enabling the algorithm to self-evolve. We establish a policy knowledge base to record the historical policy execution effects and applicable scenarios. Through case-based reasoning, we recommend the optimal policy for new scenarios, achieving intelligent recommendation and continuous optimization.

[0126] This application's embodiments acquire multidimensional data that can affect load, and perform important feature filtering and anomaly detection on the multidimensional data to obtain a high-quality load feature vector. Load forecasting is then performed based on the obtained load feature vector, thereby determining the corresponding scheduling instructions for load management decisions. Because multidimensional data affecting load, including environmental factors, user behavior, and equipment status, are collected, and high-quality data is extracted from this multidimensional data to form a load feature vector for determining load management decisions, the accuracy of load management decisions is improved.

[0127] Example 2

[0128] To address the problem of inaccurate decision-making in load management schemes based on single-dimensional power load data in existing technologies, and based on the same inventive concept as Embodiment 1, this application also provides a load collaborative management system based on multi-dimensional perception.

[0129] The specific structural diagram of the system is as follows: Figure 2 As shown, it includes the following functional units 21-25:

[0130] The data acquisition unit 21 is used to acquire multidimensional data that can affect the load through deployed edge nodes; the multidimensional data includes at least power load data, environmental parameter data, user behavior data and equipment status data.

[0131] The feature extraction unit 22 is used to extract features of the multidimensional data in different dimensions to form an initial load feature vector.

[0132] The feature extraction unit is specifically used to: extract the temporal features of the load from the multidimensional data using a long short-term memory network; extract the topological features of the load from the multidimensional data using a graph convolutional network; extract the user's behavioral features from the multidimensional data using a clustering algorithm; and perform weighted fusion of the temporal features, topological features, and behavioral features to form an initial load feature vector.

[0133] The feature processing unit 23 is used to input the initial load feature vector into a pre-trained model to obtain the load feature vector output by the model; the load feature vector is a vector for feature filtering and anomaly detection of the initial load feature vector.

[0134] The feature processing unit is specifically used to: input the initial load feature vector into a pre-trained feature extraction model to obtain an intermediate feature vector containing important features output by the feature extraction model; and input the intermediate feature vector into a pre-trained anomaly detection model to obtain a load feature vector after repairing the anomaly data output by the anomaly detection model.

[0135] Clustering unit 24 clusters load nodes based on load feature vectors, and forms corresponding group-level units for each cluster obtained from the clustering.

[0136] The clustering unit is specifically used to: monitor the intra-cluster response variance, response bias, and communication quality of each cluster obtained by clustering; and trigger a re-clustering operation when any monitoring item reaches the corresponding condition and the function value of the re-clustering decision function reaches a preset threshold.

[0137] Resource allocation unit 25 is used to select a representative node for each group-level unit, receive scheduling instructions through the representative node, and coordinate the allocation of resources in the group-level unit; the scheduling instructions are determined by the load predicted based on the load feature vector.

[0138] The resource allocation unit is specifically used to: design an evaluation system that includes response stability, communication capability, computing capability and reliability as evaluation indicators, evaluate each load node in the group-level unit; and select the load node with the largest weighted value of each evaluation indicator as the representative node of the group-level unit.

[0139] The load management system in this embodiment also includes a strategy deployment unit and a strategy evaluation unit.

[0140] The strategy deployment unit is used to determine resource allocation optimization objectives based on preset optimization dimensions; determine allocation strategies based on the resource allocation optimization objectives and generate scheduling instructions; and deploy the allocation strategies based on a preset component drag-and-drop interface and strategy template library. The strategy evaluation unit is used to establish a multi-dimensional indicator evaluation system to evaluate the determined allocation strategies and determine the optimal strategy; learn the parameters of the optimal strategy based on a preset reinforcement learning algorithm, and use the parameters as the basis for adjusting the parameters of the load forecasting model.

[0141] This application's embodiments acquire multidimensional data that can affect load, and perform important feature filtering and anomaly detection on the multidimensional data to obtain a high-quality load feature vector. Load forecasting is then performed based on the obtained load feature vector, thereby determining the corresponding scheduling instructions for load management decisions. Because multidimensional data affecting load, including environmental factors, user behavior, and equipment status, are collected, and high-quality data is extracted from this multidimensional data to form a load feature vector for determining load management decisions, the accuracy of load management decisions is improved.

[0142] Based on the same inventive concept as the foregoing embodiments of this application, this application also provides a computing device.

[0143] like Figure 3 As shown, the computing device includes a memory 31 and a processor 32. The memory 31 can be configured to store various other data to support operation on the electronic device. Examples of such data include instructions for any application or method used to operate on the electronic device. The memory 31 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0144] The processor 32, coupled to the memory 31, is used to execute a computer program stored in the memory 31 for performing a method for aggregating feasible domains of a virtual power plant as described in the foregoing embodiments.

[0145] When processor 32 executes the computer program to perform a load management method, it acquires multidimensional data that can affect the load, performs important feature filtering and anomaly detection on the multidimensional data, and obtains a high-quality load feature vector. Based on the obtained load feature vector, load forecasting is performed, and then the corresponding scheduling instructions for load management decisions are determined. Because multidimensional data affecting the load, including environmental factors, user behavior, and equipment status, are collected, and high-quality data is extracted from the multidimensional data to form a load feature vector for determining load management decisions, the accuracy of load management decisions is improved.

[0146] When the processor 32 executes the computer program in the memory 31, in addition to the functions described above, it can also perform other functions, as detailed in the descriptions of the preceding embodiments.

[0147] Furthermore, such as Figure 3 As shown, the computing device also includes other components such as a display 34, a communication component 33, a power supply component 35, and an audio component 36. Figure 3 The diagram only shows some components and does not mean that the computing device includes only these components. Figure 3 The components shown.

[0148] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a computer, can implement the methods provided in the above embodiments.

[0149] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. 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 can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0150] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, 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, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments.

[0151] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A load collaborative management method based on multi-dimensional perception, characterized in that, include: By deploying edge nodes, we can obtain multidimensional data that can affect the load; The multidimensional data includes at least power load data, environmental parameter data, user behavior data, and equipment status data; Extract the features of the multidimensional data in different dimensions to form an initial load feature vector; The initial load feature vector is input into a pre-trained model to obtain the load feature vector output by the model; the load feature vector is a vector for feature filtering and anomaly detection of the initial load feature vector. Based on the load feature vector, the load nodes are clustered, and the clusters obtained are used to form corresponding group-level units. A representative node is selected for each group-level unit. The representative node receives scheduling instructions and coordinates the allocation of resources within the group-level unit. The scheduling instructions are determined by the load predicted based on the load feature vector.

2. The method according to claim 1, characterized in that, The method further includes: Based on the preset optimization dimensions, determine the resource allocation optimization objectives; Based on the resource allocation optimization objective, an allocation strategy is determined, and the scheduling instruction is generated. The allocation strategy is deployed based on a preset component drag-and-drop interface and a strategy template library.

3. The method according to claim 2, characterized in that, The method further includes: Establish a multi-dimensional evaluation system to evaluate the determined allocation strategy and determine the optimal strategy; The parameters of the optimal strategy are learned based on a preset reinforcement learning algorithm, and these parameters are used as the basis for adjusting the parameters of the load forecasting model.

4. The method according to claim 1, characterized in that, The step of extracting features from the multidimensional data in different dimensions to form an initial load feature vector includes: Temporal features of the load are extracted from the multidimensional data using a long short-term memory network. A graph convolutional network is used to extract the topological features of the load from the multidimensional data; Clustering algorithms are used to extract user behavioral features from the multidimensional data; The time-series features, topological features, and behavioral features are weighted and fused to form an initial load feature vector.

5. The method according to claim 1, characterized in that, The step of inputting the initial load feature vector into a pre-trained model to obtain the load feature vector output by the model specifically includes: The initial load feature vector is input into a pre-trained feature extraction model to obtain an intermediate feature vector containing important features output by the feature extraction model. The intermediate feature vector is input into a pre-trained anomaly detection model to obtain the load feature vector output by the anomaly detection model after repairing the anomaly data.

6. The method according to claim 1, characterized in that, The method further includes: The clusters obtained by clustering are monitored for intra-cluster response variance, response bias, and communication quality; When any monitored item reaches the corresponding condition and the function value of the re-clustering decision function reaches the preset threshold, the re-clustering operation is triggered.

7. The method according to claim 1, characterized in that, The selection of a representative node for each group-level unit specifically includes: The design incorporates an evaluation system that includes response stability, communication capability, computing power, and reliability as evaluation indicators to assess each load node within the group-level unit. The load node with the largest weighted value among all evaluation indicators is selected as the representative node of the group-level unit.

8. A load collaborative management system based on multi-dimensional perception, characterized in that, include: The data acquisition unit is used to acquire multidimensional data that can affect the load through deployed edge nodes; The multidimensional data includes at least power load data, environmental parameter data, user behavior data, and equipment status data; The feature extraction unit is used to extract features of the multidimensional data in different dimensions to form an initial load feature vector. The feature processing unit is used to input the initial load feature vector into a pre-trained model to obtain the load feature vector output by the model; the load feature vector is a vector for feature filtering and anomaly detection of the initial load feature vector. Clustering units are formed by clustering load nodes based on load feature vectors, and the resulting clusters are then used to form group-level units. The resource allocation unit is used to select a representative node for each group-level unit, receive scheduling instructions through the representative node, and coordinate the allocation of resources in the group-level unit; the scheduling instructions are determined by the load predicted based on the load feature vector.

9. An electronic device, characterized in that, include: It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method as described in any one of claims 1-7.

10. A storage medium, characterized in that, include: The storage medium stores a program or instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1-7.

Images