A dynamic polymer visual multi-level load cooperative management method

Abstract

The application relates to a dynamic polymer visual multi-level load collaborative management method. Firstly, power system operation, user energy consumption behavior and environmental parameters are acquired and structured to obtain a data set; then, dynamic aggregates are generated according to resource characteristics and spatial distribution characteristic level division in the data set; then, time domain and frequency domain characteristic parameters of each level load are extracted to construct a layered load behavior model; then, a multi-objective optimization function is constructed based on the dynamic aggregate topology, the layered load behavior model and constraint conditions to generate a collaborative management strategy; finally, the strategy is optimized by means of real-time feedback, virtual verification and long-term learning mechanism, and layered visual mapping rules are constructed to graphically output the dynamic aggregate, the layered load behavior model and the strategy execution result. Compared with the prior art, the application has the advantages of good data integration performance, strong dynamic adaptability and multi-objective trade-off realization.

Description

Technical Field

[0001] This invention relates to the field of power system load management technology, and in particular to a dynamic aggregate visualization multi-level load collaborative management method. Background Technology

[0002] Current power system load management generally faces core challenges such as low coordination efficiency due to data heterogeneity, the inability of static hierarchical division to adapt to dynamic resource characteristics, and insufficient multi-objective strategy optimization capabilities. In existing technologies, traditional load management systems (such as centralized control based on SCADA) suffer from limited data sources and lack of cross-dimensional information fusion capabilities, making it difficult to achieve linked analysis of user behavior and environmental parameters. Distributed management schemes based on fixed hierarchical divisions (such as hierarchical demand response systems) are constrained by rigid topologies and cannot dynamically adjust hierarchical relationships based on equipment controllability and response timeliness, resulting in low resource coordination efficiency. Single-objective optimization strategies (such as energy efficiency balance or delay suppression alone) are prone to getting trapped in local optima in complex scenarios and lack multi-objective trade-off mechanisms. Furthermore, existing systems generally lack dynamic calibration capabilities, and strategy generation relies on static parameter models, making it impossible to optimize hierarchical division thresholds and coordination rules through real-time feedback.

[0003] Current existing technical solutions include:

[0004] Dynamic Aggregator Management Technology: Some studies use clustering algorithms to divide load aggregates, but these only statically divide the hierarchy based on historical load characteristics, leading to a disconnect between the hierarchy division and actual operating conditions. Multi-Objective Optimization Strategy Generation: Existing research uses weighted summation to merge energy efficiency, delay, and other objectives into a single objective function, but fixed weights result in rigid strategies that cannot adapt to the dynamic game requirements of multi-objective scenarios under disturbances. Hierarchical Load Feature Analysis: Existing methods analyze load patterns using time-domain statistics (such as mean and variance), but they neglect frequency-domain energy entropy and cross-level correlation weights, making it difficult to quantify the characteristics of sudden change response and the coupling strength between levels.

[0005] For example, the invention patent with publication number CN120013088A discloses a multi-level power market user profile dynamic collaborative management method and system. However, it only manages the association between user profiles and electricity consumption model. It does not have resource characteristic quantification standards and dynamic topology generation logic, nor does it have a collaborative management mechanism that combines multi-objective optimization and real-time feedback calibration. Furthermore, it has not built a closed-loop system covering the entire process from data processing, dynamic aggregation, load modeling to strategy optimization. This solution focuses solely on the correlation between user profiles and electricity consumption patterns, failing to establish an integration mechanism to address the heterogeneity of data. This results in fragmented information from multiple sources, hindering data support for cross-dimensional load evolution analysis. Furthermore, its hierarchical division relies on static user type classification, lacking dynamic adjustment logic based on resource characteristics such as equipment controllability and response timeliness. This makes it difficult to adapt to real-time changes in equipment status and energy demand within the power system, potentially leading to redundancy or failure in hierarchical command transmission. Additionally, its strategy optimization focuses solely on the flexibility of power supply strategies, neglecting the dynamic interplay between multiple objectives such as response delay and system stability. It also lacks calibration mechanisms such as virtual verification and long-term learning, resulting in insufficient robustness of the strategy under disturbances such as load surges and equipment failures.

[0006] In summary, existing technologies suffer from fragmented multi-source information due to data heterogeneity (e.g., the lack of correlation between power system operation data and user behavior data), making it difficult to establish cross-dimensional load evolution models and causing collaborative decision-making to rely on locally optimal strategies. Simultaneously, static hierarchical partitioning (e.g., fixed tree topologies) cannot dynamically adapt to changes in resource characteristics such as equipment controllability and response timeliness, resulting in redundancy or failure in instruction transmission between levels. Furthermore, single-objective optimization strategies (e.g., energy efficiency balance only) neglect multi-objective dynamic game dynamics (e.g., the conflict between delay suppression and stability assurance) and lack real-time feedback calibration mechanisms, significantly reducing strategy robustness under disturbance scenarios. Current technologies have shortcomings in data integration, dynamic adaptability, and multi-objective trade-offs. Summary of the Invention

[0007] The purpose of this invention is to overcome the defects of the prior art and provide a dynamic aggregate visualization multi-level load collaborative management method.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] According to one aspect of the present invention, a method for dynamic aggregate visualization and multi-level load collaborative management is provided, characterized in that the method steps include:

[0010] S1. Acquire power system operation data, user energy consumption behavior data, and environmental parameters, and perform structured processing on these three types of data to obtain a structured test dataset;

[0011] S2. Based on the resource characteristics and spatial distribution features in the structured test dataset, perform hierarchical partitioning to generate tree or network topology structures and generate dynamic aggregates.

[0012] S3. Extract the time-domain and frequency-domain characteristic parameters of the load at each level in the dynamic aggregate, and construct a hierarchical load behavior model based on the characteristic parameters;

[0013] S4. Based on the dynamic aggregate topology and hierarchical load behavior model, and combined with constraints, construct a multi-objective optimization function, and generate a collaborative management strategy based on the multi-objective optimization function;

[0014] S5. Optimize collaborative management strategies by utilizing three mechanisms: real-time feedback, virtual verification, and long-term learning. Construct hierarchical visualization mapping rules and use these rules to output the execution results of dynamic aggregates, hierarchical load behavior models, and collaborative management strategies as graphical representations.

[0015] The power system operation data in S1 includes node voltage, frequency, line power, and equipment status; user energy consumption behavior data includes user-side load curves; environmental parameters include temperature, humidity, and light intensity; structured processing includes: converting the above three types of data into a unified tensor representation, detecting and cleaning outliers, unifying data format, modeling node spatial correlation, supplementing missing data, verifying data correlation, and checking data integrity.

[0016] As a preferred technical solution, the specific process of S2 includes: establishing quantitative indicators and spatial distribution characteristics of resource characteristics based on a structured test dataset; wherein, the quantitative indicators of resource characteristics include controllability indicators and response timeliness indicators; the spatial distribution characteristics include communication cost and electrical correlation; generating a tree-like hierarchy based on the quantitative indicators of resource characteristics and spatial distribution characteristics, selecting the node with the highest controllability index and optimal geographical centrality as the root node of the entire hierarchical structure, and assigning nodes other than the root node to subtrees through K-means clustering according to communication cost and electrical correlation; adding long-range connections across levels based on the tree structure through reconnection probability to form a network structure; after the hierarchical structure is established, when running tasks and transmitting information on it, the task allocation mechanism and information transmission protocol are followed;

[0017] The task allocation mechanism is as follows: calculate the load index of each node. When allocating tasks, first sort the nodes according to the load index, and prioritize nodes with a load index less than or equal to the preset load threshold. Then, update the node priority sorting according to the real-time response timeliness index.

[0018] The information transmission protocol specifically includes intra-layer communication rules, cross-layer communication rules, and information compression rules. Intra-layer communication rules stipulate that nodes at the same level use broadcast mode, with intra-layer communication bandwidth accounting for 70% of the total bandwidth resource. Cross-layer communication rules stipulate that nodes at different levels use unicast mode, with cross-layer communication bandwidth accounting for 30% of the total bandwidth resource, and the latency of cross-layer communication must be less than or equal to 100ms. ms .

[0019] As a preferred technical solution, the time-domain and frequency-domain characteristic parameters of the load at each level extracted by S3 include steady-state fluctuation patterns and sudden change response characteristics. The steady-state fluctuation patterns include steady-state fluctuation amplitude and periodic intensity; the sudden change response characteristics include sudden change amplitude and sudden change recovery time; the frequency-domain characteristic parameters include frequency band energy distribution and frequency domain energy entropy; and the time-domain and frequency-domain characteristic parameters also include correlation weight indicators, which include cross-level correlation and spatial correlation weight.

[0020] As a preferred technical solution, the hierarchical load behavior model in S3 has a hierarchical structure, which includes a first level and a second level. In the first level, a device-level load behavior model is established to describe the start-up and shutdown patterns of a single device. In the second level, a hierarchical aggregation model is constructed by superimposing the first level and introducing a synergistic effect term. In this hierarchical load behavior model, the time scales of minutes and below use an autoregressive moving average model to predict the load recovery trend after a sudden change, while the time scales of hours and above use an LSTM network to capture nonlinear time-series dependencies.

[0021] As a preferred technical solution, the multi-objective optimization function in S4 includes a response delay penalty term and a system stability penalty term; the specific formula of the multi-objective optimization function is:

[0022] ;

[0023] ;

[0024] ;

[0025] in, To optimize decision variables; and These are the weighting coefficients; In response to the delay penalty, the greater the delay, the higher the penalty value; Total number of devices; k For equipment serial number; The current time; For the first k Response time of each device; The exponential decay coefficient; This is a penalty item for system stability. and These are the weighting coefficients; For time t System frequency deviation at that time; For time t The voltage at that time is within acceptable limits; This is the system's rated voltage; This is the voltage deviation term.

[0026] As a preferred technical solution, the collaborative management strategy in S4 is a recursive strategy, specifically including a first decision level and a second decision level. The collaborative management strategy also includes a negotiation mechanism based on game theory. The input to the first decision level is global load prediction and system constraints, and a global strategy is generated based on the input. The objective of this global strategy is to minimize the system stability penalty term. This global strategy is then transformed into a set of inter-level collaborative instructions and output. The input to the second decision level is level weights and local constraints. After receiving the inter-level collaborative instruction set, the second decision level combines the input to decompose the task into sub-levels, thereby calculating a local strategy for each device or sub-region. Finally, the local strategy is transformed into device-level operation instructions and output. The negotiation mechanism based on game theory: when there is a conflict between inter-level instructions, a non-cooperative game model is constructed, and the optimal strategy is finally solved through Nash equilibrium.

[0027] As a preferred technical solution, the real-time feedback in S5 is as follows: For the current measurement, multi-dimensional feedback data is collected in real time, including command completion rate, response delay distribution, and load deviation. Performance degradation detection is performed based on the multi-dimensional feedback data. When the command completion rate is less than a preset completion rate threshold or the response delay distribution is greater than a preset delay distribution threshold, performance degradation is determined. An isolated algorithm is used to detect abnormal feedback data points, calculate anomaly scores, and perform a policy optimization process. This policy optimization process includes: updating policy parameters using online gradient descent, global policy adjustment based on a federated learning framework, and optimization of local policies based on reinforcement learning. Virtual verification involves: constructing a power system simulation model containing disturbance scenarios, injecting the current policy into the simulation model, and running a Monte Carlo simulation. Next, the failure probability of the strategy is statistically analyzed. If the failure probability of the strategy is higher than the preset failure probability threshold, the strategy optimization process is carried out. At this time, the strategy optimization process is as follows: a robustness penalty term is added to the optimization objective, the strategy is optimized based on adversarial training, and finally the adversarial examples are solved by projective gradient descent.

[0028] The long-term learning process involves: constructing a data lake to store real-time operational data in time series; calculating the volatility of hierarchical weights and the error between predicted and actual loads based on the real-time operational data; constructing an incremental self-learning model, using a federated learning framework to collaboratively train the self-learning model across hierarchical devices, and using this self-learning model to complete the policy optimization process; constructing a closed-loop calibration and optimization process, periodically comparing the deviation between predicted and actual loads, adjusting the hierarchical division threshold based on the deviation, and using reinforcement learning to adjust parameters based on real-time operational data, updating the policy through the PPO algorithm; fitting the decay curve of the calibrated policy performance indicators based on real-time operational data, performing cross-cycle data alignment, aligning the load characteristic distribution of different time periods, and verifying data consistency through the maximum mean difference; when the decay rate at a certain point on the decay curve exceeds a preset decay rate threshold or the data consistency is less than a preset consistency threshold, using the aforementioned incremental self-learning model to perform the policy optimization process.

[0029] As a preferred technical solution, the visualization mapping rules in S5 specifically include aggregate topology mapping rules, load state mapping rules, and policy execution result mapping rules. The aggregate topology mapping rules are used to output dynamic aggregates as graphical representations. Specifically, based on the hierarchical partitioning results of the dynamic aggregates, force-directed graphs or hierarchical layout algorithms are used to dynamically render nodes and edges. Node attributes include node size and node color; node size maps to hierarchical weights, and node color encodes the node's response timeliness score. Edge attributes include edge thickness and solid / dashed lines; edge thickness represents communication bandwidth utilization, and solid / dashed lines are used to distinguish inter-hierarchical collaboration rules. Node dragging and scaling are supported, as well as triggering local topology detail expansion. The load state mapping rules are used to output the hierarchical load behavior model as a graph. The graphical representation includes the layered overlay process of time-series load curves and the highlighting process of abrupt events. When layering time-series load curves, the load curves are aggregated and displayed alongside the equipment load curves according to their hierarchical structure, using transparency layering. When highlighting abrupt events, pulse waveform markers are used to mark the abrupt response events and associate them with corresponding equipment status labels. The strategy execution result mapping rules are used to output the execution results of the collaborative management strategy as a graphical representation, specifically using a multi-objective optimization index dashboard and a strategy execution heatmap. The multi-objective optimization index dashboard dynamically displays the latency penalty and stability indicators of the current strategy, using a circular progress bar or heatmap, comparing the preset reference threshold with the actual values ​​after the current strategy execution. The strategy execution heatmap overlays the strategy execution intensity matrix onto the aggregate topology map.

[0030] The specific formula for the strategy execution strength matrix is:

[0031]

[0032] in, Let be the policy execution intensity value of the i-th and j-th topological positions in the aggregate topology at time t; N is the total number of devices participating in policy execution in the aggregate. Let k be the weight of device k; Let be the real-time active power of device k at time t; This refers to the equipment's rated maximum active power.

[0033] According to another aspect of the present invention, a dynamic aggregate visualization multi-level load collaborative management system is provided, the system including a data processing and integration module, a dynamic aggregate generation module, a hierarchical load modeling module, a collaborative strategy optimization module, and a strategy evaluation and visualization module;

[0034] The data processing and integration module is used to acquire power system operation data, user energy consumption behavior data and environmental parameters, and to perform structured processing on these three types of data to obtain a structured test dataset.

[0035] The dynamic aggregate generation module is used to generate tree or network topologies by hierarchically dividing the structured test dataset based on the resource characteristics and spatial distribution features in the dataset, thus generating dynamic aggregates.

[0036] The hierarchical load modeling module is used to extract the time-domain and frequency-domain characteristic parameters of each level of load in the dynamic aggregate, and to construct a hierarchical load behavior model based on the characteristic parameters;

[0037] The collaborative strategy optimization module is used to construct a multi-objective optimization function based on the dynamic aggregate topology and hierarchical load behavior model, combined with constraints, and to generate a collaborative management strategy based on the multi-objective optimization function.

[0038] The strategy evaluation and visualization module is used to optimize collaborative management strategies using three mechanisms: real-time feedback, virtual verification, and long-term learning. It also constructs hierarchical visualization mapping rules, which are used to output the execution results of dynamic aggregates, hierarchical load behavior models, and collaborative management strategies as graphical representations.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. This invention acquires power system operation data, user energy consumption behavior data, and environmental parameters, and performs structured processing on these three types of data to construct a hierarchical load behavior model and a multi-objective optimization function. Based on the multi-objective optimization function, a collaborative management strategy is generated, and the collaborative management strategy is optimized using three mechanisms: real-time feedback, virtual verification, and long-term learning. This provides an end-to-end, visualized, multi-level load collaborative management solution, achieving full automation from data processing, structure generation, model construction, strategy optimization to result display, thus improving the system's intelligence level and management efficiency. This management method features good data integration performance, strong dynamic adaptability, and achieves multi-objective trade-offs, completing intelligent load collaborative management and significantly improving the adaptability of the load collaborative management strategy.

[0041] 2. In this invention, by establishing quantitative indicators of resource characteristics and spatial distribution features, and then generating a tree-like hierarchy based on these, a network structure is formed by adding long-range cross-level connections based on reconnection probability. This combines the clear hierarchy of the tree structure with the high robustness of the network structure, enabling this method to achieve the dual advantages of efficient coordination and stable disturbance resistance in multi-level load management, adapting to the complex and ever-changing operating scenarios of power systems. Furthermore, after the hierarchical structure is established, tasks and information transmission follow a task allocation mechanism and an information transmission protocol. The task allocation mechanism prioritizes low-load nodes and dynamically adjusts priorities based on real-time response timeliness, achieving global load balancing and avoiding situations where some devices are overloaded while others are idle, thus maximizing resource utilization. The information transmission protocol ensures the orderliness and efficiency of information transmission. Both significantly improve operational efficiency and coordination capabilities.

[0042] 3. This invention considers both time-domain and frequency-domain characteristics to more comprehensively depict the static patterns of load. Furthermore, the introduction of correlation weighting indicators breaks down the isolation of loads at different levels, enabling the quantification and capture of the mutual influences between loads at different levels and spatial locations. The rich feature dimensions make the model less sensitive to single abnormal data, improving robustness. This invention also constructs a hierarchical load behavior model with first and second levels. The first level starts from the most basic unit, accurately describing the start-up, shutdown, and operation patterns of individual devices, providing high-quality data for upper-level aggregation. The second level superimposes on the first level and introduces a synergistic effect term to model the mutual influences between devices, making the aggregated load prediction more accurate. Moreover, by distinguishing between minute-level and hour-level multi-scale predictions, the advantages of each time scale are utilized, achieving an optimal balance between prediction performance and computational efficiency.

[0043] 4. In this invention, by constructing a multi-objective optimization function that includes response delay penalty terms and system stability penalty terms, a balanced optimization of multiple objectives is achieved while satisfying system constraints. By adopting a recursive collaborative management strategy that includes a first decision level and a second decision level, and by constructing a negotiation mechanism based on game theory, the global optimality, local feasibility, and effective coordination in conflict scenarios of the strategy are ensured.

[0044] 5. In this invention, during real-time feedback, performance degradation detection is achieved by collecting multi-dimensional feedback data in real time, enabling dynamic monitoring and rapid adjustment of the strategy execution effect. This ensures the system can adapt to changes promptly and maintain optimal performance. During virtual verification, digital twin technology is used to fully test the strategy in a virtual environment, identifying potential risks in advance and significantly enhancing the robustness and reliability of the strategy in real-world complex scenarios. During long-term learning, by constructing a data lake and a self-learning model, the strategy possesses the ability to self-evolve and continuously optimize, adapting to environmental changes and performance degradation during long-term operation and extending the effective lifespan of the strategy.

[0045] 6. In this invention, visualization is achieved by constructing a visualization mapping rule that includes aggregate topology mapping rule, load status mapping rule and strategy execution result mapping rule. This presents complex topology structures, load status and strategy execution results in an intuitive graphical way, which greatly improves the readability of information and the transparency of decision-making, making it easier for managers to understand, analyze and intervene. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the method steps of a dynamic aggregate visualization multi-level load collaborative management method in this invention;

[0047] Figure 2 This is a flowchart of the data structuring process in this invention;

[0048] Figure 3 This is a flowchart of the process of generating dynamic aggregates through hierarchical division in this invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0050] To address the core pain points in existing power system load management, such as low collaborative efficiency due to data heterogeneity, rigid hierarchical division making it difficult to adapt to dynamic resource characteristics, prediction bias caused by insufficient accuracy of load feature analysis, lack of global optimization capability in multi-objective strategy generation, inability of static strategies to cope with real-time disturbances, and insufficient system adaptability, this solution proposes a dynamic aggregate visualization multi-level load collaborative management method. It achieves cross-dimensional information association through a multi-channel data fusion framework, constructs a dynamic aggregate hierarchical structure to optimize resource collaboration efficiency, employs hierarchical feature analysis technology to improve the quantitative accuracy of load evolution patterns, combines a multi-objective optimization collaborative management strategy generation mechanism, and designs a dynamic optimization and closed-loop calibration process based on real-time feedback. Ultimately, this achieves intelligent and adaptive collaborative management of power system loads, overcoming the technical limitations of traditional methods in terms of dynamism, robustness, and multi-objective trade-offs, and solving the efficiency and reliability problems of load collaborative management in complex scenarios.

[0051] Dynamic aggregates refer to hierarchical structural units dynamically divided based on resource characteristics (controllability, response timeliness) and spatial distribution features. They achieve multi-level collaboration through tree-like or mesh topologies, supporting efficient aggregation of load resources and cross-level command transmission. Multi-level load characteristic analysis: By extracting time-domain fluctuation patterns, frequency-domain energy entropy, and correlation weight indicators, a hierarchical load behavior model is constructed to quantify load evolution characteristics at different time scales, providing accurate input for strategy generation. Collaborative management strategy generation: Combining load forecasting and system constraints, a multi-objective optimization function (energy efficiency balance, delay suppression, stability assurance) is constructed. A hierarchical recursive decision-making process enables cross-level command coordination and priority allocation. Dynamic strategy optimization and execution verification: Based on real-time feedback data, parameter correction and local iteration are triggered. Virtual simulation is used to verify the robustness of the strategy under disturbance scenarios, forming a closed-loop optimization mechanism of "perception-decision-verification." System adaptive calibration mechanism: A self-learning model is trained using long-term operating data to dynamically adjust the hierarchical division threshold and collaborative rule parameters. Continuous accuracy optimization of feature analysis and strategy generation is achieved through comparison of predicted and actual load deviations.

[0052] Example 1

[0053] In this embodiment, a dynamic aggregate visualization multi-level load collaborative management method is adopted, and the method steps are as follows: Figure 1 As shown, it specifically includes:

[0054] S1. Acquire power system operation data, user energy consumption behavior data, and environmental parameters, and perform structured processing on these three types of data to obtain a structured test dataset;

[0055] S2. Based on the resource characteristics and spatial distribution features in the structured test dataset, perform hierarchical partitioning to generate tree or network topology structures and generate dynamic aggregates.

[0056] S3. Extract the time-domain and frequency-domain characteristic parameters of the load at each level in the dynamic aggregate, and construct a hierarchical load behavior model based on the characteristic parameters;

[0057] S4. Based on the dynamic aggregate topology and hierarchical load behavior model, and combined with constraints, construct a multi-objective optimization function, and generate a collaborative management strategy based on the multi-objective optimization function;

[0058] S5. Optimize collaborative management strategies by utilizing three mechanisms: real-time feedback, virtual verification, and long-term learning. Construct hierarchical visualization mapping rules and use these rules to output the execution results of dynamic aggregates, hierarchical load behavior models, and collaborative management strategies as graphical representations.

[0059] The main steps of implementing this method are as follows:

[0060] 1. Data Acquisition and Structured Processing: Basic data streams of power system operation, user energy consumption behavior, and environmental parameters are captured through multi-channel input interfaces to construct a unified data representation format. Data correlation mapping rules are established to achieve logical association and dynamic supplementation of cross-dimensional information.

[0061] (1) Data source definition and access: Real-time capture of three types of data streams through multi-channel input interfaces (such as sensor networks, SCADA systems, smart meters, and user behavior log interfaces):

[0062] Power system operating data: including node voltage (V) t ), frequency (f) t ), line power (P) ij Q ij Equipment status (such as transformer load rate) User energy consumption behavior data: covering user-side load curves. (s) u,k For device start / stop status, P u,k (Equipment power), time-of-use electricity price response coefficient Environmental parameters: Temperature (T) t ),humidity( ), light intensity (I) t Data is acquired through meteorological sensors or public APIs. Data synchronization mechanism: A timestamp alignment algorithm (such as linear interpolation) is used to ensure consistent temporal resolution across multiple data sources (second-level synchronization error Δt ≤ 1s).

[0063] The specific processing steps for data standardization and structuring are as follows: Figure 2As shown, the process includes converting heterogeneous data into a unified tensor representation, outlier detection and cleaning, data format standardization, data association, hierarchical storage, and data integrity verification. By standardizing and structuring multi-source heterogeneous data, data quality and consistency are ensured, providing a solid data foundation for subsequent accurate analysis and decision-making.

[0064] (2) Data representation construction, transforming heterogeneous data into a unified tensor representation:

[0065]

[0066] Where m is the data dimension (such as power parameters, environmental parameters, etc.) and n is the time series length.

[0067] (3) Outlier detection and cleaning, Grubbs test: identifying outliers and hypothesis testing statistic. When G>G threshold (e.g. G) 0.95 When the value is 0, it is considered an outlier.

[0068] Sliding window mean fill: for bad data (Window width w=5).

[0069] (4) Unified data format, unstructured data processing: Parsing user behavior logs into structured fields, for example:

[0070] User behavior tags: b u =[Appliance type, usage period, energy efficiency rating].

[0071] Dynamic Feature Extraction: Calculating the Short-Time Fourier Transform (STFT) Frequency Domain Energy Entropy of the Load:

[0072]

[0073] Where X k H(f) is the STFT coefficient, which quantifies the complexity of load fluctuations (the larger H(f) is, the more disordered the fluctuations).

[0074] (5) Spatiotemporal correlation rules and spatial correlation modeling: based on power flow calculation sensitivity Define the electrical distance matrix between nodes. ( (This is the attenuation coefficient).

[0075] Dynamic supplementation mechanism: When the data source is missing, imputation is performed using Markov Chain Monte Carlo (MCMC).

[0076]

[0077] Among them, X missFor missing data, X obs The data is known.

[0078] (6) Data correlation verification and consistency check: Evaluate the consistency of multi-source data through cross-entropy.

[0079]

[0080] When CE > 1.5, an abnormal alarm is triggered and redundant data source switching is initiated.

[0081] (7) Layered storage architecture, raw data layer: stores raw time-series data (compression rate 80%, storage period 1 year).

[0082] Feature data layer: Extracting load features (such as peak-to-valley difference) ), storage period of 3 years.

[0083] API Interface Design: Provides a standardized data retrieval protocol.

[0084] def fetch_data(start_time, end_time, data_type="power"):

[0085] return X_t[start_time:end_time][data_type]

[0086] It supports flexible invocation based on time range and data type, with an interface response time of <100ms.

[0087] (8) Data integrity verification and temporal continuity test: Calculate the mean and variance of the differences between adjacent timestamp data using a sliding window.

[0088]

[0089] when The data was determined to be broken at that time.

[0090] Anomaly feedback mechanism: This mechanism will process data integrity verification results (such as CE values, etc.) Input to the system's adaptive calibration module to dynamically adjust data cleaning parameters (such as the sliding window width W).

[0091] 2. Dynamic Aggregate Hierarchical Division: Based on resource characteristics (such as controllability and response timeliness) and spatial distribution features, define the hierarchical division principles of the aggregate to form a tree-like or network-like topology. Design inter-level collaboration rules to clarify the interaction logic of different levels in task allocation, information transmission, and decision-making authority.

[0092] (1) Quantitative indicators of resource characteristics, controllability indicators: Based on historical data of the controllable state of equipment, the response probability p is calculated. ctrl :

[0093] Number of times to respond to commands

[0094] Controllability threshold set to Devices with values ​​below this threshold are classified as low response level devices.

[0095] Response timeliness metrics: Data on instruction execution delay Calculate the timeliness score:

[0096] The attenuation coefficient ( ≥0.1)

[0097] Timeliness score It is positively correlated with the weight of hierarchical division.

[0098] (2) Spatial distribution feature modeling, geographic distance weighting: based on node coordinates Calculate Euclidean distance Converted into communication costs (ε is a constant that is removed from zero).

[0099] Electrical coupling strength: using power flow sensitivity matrix Define spatial correlation Highly correlated nodes are prioritized to be assigned to the same level.

[0100] (3) Tree-like hierarchical generation, divided from top to bottom:

[0101] Root node selection: Select the node with the highest controllability. And the geographic centrality is optimal (e.g., betweenness centrality). The node is taken as the root node.

[0102] Child node allocation: based on communication cost and electrical correlation The remaining nodes are assigned to subtrees using K-means clustering, with the objective function being: .

[0103] Where w is the weighting coefficient, The communication cost from the node to the root. This represents the degree of association between a node and its parent node.

[0104] (4) Mesh hierarchy optimization, small-world network modeling: using reconnection probability p rewire Adjust the network topology to balance local connectivity with global propagation efficiency.

[0105] Reconnection probability (p)rewire ) is the core parameter for adjusting network topology (value is 0~1), p rewire The value is determined in the following way:

[0106] Method 1: Experience Setting

[0107] Power network optimization: p=0.1~0.3 (balancing local power supply stability and fault propagation control)

[0108] Method 2: Dynamic Optimization

[0109] Reinforcement learning: p is used as a policy parameter and dynamically adjusted through a reward function (e.g., network efficiency, robustness). Adaptive algorithms: p is adjusted in real-time based on network load or failure events.

[0110] Method 3: Data-driven

[0111] Historical data analysis: Fit the optimal value of p by historical network performance (such as propagation delay, failure rate).

[0112] Machine learning models: use regression or neural networks to predict p, with input variables including node degree distribution, network structure, etc.

[0113] Local connection weights: Preserving edges between highly associated nodes (A) ij >θ A ).

[0114] Random long-range connections: with probability p rewire Add cross-region edges to enhance system robustness.

[0115] (5) Task allocation mechanism, load balancing based on resource capacity: Define node load index ( For the current workload, (For maximum capacity), task allocation prioritizes node load index L. i Nodes with a value ≤0.8.

[0116] Dynamic priority adjustment: based on real-time response latency τ real (t) Update node priority:

[0117]

[0118] in, This represents the system's average latency.

[0119] (6) Information transmission protocol, intra-layer communication: Broadcast mode is used between nodes at the same level, and the bandwidth allocation ratio is... Cross-layer communication: Unicast mode is used between upper and lower level nodes, bandwidth... And it must meet the time delay constraint. .

[0120] Information compression rules: Differential encoding is used for non-critical data (such as environmental parameters), with a compression ratio of [missing information]. .

[0121] (7) Adaptive repartitioning trigger condition, performance degradation detection: when the average response delay within the layer... Or node failure rate Refactoring is triggered at certain times.

[0122] The average response latency within a hierarchy is the average response time (such as the sum of network latency and computation latency) of all nodes within a certain hierarchy when processing requests.

[0123] This is a preset threshold.

[0124] According to the formula's judgment logic, when the average response latency within a level exceeds a preset threshold, a refactoring is triggered.

[0125] The failure rate is the proportion of failed nodes in the current level or system to the total number of nodes. The number of failed nodes N is counted by periodically checking node status (e.g., through a heartbeat mechanism). fail Total number of nodes N total

[0126] 。 ;

[0127] The average response latency within a hierarchy is the average response time (such as the sum of network latency and computation latency) of all nodes within a certain hierarchy when processing requests. This is a preset threshold. According to the formula's judgment logic, when the average response latency within a level exceeds the preset threshold, a refactoring is triggered.

[0128] Incremental adjustment algorithm: performs local optimization only on the affected subtrees, reducing computational overhead.

[0129]

[0130] Where S is the set of affected nodes, and the objective is to minimize .

[0131] This represents the change in energy or error, that is, the change in the overall objective function (such as the energy function or loss function) after the algorithm performs local optimization on the affected subtree. Incremental algorithms reduce computational overhead by calculating this local change, thus avoiding recalculating the objective function of the entire system. This represents the weight coefficient of node i, used to measure the importance or influence of that node in local optimization. The weight can be set based on the saliency of the node's characteristics, historical optimization experience, or domain knowledge. A larger weight means that changes to that node have a more significant impact on the overall objective function. This represents the local energy value or local error value of node i after local optimization. It reflects the optimal state of the adjusted node and its associated subtrees (such as minimum energy configuration or minimum prediction error). This represents the local energy value or local error value of node i before optimization, i.e., the initial state before adjustment. The subscript i belongs to the set S, where S is defined as the set of affected subtree nodes. The incremental algorithm only calculates the local region (not the entire system) directly related to the current adjustment operation. The range of S is determined by the algorithm's locality constraints (such as dependencies and neighborhood rules) to ensure computational efficiency.

[0132] (8) Topological evolution model, policy update based on reinforcement learning: defining the state space The action space A = {merging levels, splitting levels}, and the reward function is:

[0133]

[0134] in To delay the improvement amount, Due to changes in communication costs, For robustness gain.

[0135] 3. Multi-level load characteristic analysis: Extract time-domain and frequency-domain characteristic parameters of load at each level, including steady-state fluctuation patterns, sudden change response characteristics, and correlation weighting indicators. Construct a hierarchical load behavior model based on the characteristic parameters to quantify the load evolution patterns at different time scales.

[0136] 3.1 Multi-level load feature extraction:

[0137] (1) Time-domain characteristic parameters, steady-state fluctuation law: Define the load curve L u The steady-state fluctuation amplitude σ of (t) steady and periodic intensity C period :

[0138]

[0139]

[0140] in, Reflects the long-term stability of the load. Quantify periodic characteristics (such as daily cycle, weekly cycle).

[0141] Mutation response characteristics: Detection of load mutation points And calculate the mutation magnitude and recovery time :

[0142]

[0143] in, The mutation initiation time. This is the moment when the load recovers to within the threshold.

[0144] (2) Frequency domain characteristic parameters, frequency band energy distribution: perform wavelet transform on the load curve (Daubechies wavelet basis) ), calculate the energy percentage of each frequency band:

[0145]

[0146] in, Let be the wavelet coefficients of the j-th layer. This represents the energy contribution rate of frequency band j.

[0147] Frequency domain energy entropy: quantifying the complexity of load fluctuations.

[0148]

[0149] The larger the load, the more disordered the load fluctuations (such as sudden start-up and shutdown of electrical appliances).

[0150] (3) Correlation weight index, cross-level correlation: calculate the load correlation between level i and parent level p(i). :

[0151]

[0152] in, For covariance, The standard deviation is denoted as .

[0153] Spatial association weights: combined with the electrical distance matrix Time-series correlation Define the overall association weight:

[0154]

[0155] 3.2 Modeling of Stratified Load Behavior:

[0156] (1) Hierarchical model structure, lower level (equipment level): establish equipment-level load behavior model Describe the start-up and shutdown patterns of a single device:

[0157]

[0158] in, Let m be the amplitude of the m-th activity mode of device k. The attenuation coefficient is... It is noise.

[0159] Mid-to-high level (regional / system level): By overlaying lower-level models and introducing synergistic effect terms, a hierarchical aggregation model is constructed.

[0160]

[0161] in, Hierarchical weights (as shown in the formula) ), This indicates cross-level synergy (such as demand response strategies).

[0162] (2) Time-scale quantification, short time scale (minute level): Autoregressive moving average (ARMA) model is used to predict the load recovery trend after the mutation:

[0163]

[0164] Where p is the autoregressive order and q is the moving average order.

[0165] Long time scale (hours and above): Capturing nonlinear temporal dependencies based on LSTM networks:

[0166]

[0167]

[0168] Among them, h t It is in a hidden state, c t It is in the cellular state.

[0169] 3.3 Dynamic calibration of correlation weights:

[0170] (1) Weight update mechanism based on Granger causality test: test the predictive ability of level i on j, and dynamically adjust the weight W. ij :

[0171]

[0172] Where RSS is the sum of squared residuals, and p and q are the number of model parameters. If F Granger >F threshold Then enhance W ij .

[0173] Weight optimization based on reinforcement learning: defining the state space Action space The reward function is:

[0174]

[0175] Weight update strategy using Q-learning:

[0176]

[0177] 3.4 Cross-layer transfer of features and strategies:

[0178] (1) Feature upward propagation rule, low level → mid-to-high level: device-level features By weight After aggregation, an additional delay compensation term is added. Passed to the upper level:

[0179]

[0180] in, Due to communication delay, For compensation functions (such as linear interpolation).

[0181] (2) Logic for downward decomposition of strategy, from high-level to low-level: optimize the strategy Decomposed into device-level instructions To satisfy resource constraints:

[0182]

[0183] in, This is the execution cost penalty for device k.

[0184] 4. Collaborative Management Strategy Generation: Combining load forecasting results with system constraints, a multi-objective optimization function is constructed, covering energy efficiency balance, response delay suppression, and system stability assurance. A hierarchical decision-making process is designed, using recursive strategy decomposition to achieve coordinated generation and priority allocation of cross-level instructions.

[0185] 4.1 Construction of Multi-Objective Optimization Function:

[0186] (1) Objective function design based on load forecast results Given system constraints, construct a multi-objective optimization function that covers the following core metrics:

[0187] Response latency suppression: Limiting command response time Introduce a delay penalty term:

[0188]

[0189] in, This represents the actual response time of device k.

[0190] System stability assurance: through frequency deviation and voltage qualification rate Constraint system stability:

[0191]

[0192] in, These are the weighting coefficients.

[0193] Overall optimization goals:

[0194]

[0195] in, For multi-objective weights, satisfying .

[0196] (2) Constraints, Equipment Capacity Constraints:

[0197]

[0198] Communication bandwidth constraints:

[0199] Cross-level instruction transfer rate

[0200] Power balance constraints:

[0201]

[0202] 4.2 Hierarchical Decision-Making Process Design:

[0203] (1) Recursive strategy decomposition, high-level (system-level) decision-making: input global load forecast Given system constraints, generate an initial policy. The goal is to achieve equilibrium stability under constraints:

[0204]

[0205] The output is a set of inter-level collaborative instructions. .

[0206] Mid-to-low level (regional / equipment level) decision-making: based on hierarchical weights And local constraints, decompose global instructions into local policies :

[0207]

[0208]

[0209] The output is a device-level operation command. .

[0210] (2) Priority allocation mechanism, based on response timeliness: defining device priority. i Priority score ( (Based on historical average response time), high-scoring devices are prioritized for execution.

[0211] Dynamic adjustment rule: When a sudden increase in the response latency of device i is detected ( ), trigger priority reordering:

[0212]

[0213] 4.3 Cross-level command coordination:

[0214] (1) Instruction mapping and compensation, inter-level instruction mapping: mapping high-level instructions By weight Decompose to sub-levels:

[0215]

[0216] Communication delay compensation: Introducing a delay compensation term for cross-level instructions. The corrected instructions are:

[0217]

[0218] in, (t) represents the instruction transmission delay, and α is the compensation coefficient.

[0219] (2) Conflict resolution strategy, based on the negotiation mechanism of game theory: When there is a conflict between instructions between levels (such as the parent node requesting load reduction while the child node needs to maintain power supply), a non-cooperative game model is constructed:

[0220] , Let i be the utility function for level i. represents the conflict penalty coefficient. The optimal strategy is determined using Nash equilibrium.

[0221] 4.4 Strategy Dynamic Adjustment Interface:

[0222] (1) Real-time feedback access and deviation detection: calculate the actual load Compared with the predicted value The residual:

[0223]

[0224] when The policy is adjusted when the time is right.

[0225] (2) Rolling time-domain optimization: Model predictive control (MPC) framework is adopted, within the time window. Inner recursive optimization strategy:

[0226]

[0227] in, This is the optimal strategy for the next moment.

[0228] 5. Dynamic Strategy Optimization and Execution Verification: Based on real-time feedback data, the strategy execution effect is dynamically evaluated, triggering parameter adjustments and local strategy iterations. A virtual verification mechanism is established to verify the effectiveness of the strategy under perturbation scenarios through simulation and to optimize the strategy's robustness.

[0229] 5.1 Real-time feedback data evaluation and strategy performance quantification:

[0230] (1) Multi-dimensional feedback data collection and performance indicators:

[0231] Instruction completion rate: Calculates the device-level instruction execution success rate η exec :

[0232]

[0233] in, For indicator functions, This represents the allowable error range.

[0234] Response latency distribution: statistical command response time percentiles (e.g.) ):

[0235]

[0236] Load deviation penalty: Calculate the actual load Compared with the predicted value Root mean square error (RMSE):

[0237]

[0238] (2) Performance degradation detection, dynamic threshold triggering mechanism: when or When this happens, strategy optimization is triggered. For example, setting... .

[0239] Anomaly pattern recognition: The Isolation Forest algorithm is used to detect anomaly feedback data points and calculate anomaly scores. :

[0240]

[0241] Where, μ and Let be the feature mean and standard deviation, and n be the feature dimension. It is marked as an exception.

[0242] 5.2 Dynamic Parameter Adjustment and Local Strategy Iteration:

[0243] (1) Online learning and parameter update, incremental model update: Based on real-time feedback data, online gradient descent (OGD) is used to update parameters. :

[0244]

[0245] in, For loss function, For learning rate, This is real-time data.

[0246] Federated Collaborative Optimization: In cross-level policy adjustments, a federated learning framework is used to protect privacy by sharing only model gradients.

[0247]

[0248] in, This represents the intensity of differential privacy noise.

[0249] (2) Local policy re-optimization, based on fast adjustment of reinforcement learning: defining the state space Action space The reward function is:

[0250]

[0251] Update the policy network using the Proximal Policy Optimization (PPO) algorithm:

[0252]

[0253] in, ε represents the probability ratio between the old and new strategies, and ε is the pruning coefficient.

[0254] 5.3 Virtual Verification Mechanism and Robustness Enhancement:

[0255] (1) Construction of digital twin simulation environment and dynamic modeling of system: Establish a power system simulation model, including disturbance scenarios such as load changes and equipment failures:

[0256]

[0257] in, For random disturbance terms, These are system parameters.

[0258] Strategy Injection and Deduction: Introducing the strategy to be verified Inject the simulation model and run the Monte Carlo simulation. Next, the probability of statistical strategy failure:

[0259]

[0260] (2) Robust optimization method, robust constraint embedding: add a robust penalty term to the optimization objective, such as minimizing the cost in the worst-case scenario:

[0261]

[0262] in, For the perturbation set, For robustness weights.

[0263] Policy enhancement based on adversarial training: generating adversarial perturbations To simulate extreme scenarios and optimize the strategy's resistance to interference:

[0264] st

[0265] Adversarial examples are solved using projective gradient descent (PGD).

[0266] 5.4 Closed-loop feedback and strategy version management:

[0267] (1) Strategy version iteration record, version comparison and rollback mechanism: record strategy version Performance metrics (such as) When the performance of the new version drops below a threshold Automatic rollback occurs.

[0268] Incremental policy patching: Updates are only applied locally to submodules experiencing performance degradation, reducing global policy fluctuations. For example, if a region-level policy fails, only the weights of that region are adjusted. .

[0269] (2) Long-term performance evaluation and decline trend analysis: fitting the decline curve of strategy performance over time (such as the exponential decay model):

[0270]

[0271] in, For the rate of decay, This is the noise term.

[0272] Adaptive calibration trigger: when decay rate At that time, the seventh part of the system's adaptive calibration mechanism is triggered, and the hierarchical division rules and collaborative parameters are updated globally.

[0273] 6. Visual Interaction Logic Design: Construct hierarchical visualization mapping rules to transform aggregate topology, load status, and strategy execution results into graphical representations. Design an interactive analysis framework that allows users to drive strategy optimization direction through parameter adjustments and hypothetical scenario inputs.

[0274] 6.1 Layered Visualization Mapping Rules:

[0275] (1) Aggregate topology visualization and tree / mesh structure rendering: Based on the dynamic aggregation hierarchy partitioning results (such as tree or mesh topology), force-directed graphs or hierarchical layout algorithms (such as the Reingold-Tilford algorithm) are used to dynamically render nodes and edges:

[0276] Node attributes: Node size maps to hierarchical weight W ij Color coding response timeliness score time (For example, cool colors indicate high latency).

[0277] Edge attribute: Edge thickness represents the communication bandwidth utilization rate. Solid and dashed lines distinguish between collaborative rules at different levels (e.g., solid lines represent mandatory constraints, while dashed lines represent suggested strategies).

[0278] Dynamic interaction: Supports node dragging and zooming, triggering the expansion of local topology details (such as clicking the parent node to display the load distribution of child nodes).

[0279] (2) Load status visualization, with time-series load curves layered and overlaid:

[0280] Hierarchical Aggregate Load Curve Equipment-level load curve Overlay display, using alpha blending to avoid visual confusion:

[0281]

[0282] Where k is the adjustment factor. This represents the average load of the system.

[0283] Mutation event highlighting: Highlighting mutation response events (such as...) Pulse waveform marking is used, and it is associated with device status labels (such as fault codes). ).

[0284] 3) Visualization of strategy execution results, multi-objective optimization indicator dashboard: dynamically displays the latency penalty of the current strategy. and stability indicators A circular progress bar or heatmap is used to compare the threshold with the actual value.

[0285] Policy execution heatmap: A policy execution intensity matrix overlaid on the aggregate topology graph. The calculation formula is:

[0286]

[0287] in, Let k be the weight of device k. This represents real-time power.

[0288] 6.2 Interactive Analysis Framework Design:

[0289] (1) Parameter adjustment interface, slider and input box linkage: Users can adjust the multi-objective optimization weights through the slider. Real-time triggering of strategy regeneration (as shown in formula) The Pareto frontier changes are illustrated using 3D surface plots.

[0290] Assuming the scenario input module supports user-defined disturbance scenario parameters (such as extreme temperatures). Electricity price fluctuations Load forecast bias is generated through Monte Carlo simulation. And automatically triggers a policy robustness assessment.

[0291] (2) Dynamic deduction and feedback closed loop, real-time strategy simulator: After the user adjusts the parameters, the strategy generation module in Part 5 is called to run incremental MPC optimization (time window) in the virtual simulation environment. ), returns the predicted load curve and system stability indices.

[0292] Interactive annotation and commentary: Allows users to mark anomalous events in the visualization interface (such as clicking on load mutation points to add comments), and feeds the annotated data back to the calibration mechanism in Part VII to optimize the feature parsing threshold.

[0293] 6.3 Data-driven visualization logic:

[0294] (1) Multi-source data mapping rules, time-series data to color mapping: mapping environmental parameters (such as temperature) After normalization, it is mapped to the background color temperature (cool colors correspond to low temperatures, warm colors correspond to high temperatures), and the formula is:

[0295]

[0296] Hierarchical weights to layout transformation: based on hierarchical weights Dynamically adjust node positions; high-weight nodes are automatically clustered to the topology center (e.g., the force-directed parameter in the Fruchterman-Reingold algorithm). ).

[0297] (2) User behavior feedback modeling and interaction hotspot analysis: Visualize the areas where users frequently operate (e.g., weight parameters that have been adjusted multiple times). ), generate the interaction frequency matrix Guide to interface optimization:

[0298]

[0299] in, For indicator functions, Let i be the operating region of user i at time t.

[0300] 6.4 Visualization and System Closed-Loop Linkage:

[0301] (1) Visualized results drive strategy optimization and user preference learning: Record users' preferences for adjusting specific strategies (such as reducing multiple times). User profile vectors are generated using a collaborative filtering algorithm. Enter the sixth part to automatically generate strategy suggestions.

[0302] Anomaly alarm linkage: When the visual interface detects a load deviation An automatic pop-up strategy correction suggestion panel will recommend adjusting the parameter range (e.g., ...). ).

[0303] (2) Long-term performance feedback to the calibration mechanism, and archiving of visualized performance indicators: This includes policy performance indicators after user interaction (such as adjusted performance indicators). The data is stored in the seventh part of the calibration database, triggering the update of hierarchical division rules and collaborative parameters.

[0304] Dynamic interface version management: Based on long-term operation data, automatically switch the visualization mode (such as switching from a detailed topology view to a simplified dashboard view) to adapt to different user roles (such as operations and maintenance personnel vs. management).

[0305] 7. System Adaptive Calibration Mechanism: A self-learning model is built based on long-term operational data to dynamically adjust the hierarchical division threshold and collaborative rule parameters. A closed-loop calibration process is designed to continuously optimize the accuracy of feature parsing and strategy generation by comparing the deviation between predictions and actual loads.

[0306] 7.1 Self-learning model construction and dynamic parameter updating:

[0307] (1) Long-term operational data storage and feature extraction, data lake architecture: Long-term operational data is stored in time series, including load curve L(t) and hierarchical partitioning results. Strategy execution indicators Columnar storage (such as Parquet format) is used to support efficient querying.

[0308] Feature engineering: Extracting dynamic features from raw data, for example:

[0309] Hierarchical stability index: Calculating hierarchical weights volatility :

[0310]

[0311] Strategy Failure Modes: Statistics on Strategy Failure Events The spatiotemporal distribution characteristics, such as the spatial clustering radius of high-frequency failure regions. .

[0312] (2) Incremental self-learning model, federated learning framework: Self-learning models are collaboratively trained across different device levels while protecting privacy. For example, device-level models. Aggregate global model using federated average (FedAvg) :

[0313]

[0314] in, For local datasets, This is the gradient of the loss function.

[0315] Concept drift detection: The ADWIN algorithm is used to dynamically detect changes in data distribution, and when the drift amplitude... Time-triggered model retraining:

[0316]

[0317] Where p(t) is the predicted probability distribution at the current time.

[0318] 7.2 Closed-loop calibration process design:

[0319] (1) Prediction-actual deviation analysis and multi-dimensional error calculation: Comparison of predicted load Compared with actual load Calculate the timing error index:

[0320] Absolute percentage error (MAPE):

[0321]

[0322] Dynamic Time Warping (DTW) Distance: Measures the difference in load curve shape.

[0323]

[0324] (2) Dynamic adjustment of hierarchical division rules and adaptive updating of thresholds: The hierarchical division thresholds are dynamically adjusted based on the deviation analysis results. For example, when At this time, reduce the granularity of hierarchical division:

[0325]

[0326] in, This is the attenuation coefficient.

[0327] Topology reconfiguration trigger: When a high-frequency failure region (such as...) is detected... This triggers local topology optimization, such as merging adjacent inefficient layers.

[0328] (3) Collaborative rule parameter optimization, parameter adjustment based on reinforcement learning: defining the state space Action space (Optimize target weight adjustment), the reward function is:

[0329]

[0330] The policy network is updated using the PPO algorithm, and the optimal parameter combination is output. .

[0331] 7.3 Dynamic Parameter Optimization and Strategy Iteration:

[0332] (1) Online Bayesian optimization, objective function modeling: the accuracy of policy generation Modeled as a Gaussian process (GP):

[0333]

[0334] Select the next optimization point by collecting surrogate models (such as ExpectedImprovement). .

[0335] Parameter space reduction: Using Markov chain Monte Carlo (MCMC) sampling, the weight parameters are reduced. Search scope:

[0336]

[0337] Where D represents historical calibration data.

[0338] (2) Layered strategy retraining and cross-level joint optimization: In a tree-like hierarchical structure, subtree models are retrained from bottom to top. For example, for leaf nodes that fail frequently, transfer learning is used to reuse the features of the parent node:

[0339]

[0340] in, This is an incremental update item.

[0341] Conflict resolution mechanism: When there is a conflict between strategies at different levels (such as a parent node requesting load reduction while a child node needs to maintain power supply), a game theory model is constructed to solve for the Nash equilibrium.

[0342]

[0343] 7.4 Calibration effect verification and feedback loop:

[0344] (1) Virtual simulation stress test, extreme scenario injection: Simulate extreme weather (such as a sudden rise in temperature) in a digital twin environment. If there is a device malfunction, verify the robustness of the calibrated strategy:

[0345]

[0346] Multi-objective Pareto front analysis: Comparing the Pareto fronts of the strategies before and after calibration to quantify the optimization effect of the energy efficiency-delay-stability trade-off.

[0347]

[0348] (2) Long-term performance degradation monitoring and degradation rate modeling: fitting the calibrated strategy performance indicators (such as...) The decline curve of )

[0349]

[0350] When the rate of decay At that time, the full retraining of the self-learning model in Part 7 is triggered.

[0351] Cross-period data alignment: Aligning load characteristic distributions across different time periods and verifying data consistency using the maximum mean difference (MMD).

[0352]

[0353] in, This is the kernel function.

[0354] This method provides an end-to-end, visualized, multi-level load coordination management solution, achieving full automation from data processing, structure generation, model building, strategy optimization to result display, thus improving the system's intelligence level and management efficiency. This management method boasts excellent data integration performance, strong dynamic adaptability, and achieves multi-objective trade-offs, completing intelligent load coordination management and significantly improving the adaptability of load coordination management strategies.

[0355] Example 2

[0356] In this embodiment, a dynamic aggregate visualization multi-level load collaborative management system is adopted. The system includes a data processing and integration module, a dynamic aggregate generation module, a hierarchical load modeling module, a collaborative strategy optimization module, and a strategy evaluation and visualization module.

[0357] The data processing and integration module acquires power system operation data, user energy consumption behavior data, and environmental parameters, and performs structured processing on these three types of data to obtain a structured test dataset. The dynamic aggregate generation module generates a dynamic aggregate by hierarchically partitioning the data based on resource characteristics and spatial distribution features in the structured test dataset, creating a tree-like or mesh-like topology. The hierarchical load modeling module extracts the time-domain and frequency-domain characteristic parameters of each level of load in the dynamic aggregate and constructs a hierarchical load behavior model based on these parameters. The collaborative strategy optimization module constructs a multi-objective optimization function based on the dynamic aggregate topology and the hierarchical load behavior model, combined with constraints, and generates a collaborative management strategy based on this function. The strategy evaluation and visualization module optimizes the collaborative management strategy using real-time feedback, virtual verification, and long-term learning mechanisms, and constructs a hierarchical visualization mapping rule. This rule is used to output the execution results of the dynamic aggregate, the hierarchical load behavior model, and the collaborative management strategy as a graphical representation.

[0358] In this system, each module operates on the core logic of data-driven, functional progression, and closed-loop optimization, forming a tight unidirectional flow and bidirectional feedback connection, as detailed below:

[0359] Basic data flow chain: The data processing and integration module provides the core input to the system. Its output structured test dataset is directly transmitted to the dynamic aggregate generation module, supporting topology partitioning and dynamic aggregate generation. The dynamic aggregate topology output by the dynamic aggregate generation module, together with the time-domain / frequency-domain feature parameters extracted by the hierarchical load modeling module, serves as the input to the collaborative strategy optimization module, providing the data and model foundation for the construction of multi-objective optimization functions. The collaborative management strategy generated by the collaborative strategy optimization module is finally passed to the strategy evaluation and visualization module for subsequent optimization and graphical output.

[0360] Closed-loop optimization feedback chain: The strategy evaluation and visualization module transmits deviation data in strategy execution back to the collaborative strategy optimization module through real-time feedback and virtual verification mechanisms, realizing dynamic adjustment of the strategy; at the same time, the historical optimization experience accumulated by its long-term learning mechanism will be synchronously fed back to the hierarchical load modeling module to update the parameters of the hierarchical load behavior model, improve the model prediction accuracy, and form a closed loop of data-model-strategy-evaluation-optimization.

[0361] The specific implementation method of this system is the same as that in Example 1; the data processing and integration module realizes the standardized integration of multi-source heterogeneous data (power operation, user behavior, environmental parameters) to avoid the problem of data silos; the hierarchical load modeling module combines time domain / frequency domain characteristics with hierarchical structure to accurately depict the operating rules of different load levels, and can adapt to changes in load characteristics through long-term learning and dynamic iterative model; the collaborative strategy optimization module is based on multi-objective optimization function to balance core requirements such as delay and stability, and generates scientific management strategies, which are more in line with the actual power grid operation scenario compared with traditional single-objective strategies.

[0362] The dynamic aggregate generation module clearly presents the hierarchy and spatial relationship of loads through tree / mesh topology, making it easy for managers to quickly locate key load nodes; the strategy evaluation and visualization module transforms abstract models and strategies into graphical representations, reducing the technical threshold for managers and reducing decision-making time; the closed-loop optimization mechanism reduces the frequency of manual intervention, lowers the labor costs of operation and maintenance, and avoids power grid risks caused by human decision-making bias.

[0363] Real-time feedback and virtual verification mechanisms can promptly correct strategy deviations, effectively suppress the impact of load fluctuations on grid frequency and voltage, and improve system stability. Precise load behavior models can optimize power allocation, reduce redundant power supply, and lower energy losses. In addition, visualized outputs facilitate collaboration among multiple entities such as the grid dispatch center and user-side management platform, helping to achieve integrated power generation, grid, load, and storage, and promoting the transformation of the power system towards low-carbon and intelligent operation.

[0364] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for dynamic aggregate visualization and multi-level load collaborative management, characterized in that, The method steps include: S1. Acquire power system operation data, user energy consumption behavior data, and environmental parameters, and perform structured processing on these three types of data to obtain a structured test dataset; S2. Based on the resource characteristics and spatial distribution features in the structured test dataset, hierarchical division is performed to generate a tree structure. On the basis of the tree structure, long-range connections across levels are added through reconnection probability to form a network structure and generate a dynamic aggregate. S3. Extract the time-domain and frequency-domain characteristic parameters of the load at each level in the dynamic aggregate, and construct a hierarchical load behavior model based on the characteristic parameters; S4. Based on the dynamic aggregate topology and hierarchical load behavior model, and combined with constraints, construct a multi-objective optimization function, and generate a collaborative management strategy based on the multi-objective optimization function; S5. Optimize collaborative management strategies using three mechanisms: real-time feedback, virtual verification, and long-term learning. Construct hierarchical visualization mapping rules and use these rules to output the execution results of dynamic aggregates, hierarchical load behavior models, and collaborative management strategies as graphical representations. The hierarchical load behavior model in S3 has a hierarchical structure, consisting of a first level and a second level. In the first level, a device-level load behavior model is established to describe the start-up and shutdown patterns of individual devices. In the second level, the mutual influence between devices is modeled by superimposing the first level and introducing a synergistic effect term, thus constructing a hierarchical aggregation model. In this hierarchical load behavior model, the autoregressive moving average model is used to predict the load recovery trend after abrupt changes at time scales of minutes and below, while the LSTM network is used to capture nonlinear time-series dependencies at time scales of hours and above.

2. The dynamic aggregate visualization multi-level load collaborative management method according to claim 1, characterized in that, The power system operation data in S1 includes node voltage, frequency, line power, and equipment status; User energy consumption behavior data includes user-side load curves; environmental parameters include temperature, humidity, and light intensity; the structured processing includes: converting the above three types of data into a unified tensor representation, detecting and cleaning outliers, unifying the data format, modeling the spatial correlation of nodes, supplementing missing data, verifying the correlation of data, and checking the integrity of data.

3. The dynamic aggregate visualization multi-level load collaborative management method according to claim 1, characterized in that, The specific process of S2 includes: Based on a structured test dataset, quantitative indicators and spatial distribution characteristics of resource characteristics are established. Among them, the quantitative indicators of resource characteristics include controllability indicators and response timeliness indicators; the spatial distribution characteristics include communication cost and electrical correlation. A tree-like hierarchy is generated based on resource characteristic quantitative indicators and spatial distribution features. The node with the highest controllability index and the best geographical centrality is selected as the root node of the entire hierarchy. Based on communication cost and electrical correlation, nodes other than the root node are assigned to subtrees through K-means clustering. Based on the tree structure, long-range connections across levels are added by reconnection probability to form a network structure; Once the hierarchical structure is established, tasks and information transfer on it must follow the task allocation mechanism and information transfer protocol. The task allocation mechanism is as follows: calculate the load index of each node; when allocating tasks, first sort the nodes according to the load index, and prioritize nodes with load indices less than or equal to a preset load threshold; then update the node priority sorting according to real-time response timeliness indicators. The information transmission protocol specifically includes intra-layer communication rules, cross-layer communication rules, and information compression rules. The intra-layer communication rules stipulate that nodes at the same level use a broadcast mode, with intra-layer communication bandwidth accounting for 70% of the total bandwidth resource. The cross-layer communication rules stipulate that nodes at different levels use a unicast mode, with cross-layer communication bandwidth accounting for 30% of the total bandwidth resource, and the cross-layer communication latency must be less than or equal to 100ms. ms .

4. The dynamic aggregate visualization multi-level load collaborative management method according to claim 1, characterized in that, The time-domain and frequency-domain characteristic parameters of each load level extracted in step S3 include steady-state fluctuation patterns and sudden change response characteristics. The steady-state fluctuation patterns include steady-state fluctuation amplitude and periodic intensity; the sudden change response characteristics include sudden change amplitude and sudden change recovery time. The frequency domain feature parameters include frequency band energy distribution and frequency domain energy entropy; and the time domain and frequency domain feature parameters also include correlation weight indicators, which include cross-level correlation and spatial correlation weight.

5. The dynamic aggregate visualization multi-level load collaborative management method according to claim 1, characterized in that, The multi-objective optimization function in S4 includes a response delay penalty term and a system stability penalty term; the specific formula of the multi-objective optimization function is as follows: ； ； ； in, To optimize decision variables; and These are the weighting coefficients; In response to the delay penalty, the greater the delay, the higher the penalty value; Total number of devices; k For equipment serial number; The current time; For the first k Response time of each device; The exponential decay coefficient; This is a penalty item for system stability. and These are the weighting coefficients; For time t System frequency deviation at that time; For time t The voltage at that time is within acceptable limits; This is the system's rated voltage; This is the voltage deviation term.

6. The dynamic aggregate visualization multi-level load collaborative management method according to claim 5, characterized in that, The collaborative management strategy in S4 is a recursive strategy, specifically including a first decision level and a second decision level, and the collaborative management strategy also includes a negotiation mechanism based on game theory; The inputs to the first decision level are global load forecasts and system constraints, and a global strategy is generated based on these inputs. The goal of this global strategy is to minimize the system stability penalty term; then, this global strategy is transformed into a set of inter-level collaborative instructions and output. The input to the second decision level is the hierarchy weight and local constraints. After receiving the inter-hierarchical collaborative instruction set, the second decision level combines the input to decompose the task into sub-levels, thereby calculating the local policy for each device or sub-region. Finally, the local policy is converted into device-level operation instructions and output. A negotiation mechanism based on game theory: When there is a conflict between instructions at different levels, a non-cooperative game model is constructed, and the optimal strategy is finally solved through Nash equilibrium.

7. The dynamic aggregate visualization multi-level load collaborative management method according to claim 6, characterized in that, The real-time feedback in S5 is as follows: for the current measurement, multi-dimensional feedback data is collected in real time, including instruction completion rate, response delay distribution and load deviation; performance degradation detection is performed based on multi-dimensional feedback data. When the instruction completion rate is less than the preset completion rate threshold or the response delay distribution is greater than the preset delay distribution threshold, it is determined to be performance degradation. An isolated algorithm is used to detect abnormal feedback data points, calculate the abnormal score, and optimize the strategy. The optimization process of the policy at this time includes: updating the policy parameters using online gradient descent, adjusting the global policy based on the federated learning framework, and optimizing the local policy based on reinforcement learning. Virtual verification involves constructing a power system simulation model that includes disturbance scenarios, injecting the current strategy into the simulation model, and running a Monte Carlo simulation. Next, the failure probability of the strategy is statistically analyzed. If the failure probability of the strategy is higher than the preset failure probability threshold, the strategy optimization process is carried out. At this time, the strategy optimization process is as follows: a robustness penalty term is added to the optimization objective, the strategy is optimized based on adversarial training, and finally the adversarial examples are solved by projective gradient descent. The long-term learning process involves: constructing a database to store real-time operational data in time series; calculating the volatility of hierarchical weights and the error between predicted and actual loads based on the real-time operational data; constructing an incremental self-learning model, using a federated learning framework to collaboratively train the self-learning model across hierarchical devices, and using this self-learning model to complete the policy optimization process; constructing a closed-loop calibration and optimization process, periodically comparing the deviation between predicted and actual loads, adjusting the hierarchical division threshold based on the deviation, and using reinforcement learning to adjust parameters based on real-time operational data, updating the policy through the PPO algorithm; fitting the decay curve of the calibrated policy performance index based on real-time operational data, performing cross-cycle data alignment, aligning the load characteristic distribution of different time periods, and verifying data consistency through the maximum mean difference; when the decay rate at a certain point on the decay curve exceeds a preset decay rate threshold or the data consistency is less than a preset consistency threshold, the above incremental self-learning model is used to optimize the policy.

8. The dynamic aggregate visualization multi-level load collaborative management method according to claim 1, characterized in that, The visualization mapping rules in S5 specifically include aggregate topology mapping rules, load status mapping rules, and strategy execution result mapping rules; The aggregate topology mapping rule is used to output a graphical representation of the dynamic aggregate. Specifically, based on the hierarchical partitioning result of the dynamic aggregate, a force-directed graph or hierarchical layout algorithm is used to dynamically render nodes and edges. Node attributes include node size and node color; node size maps to hierarchical weights, and node color encodes the node's response timeliness score. Edge attributes include edge thickness and solid / dashed lines; edge thickness represents communication bandwidth utilization, and solid / dashed lines are used to distinguish inter-hierarchical collaboration rules. Node dragging and scaling are supported, as is triggering local topology detail expansion. The load status mapping rule is used to output the hierarchical load behavior model as a graphical representation, including the process of layering and overlaying time-series load curves and the process of highlighting abrupt events. When layering and overlaying time-series load curves, the load curves are aggregated according to the hierarchy and displayed on top of the equipment load curves, with transparency layering. When highlighting abrupt events, the abrupt response events are marked with pulse waveforms and associated with the corresponding equipment status labels. The strategy execution result mapping rule is used to output the execution result of the collaborative management strategy as a graphical representation, specifically using a multi-objective optimization index dashboard and a strategy execution heatmap. The multi-objective optimization index dashboard dynamically displays the delay penalty and stability index of the current strategy, using a circular progress bar or heatmap, and compares the preset reference threshold with the actual value after the current strategy is executed. The strategy execution heatmap is superimposed on the aggregate topology map with a strategy execution intensity matrix.

9. The dynamic aggregate visualization multi-level load collaborative management method according to claim 8, characterized in that, The specific formula for the strategy execution strength matrix is ​​as follows: in, Let be the policy execution intensity value of the i-th and j-th topological positions in the aggregate topology at time t; N is the total number of devices participating in policy execution in the aggregate. Let k be the weight of device k; Let be the real-time active power of device k at time t; This refers to the equipment's rated maximum active power.

Images