A power consumption load feature extraction system based on multi-source information agent

Abstract

The application discloses a kind of based on multi-source information intelligent agent's electric load feature extraction system, it is related to load sampling technical field.The system includes multiple information processing intelligent agents and at least one comprehensive intelligent agent.For receiving original electric power related data and extracting the basic feature vector of this source domain.Task perception interface is used to obtain the feature requirement information of target analysis model.Dynamic collaborative scheduler dynamically selects the intelligent agent set participating in collaboration according to feature requirement information, and allocates real-time computing resources and data sampling strategy for each selected intelligent agent.Feature collaborative recombination engine receives the basic feature vector output by selected intelligent agent, and carries out cross-domain collaborative fusion, generates the electric load feature tensor adapted to target analysis model.Effectively solve the problem of insufficient multi-source information fusion depth and low feature recombination accuracy, significantly improve the flexibility of feature extraction and the adaptation ability to complex analysis model.

Description

Technical Field

[0001] This invention relates to load sampling technology, and more specifically, to a power load feature extraction system based on a multi-source information intelligent agent. Background Technology

[0002] With the deepening construction of new power systems, electricity loads are exhibiting high randomness, volatility, and complexity. To achieve accurate load forecasting, power quality analysis, and non-intrusive load monitoring, extracting high-dimensional, multi-scale electricity load characteristics has become crucial. Traditional load acquisition and extraction methods often rely on a single data source and a fixed sampling frequency, making it difficult to consider both steady-state and transient characteristics, and unable to dynamically adjust extraction strategies according to the actual needs of different analysis tasks.

[0003] Publication number CN105373968A discloses an intelligent power data acquisition system. By establishing a multi-level server architecture at the provincial, municipal, district / county, and large enterprise / residential community levels, it achieves hierarchical collection and real-time feedback of user power data, providing data support for the dispatch system's analysis of operational modes. However, this technical solution primarily focuses on the physical transmission architecture and storage of data, lacking deep integration of multi-source information (such as environment, user behavior, distributed power sources, etc.). In terms of feature extraction, this solution typically only provides basic power consumption data, making it difficult to output high-dimensional feature tensors for complex analysis models, and it cannot achieve dynamic allocation of computing resources under task-driven conditions.

[0004] Publication number CN106952464A discloses an intelligent data acquisition system and method. This system connects various instruments such as water, electricity, and gas meters via an intelligent data acquisition device, and utilizes RS485 communication and multiple protocol tests to achieve unified acquisition and monitoring of heterogeneous instrument data. However, this technical solution only addresses simple aggregation and display of multi-source data, leaving the data sources in an "information silo" state and lacking a cross-domain feature fusion mechanism based on intelligent agent collaboration. Especially when facing transient events such as load surges or grid failures, this solution cannot achieve time axis alignment and collaborative reorganization of multi-source features, resulting in insufficient representational ability of the generated feature vectors for complex operating conditions, making it difficult to meet the stringent requirements of modern power intelligent analysis models for feature quality and adaptability.

[0005] The aforementioned problems indicate that existing electricity load acquisition and processing systems still have significant shortcomings in terms of multi-source information fusion depth, task awareness flexibility, feature recombination accuracy, and edge computing collaboration capabilities. Therefore, this invention provides an electricity load feature extraction system based on multi-source information intelligent agents. It aims to achieve dynamic resource scheduling and cross-domain feature fusion based on task requirements through a collaborative mechanism of multiple specialized information processing intelligent agents and a comprehensive intelligent agent, thereby generating a highly adaptable electricity load feature tensor to support advanced analysis and intelligent decision-making in the power grid. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a power load feature extraction system based on multi-source information intelligent agents.

[0007] To solve the above-mentioned technical problems, the technical solution of the present invention is: a power load feature extraction system based on a multi-source information intelligent agent, comprising: Multiple information processing agents, each uniquely corresponding to a preset type of information port, are used to receive raw electricity-related data collected by their corresponding port and extract the basic feature vector that forms the source domain. At least one integrated intelligent agent is communicatively connected to multiple information processing intelligent agents; The integrated intelligent agent includes: The task-aware interface is used to obtain feature requirement information for at least one target analysis model. A dynamic collaborative scheduler is connected to the task awareness interface and multiple information processing agents respectively. It is used to dynamically select a set of agents to participate in the collaboration from the multiple information processing agents according to the feature requirement information, and allocate real-time computing resources and data sampling strategies to each selected agent. The feature-coordinated recombination engine, connected to the dynamic coordinated scheduler, is used to receive the basic feature vectors output by the selected agent according to the data sampling strategy, and to perform cross-domain coordinated fusion of these feature vectors from different source domains based on the feature requirement information to generate an electricity load feature tensor adapted to the target analysis model.

[0008] Furthermore: the plurality of said information processing agents include at least two of the following types: The low-frequency steady-state intelligent agent is connected to the smart meter's data acquisition port and is used to extract the daily load curve, peak-valley difference, and load rate statistical features as basic feature vectors. An environmental context intelligent agent connects to a meteorological data interface, a calendar interface, or an electricity price system interface to extract temperature, humidity, holiday markers, and electricity price signal context features as basic feature vectors. A high-frequency transient intelligent agent is connected to the power grid transient waveform acquisition port to extract harmonic, interharmonic, and transient event waveform features as basic feature vectors. The user behavior intelligent agent connects to the user-side intelligent terminal or demand response system to extract user response behavior and electricity consumption habit behavior features as basic feature vectors. Distributed power generation intelligent agents, connected to the distributed power generation monitoring port, are used to extract output fluctuations and power prediction characteristics of photovoltaic and wind power as basic feature vectors. This ensures the system covers diverse data sources, thus addressing the issue of insufficient feature extraction comprehensiveness. The system is forced to integrate data from multiple key domains, avoiding feature gaps caused by a single data source. Each agent type, based on its connection port and extracted features, specifically addresses the coverage problem of particular data sources, collectively ensuring the comprehensiveness and adaptability of the basic feature vectors, providing multi-dimensional support for subsequent cross-domain collaborative integration.

[0009] Furthermore: the dynamic cooperative scheduler includes an event-triggered alignment module; The event triggering alignment module is configured to: monitor the power grid operation status in real time; when a preset type of event is detected, take the time of the event as the time reference, force the data sampling strategies of multiple information processing agents to be synchronously adjusted to a preset time window sampling mode centered on the event, and align the basic feature vectors output by each agent within the time window in the feature collaborative recombination engine on the time axis. The preset event types include at least one of power grid failure, voltage sag, and load surge.

[0010] The event-triggered alignment module addresses the issue of feature time alignment when power grid events occur, ensuring the accuracy and consistency of feature fusion. This module monitors the power grid's operational status in real time, promptly capturing preset events and preventing feature acquisition delays due to monitoring latency. When a preset type of event is detected, the module uses the event's occurrence time as the time reference and forcibly adjusts the data sampling strategies of all information processing agents to ensure all data is collected around the same event time point, eliminating sampling time bias. Furthermore, the module aligns the basic feature vectors output by each agent within a preset time window in the feature collaborative reorganization engine, directly resolving inconsistencies in the feature vectors over time and improving the reliability of the fusion results.

[0011] Furthermore: the information processing intelligent agent has a built-in data quality perception unit; The data quality sensing unit is used to evaluate the integrity or signal-to-noise ratio of the input data at its corresponding port in real time, and generate a corresponding confidence coefficient to be added to the output basic feature vector. When allocating computing resources, the dynamic collaborative scheduler dynamically adjusts the sampling weight of the corresponding agent based on the confidence coefficient. For agents with a confidence coefficient lower than a preset threshold, its data sampling frequency is reduced or its participation in collaboration is suspended. The dynamic collaborative scheduler is also used to pass the confidence coefficient to the feature collaborative recombination engine, so that it can be used as the basis for feature weighting during cross-domain collaborative fusion.

[0012] Each information processing agent incorporates a data quality awareness unit to monitor input data quality in real time and dynamically adjust resource allocation and feature fusion based on the generated confidence coefficient, thereby addressing feature inaccuracies caused by data quality issues. When allocating computing resources, the dynamic collaborative scheduler dynamically adjusts the sampling weights of corresponding agents based on the confidence coefficient. For agents with confidence coefficients below a preset threshold, their data sampling frequency is reduced or their participation in collaboration is suspended. This prioritizes high-quality data sources, optimizes resource utilization efficiency, and prevents computing resources from being wasted on unreliable data. The dynamic collaborative scheduler also transmits the confidence coefficient to the feature collaborative reorganization engine, which uses it as a basis for feature weighting during cross-domain collaborative fusion. This ensures that high-quality features are given higher weights during the fusion process, improving the accuracy and reliability of the final feature tensor.

[0013] Furthermore: the plurality of information processing agents include high-frequency transient agents deployed at edge acquisition terminals; The high-frequency transient agent has a built-in waveform fingerprint encoder, which is used to compress the original transient waveform into a low-dimensional waveform fingerprint vector in real time at the edge. The complete transient waveform features are only uploaded to the integrated agent when the waveform fingerprint vector matches the templates in the pre-stored fault template library more than a preset threshold. The dynamic collaborative scheduler determines whether to include the high-frequency transient agent in the current collaborative task based on the matching degree information, and allocates a higher sampling frequency to it when it is included.

[0014] By introducing a waveform fingerprint encoder and matching mechanism into the high-frequency transient agent, the complete waveform features are only uploaded when a potential fault is detected, thereby effectively reducing unnecessary data transmission and computational resource consumption.

[0015] By introducing a waveform fingerprint encoder and matching mechanism into the high-frequency transient agent, complete waveform features are uploaded only when a potential fault is detected, effectively reducing unnecessary data transmission and computational resource consumption. The high-frequency transient agent deployed at the edge acquisition terminal uses the waveform fingerprint encoder to compress the original transient waveform into a low-dimensional waveform fingerprint vector in real time at the edge, avoiding direct transmission of the original waveform data and reducing network burden. When the waveform fingerprint vector matches a pre-stored fault template in a library with a pre-set threshold, the complete transient waveform features are uploaded, ensuring that critical data is transmitted only when a fault event occurs, reducing redundant information processing. The dynamic collaborative scheduler determines whether to include the agent in a collaborative task based on the matching degree information and allocates a higher sampling frequency upon inclusion, optimizing resource allocation and ensuring high-precision data acquisition during important events while avoiding resource waste in non-fault scenarios.

[0016] Furthermore: when the feature demand information obtained by the task perception interface comes from the load prediction model, the dynamic collaborative scheduler selects a first information processing agent corresponding to the meteorological port and a second information processing agent corresponding to the historical load port; The feature collaborative reorganization engine adopts a temporal attention interaction mechanism, using the meteorological feature vector output by the first information processing agent as the query sequence and the historical load feature vector output by the second information processing agent as the key value sequence to calculate the meteorological-load cross-attention map for future periods, which is then output to the load prediction model as the electricity load feature tensor.

[0017] Triggered through the task-aware interface, the dynamic collaborative scheduler accurately selects the agents corresponding to the meteorological and historical load ports, ensuring that relevant data sources are prioritized and avoiding interference from irrelevant features. The feature collaborative reorganization engine adopts a temporal attention interaction mechanism, using meteorological feature vectors as query sequences and historical load feature vectors as key-value sequences. This design utilizes the query sequence to actively guide the attention allocation of the key-value sequences, efficiently capturing the temporal dependency between meteorological changes and load fluctuations. The calculated meteorological-load cross-attention map directly maps the key interaction patterns of future periods, thereby generating a highly adapted feature tensor output to the load prediction model. This solves the problem of insufficient temporal correlation capture in traditional fusion mechanisms, improving feature adaptability and prediction accuracy.

[0018] Furthermore: when the task perception interface simultaneously acquires feature requirement information from the power quality analysis model and the non-intrusive load decomposition model, the feature collaborative recombination engine performs frequency band segmentation processing on the basic feature vector output by the same high-frequency transient agent; The frequency band segmentation process includes: decomposing the basic feature vector into high-frequency detail components and low-frequency envelope components through wavelet packet decomposition; routing the high-frequency detail components to the output channel corresponding to the power quality analysis model; and routing the low-frequency envelope components to the output channel corresponding to the non-intrusive load decomposition model, thereby generating two different power load feature tensors.

[0019] By employing a frequency band segmentation mechanism, the efficiency issue of feature processing when simultaneously addressing the needs of multiple models is resolved. Specifically, when the task-aware interface detects the simultaneous existence of feature requirements from both the power quality analysis model and the non-intrusive load decomposition model, the feature collaborative recombination engine performs frequency band segmentation on the output of the same high-frequency transient agent. This segmentation is achieved through wavelet packet decomposition, decomposing the basic feature vector into high-frequency detail components and low-frequency envelope components. The high-frequency detail components are routed to the power quality analysis model, as power quality analysis requires high-frequency details to capture transient events such as harmonics, while the low-frequency envelope components are routed to the non-intrusive load decomposition model, as load decomposition relies more on low-frequency envelopes to identify electricity usage habits. In this way, the engine can efficiently generate two independent feature tensors, each adapting to the needs of different models, avoiding feature redundancy and wasted computational resources, and improving the system's flexibility and feature adaptability.

[0020] Furthermore, the dynamic collaborative scheduler also includes a resource monitoring unit; The resource monitoring unit is used to obtain the CPU and memory load of the computing node where the integrated intelligent agent is located in real time; When the load exceeds a preset threshold, the dynamic collaborative scheduler automatically shuts down the calls to one or more information processing agents with the highest data sampling frequency, generates simplified feature tensors based only on the remaining agents, and resumes full collaboration after the load returns to normal. The simplified feature tensor is generated by reducing the feature dimension or by using a lightweight fusion algorithm.

[0021] By introducing a resource monitoring unit and an adaptive adjustment mechanism, the resource optimization problem under high computing node load is solved, ensuring that the system can still operate efficiently when resources are scarce. The system can perceive the resource status of computing nodes in real time, providing a basis for subsequent adjustments. The resource monitoring unit is used to obtain the CPU and memory load of the computing nodes where the integrated agents reside in real time. By continuously monitoring load data, the system can promptly identify resource bottlenecks. When the load exceeds a preset threshold, the dynamic collaborative scheduler automatically shuts down calls to one or more information processing agents with the highest data sampling frequency. This prioritizes reducing the resource consumption of high-frequency sampling agents and avoids system overload. Simplified feature tensors are generated only based on the remaining agents, and simplified features are output using the remaining agents, ensuring uninterrupted feature extraction when resources are insufficient. Full collaboration is restored after the load returns to normal, and all agent calls are restored when the load decreases, maintaining the integrity of feature extraction. Simplified feature tensors are generated by reducing feature dimensions or using lightweight fusion algorithms. By simplifying feature generation methods, computational complexity is reduced, adapting to low-resource environments.

[0022] Furthermore, it also includes a third information processing intelligent agent corresponding to the distributed power information port, used to extract the output fluctuation characteristics of photovoltaic or wind power; When the task perception interface obtains the feature demand information from the source-grid-load-storage coordinated control model, the feature collaborative reorganization engine performs correlation analysis on the output fluctuation characteristics and load characteristics of the third information processing agent, generates source-load matching degree characteristics, and incorporates them into the power load feature tensor. The correlation analysis includes calculating the Pearson correlation coefficient or mutual information between the output fluctuation characteristics and the load characteristics over time.

[0023] By introducing a specialized agent to handle distributed power sources and performing correlation analysis under specific requirements, the problem of missing source-load matching characteristics is addressed, thereby improving the adaptability of the feature tensor to the source-grid-load-storage coordination model. The feature collaborative reorganization engine performs correlation analysis between output fluctuation characteristics and load characteristics to generate source-load matching characteristics. This enhances the depth and practicality of feature representation by quantifying the statistical relationship between distributed power source output and load. Incorporating source-load matching characteristics into the electricity load feature tensor allows the output characteristics to directly adapt to the input requirements of the source-grid-load-storage coordination model, improving the overall system compatibility. The correlation analysis includes calculating the Pearson correlation coefficient or mutual information between output fluctuation characteristics and load characteristics over time, providing a reliable measurement method based on time-series statistics to ensure the accuracy and robustness of the matching degree calculation.

[0024] Furthermore, the integrated intelligent agent also includes an adaptation monitor connected to the output of the feature collaborative recombination engine; The fit monitor is used to periodically calculate the mutual information entropy between the generated electricity load characteristic tensor and the input of the target analysis model. When the mutual information entropy is lower than a preset threshold, the adaptability monitor triggers the dynamic collaborative scheduler to readjust the selection strategy and resource allocation weights for information processing agents until the mutual information entropy recovers to above the threshold.

[0025] By introducing a fit monitor, the problem of feature tensors not being adequately adapted to the target analysis model is solved, ensuring that the generated features have high fit and guaranteeing that the feature tensors ultimately meet the model requirements, thereby improving the analysis accuracy.

[0026] The main technical effects of this invention are reflected in the following aspects: By constructing a collaborative architecture of multiple specialized information processing agents and a comprehensive agent, task-driven dynamic resource scheduling and cross-domain feature fusion are achieved, effectively solving the problems of insufficient depth of multi-source information fusion, poor task perception flexibility, low accuracy of feature recombination, and weak edge computing collaboration capabilities. Specifically, each of the multiple information processing agents uniquely corresponds to a preset type of information port, ensuring that each agent focuses on data processing in a specific source domain. By receiving raw electricity-related data and extracting basic feature vectors, the problem of insufficient depth of multi-source information fusion is solved because this specialized design avoids data silos and improves the targeting and completeness of feature extraction. The comprehensive agent communicates with multiple information processing agents, acting as a coordination center to integrate the outputs of each agent, enhancing the overall collaborative capability of the system. The task perception interface obtains feature requirement information from the target analysis model, enabling the system to perceive external task requirements and solving the problem of poor task perception flexibility, because this interface mechanism allows the system to dynamically adjust the feature extraction strategy according to the specific analysis model. The dynamic collaborative scheduler dynamically selects a set of agents to participate in the collaboration based on feature demand information, and allocates real-time computing resources and data sampling strategies to each selected agent. This solves the problems of poor task awareness flexibility and weak edge computing collaboration capabilities, because the selection and allocation mechanism based on feature demand information achieves resource optimization and adaptive scheduling, avoiding the rigidity of fixed strategies. The feature collaborative recombination engine receives the basic feature vectors output by the selected agents according to the data sampling strategy, and performs cross-domain collaborative fusion based on feature demand information to generate an electricity load feature tensor. This solves the problem of low feature recombination accuracy, because the demand-driven fusion mechanism ensures efficient integration and adaptability of features from different source domains, improving the matching degree of the feature tensor to the target model. Attached Figure Description

[0027] Figure 1 This invention presents a schematic diagram of the overall architecture of a power load feature extraction system based on a multi-source information intelligent agent. Figure 2: Workflow diagram of the dynamic collaborative scheduler in event-triggered mode; Figure 3 : A schematic diagram of frequency band segmentation processing performed by the feature-coordinated recombination engine. Detailed Implementation

[0028] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, so that the technical solution of the present invention can be more easily understood and mastered.

[0029] like Figure 1 As shown, the present invention describes a power load feature extraction system based on multi-source information agents. Its core architecture consists of multiple information processing agents and at least one integrated agent. The multiple information processing agents are logically independent, and each agent uniquely corresponds to a pre-defined type of information port. Here, "information port" refers to various physical or logical interfaces capable of collecting or receiving raw power consumption-related data, such as the digital port of a smart meter, the network interface for meteorological data, and the analog input port of a transient waveform acquisition card. Each information processing agent is responsible for receiving the raw data collected by its corresponding port and using a built-in specific algorithm to perform preliminary processing on the data from the source domain to extract a basic feature vector that characterizes the core attributes of the data source. For example, the low-frequency steady-state agent connected to the smart meter acquisition port is responsible for extracting statistical features such as daily load curves, peak-to-valley differences, and load rates from long-term power consumption data, organizing this information into a multi-dimensional vector; while the high-frequency transient agent connected to the power grid transient waveform acquisition port is responsible for extracting harmonic and interharmonic content, as well as transient event waveform parameters, from high-frequency sampled data as its basic feature vector.

[0030] The integrated intelligent agent, acting as the central coordination and processing unit of the system, establishes communication connections with all information processing agents. It integrates three key modules: a task-aware interface, a dynamic cooperative scheduler, and a feature cooperative reorganization engine. The task-aware interface serves as the window for interaction between the system and external analysis models, used to obtain feature requirement information for one or more target analysis models. This "feature requirement information" can be a specific instruction specifying the type of features needed, such as "provide features for the next 24 hours of load forecasting," or it can be metadata about the analysis model itself, such as model structure and input layer dimensions. The dynamic cooperative scheduler connects the task-aware interface and all information processing agents, and it is the core of dynamic resource allocation. This scheduler parses the requirement information from the task-aware interface and dynamically selects the set of agents from multiple information processing agents to participate in the current cooperative task based on this information. For example, if the task requirement only involves steady-state analysis, high-frequency transient agents may not be selected. Simultaneously, the dynamic cooperative scheduler allocates real-time computing resources to each selected agent, such as CPU time slices or memory buffer sizes, and issues specific data sampling strategies, such as sampling frequency and sampling time windows. The feature collaborative recombination engine, connected to the dynamic collaborative scheduler, is responsible for receiving the basic feature vectors output by the selected agent according to a predetermined sampling strategy. Since these vectors come from different source domains and are heterogeneous, the engine needs to execute a cross-domain collaborative fusion algorithm based on the feature requirement information provided by the task-aware interface. This "cross-domain collaborative fusion" is not a simple vector concatenation, but rather the integration of features from different dimensions into a unified, information-rich electricity load feature tensor through mathematical transformations, alignment, and weighting. This tensor is directly accepted and used by the target analysis model. Through this architecture, the system achieves end-to-end task-driven and dynamic optimization from data acquisition to feature generation, significantly improving the targeting and effectiveness of feature extraction.

[0031] To ensure the system can cover comprehensive data sources, the information processing agent can specifically include, but is not limited to, the following types. A low-frequency steady-state agent, connected to the smart meter's data acquisition port, can extract a basic feature vector that can be denoted as... Here, each element represents average load, peak load, valley load, peak occurrence time, and load factor, respectively. The environmental context agent, connected to a meteorological data interface, calendar interface, or electricity pricing system interface, extracts a feature vector that can be denoted as... These represent temperature, humidity, holiday indicators, and real-time electricity price signals, respectively. The high-frequency transient intelligent agent, connected to the power grid transient waveform acquisition port, can extract feature vectors that may include the amplitudes of various harmonics. With phase subscript Represents the harmonic order and the energy of transient events. and duration A user behavior intelligent agent, connected to a user-side intelligent terminal or demand response system, outputs a feature vector that quantifies the degree of user response to stimuli, such as... ,in The probability of user participation in demand response. The system encodes typical user electricity consumption patterns. This pattern encoding can be achieved by classifying users' historical daily load curves using clustering algorithms to obtain category labels, or by representing them using low-dimensional embedding vectors extracted by an autoencoder. The distributed power generation agent, connected to the distributed power generation monitoring port, can extract feature vectors that may include the volatility of photovoltaic output. And the predicted value of wind power output The configuration of these five intelligent agents ensures that the system can acquire information from multiple dimensions, including steady state, environment, transient state, behavior, and source, thus guaranteeing the comprehensiveness of the basic feature vector.

[0032] like Figure 2 As shown, the dynamic coordinated dispatcher integrates an event-triggered alignment module to address the multi-source characteristic time synchronization problem under power grid transient events. This module continuously monitors the power grid's operating status parameters, such as RMS voltage, frequency, and total harmonic distortion (THD). When a preset type of event is detected, such as a voltage sag exceeding 10% or a load change exceeding 20% ​​within one second, the module immediately uses the event occurrence time as the time of the event. Using [a specific time reference], a mandatory synchronization command is broadcast to all activated information processing agents. Upon receiving the command, the data sampling strategies of each agent are uniformly adjusted to [a specific time reference]. Centered preset time window Sampling mode. Subsequently, each agent will collect and process data within this time window, outputting its basic feature vector. Since all vectors correspond to the same absolute time window, the feature collaborative recombination engine can directly align these vectors on the time axis after receiving them, eliminating timing deviations caused by sampling delays or clock asynchrony of various data sources. The key effect of this technique is that when analyzing transient processes such as power grid faults, it can ensure that data with different rates of change, such as meteorological conditions, user behavior, and distributed power output, accurately correspond to the power grid transient process, thereby fusing a load feature tensor that can accurately characterize the system state during the event and avoiding distortion of analysis conclusions due to feature misalignment.

[0033] To further enhance the system's robustness, each information processing agent also incorporates a data quality awareness unit. This unit assesses data integrity in real time—whether data packets are continuous and without loss—and the signal-to-noise ratio while the agent receives raw data. The evaluation result is expressed as a confidence coefficient. The form of manifestation, The value range is [0, 1], where 1 represents completely reliable data quality. This confidence coefficient is appended as metadata to the basic feature vector and sent together with it to the integrated intelligent agent. The dynamic cooperative scheduler makes decisions based on the received confidence coefficient when allocating computing resources. For example, for a confidence coefficient lower than a preset threshold... If an agent's data is deemed unreliable by the scheduler, the scheduler will dynamically reduce its data sampling frequency or even suspend its participation in the current collaborative task, thus prioritizing valuable computing resources for agents providing high-quality data. Simultaneously, the dynamic collaborative scheduler also passes all received confidence coefficients to the feature collaborative recombination engine. The engine uses these coefficients as the basis for feature weighting during cross-domain collaborative fusion. Specifically, the fused feature tensor... It can be generated through weighted averaging or weighted concatenation, for example. ,in For the first The system generates a basic feature vector for each agent. Through this mechanism, the system can proactively avoid the influence of unreliable data sources, ensuring that the final generated feature tensor has high reliability and optimizing the overall utilization efficiency of computing resources.

[0034] To address the bandwidth and computational burden of high-frequency data processing, the system has specifically optimized the high-frequency transient agent deployed at the edge acquisition terminal. This agent incorporates a waveform fingerprint encoder, the core of which is a pre-trained autoencoder neural network. The training process of this autoencoder is as follows: a large amount of historical transient waveform data is collected as training samples, each sample... The original waveform corresponds to a time window; the sample is input into the encoder to obtain a low-dimensional fingerprint vector. The waveform is then reconstructed using a decoder. The training objective is to minimize the reconstruction error, such as the mean squared error. Once training is complete, the encoder can be used online to process the raw transient waveforms acquired in real time. Compressed into fingerprint vector At the edge, the agent locally computes the fingerprint vector and compares it with template vectors pre-stored in the local fault template library. The matching degree, for example, by calculating cosine similarity. Only when the maximum matching degree Exceeding a preset threshold Only then does the agent determine that the current waveform may correspond to a potential fault, and uploads the complete transient waveform features to the integrated agent. Conversely, only the low-dimensional fingerprint vector is uploaded. Alternatively, no data may be uploaded. Upon receiving matching information, the dynamic collaborative scheduler determines subsequent resource allocation based on this information. If the matching degree is high, the scheduler will include the high-frequency transient agent in a higher-priority collaborative task and allocate a higher sampling frequency to it for more detailed event tracking. This technique, by filtering and compressing data at the edge, effectively reduces the impact of massive transient data on the communication network and central computing nodes, achieving event-driven, precise data acquisition and resource focus.

[0035] The feature collaborative reorganization engine operates with high flexibility to meet the specific needs of different target analysis models. For example, when the feature requirement information obtained by the task-aware interface comes from the load forecasting model, the dynamic collaborative scheduler selects the first information processing agent corresponding to the meteorological port (i.e., the environmental context agent) and the second information processing agent corresponding to the historical load port (i.e., the low-frequency steady-state agent). The feature collaborative reorganization engine does not perform simple vector concatenation at this point, but instead uses a temporal attention interaction mechanism to generate features. Specifically, the engine uses the meteorological feature vectors for future time periods output by the first agent as the query sequence. The historical load feature vector output by the second agent is used as the key sequence. Sum sequence Through the attention calculation formula A weather-load cross-attention map was calculated, in which... Key vector The dimension is used to scale the dot product result to prevent gradient vanishing. Each element in this graph represents the weight of the influence of future weather conditions on historical load at various time points, effectively capturing the temporal dependency between weather changes and load fluctuations. The final output electricity load feature tensor is this attention graph rich in interactive information, which serves as the direct input to the load forecasting model, significantly improving forecast accuracy.

[0036] For example, such as Figure 3 As shown, when the task-aware interface simultaneously acquires feature requirement information from the power quality analysis model and the non-intrusive load decomposition model, the feature collaborative recombination engine performs frequency band segmentation processing on the basic feature vector output by the same high-frequency transient agent. This processing is achieved through wavelet packet decomposition. Let the original feature vector output by the high-frequency transient agent be... It is essentially a frequency domain or time domain representation of the original waveform. The engine executes it. Layer wavelet packet decomposition yields Each component has a frequency component of equal bandwidth. These components are then recombined according to the requirements of the two target models. For example, high-frequency detail components containing rich information (such as coefficients corresponding to high-frequency subbands in wavelet packet decomposition) are reconstructed to form a feature tensor. This signal is then routed to the output channel of the power quality analysis model for analyzing transient events such as harmonics and interharmonics. Simultaneously, the low-frequency envelope component (corresponding to the coefficients of the low-frequency subband), which contains the overall signal trend, is reconstructed to form a characteristic tensor. This data is then routed to the output channel of a non-intrusive load decomposition model to identify the switching and operating status of electrical equipment. In this way, the same raw data source is efficiently parsed and adapted to two completely different analysis tasks, avoiding the waste of resources from repeated data collection and feature calculation.

[0037] Considering the limited computing resources in real-world deployment environments, the dynamic cooperative scheduler also includes a resource monitoring unit. This unit monitors the hardware resource usage of the computing node where the integrated intelligent agent resides in real time, primarily the CPU load rate. and memory usage When the comprehensive load index Exceeding the preset overload threshold At that time, the dynamic collaborative scheduler initiates resource throttling mode. Among these, and The weighting coefficients are preset based on the system's sensitivity to CPU and memory resources, and satisfy the following conditions: For example, all values ​​can be set to 0.5. In this mode, the scheduler automatically identifies one or more information processing agents with the highest data sampling frequency in the current collaborative task and temporarily disables calls to these agents to free up computing resources. The system generates a simplified feature tensor based only on the remaining agents with lower resource requirements. This simplified feature tensor can be generated by reducing the feature dimension, for example, by outputting only the first few principal components after principal component analysis, or by using a lighter-weight fusion algorithm such as direct averaging instead of a complex attention mechanism. The resource monitoring unit continuously monitors the load, and once... Once the data drops below the normal threshold, the scheduler automatically resumes full collaborative calls to all agents, regenerating the complete feature tensor. This adaptive mechanism ensures that the system can still provide basic services when resources are scarce, achieving stability and reliability in complex environments.

[0038] In scenarios involving distributed energy resources, the system uses a specialized distributed power source agent, namely a third information processing agent, to extract the output fluctuation characteristics of photovoltaic or wind power, such as the output variance. or range When the task-aware interface acquires characteristic demand information from the source-grid-load-storage coordinated control model, the feature collaborative reorganization engine, in addition to fusing regular load characteristics, will also perform a specialized correlation analysis. It will process the output fluctuation characteristic sequence output by the third information processing agent. With load feature sequence from low-frequency steady-state agent Alignment is performed temporally. Then, the statistical correlation between the two is calculated, such as the Pearson correlation coefficient: Or calculate their mutual information. This was used to quantify the nonlinear dependency between the two. The calculated... Alternatively, the mutual information value can be considered the source-load matching characteristic, which is incorporated as a new dimension into the final output load characteristic tensor. This characteristic intuitively reflects the degree of synchronization between distributed power generation output and local power load in terms of changing trends, and is a crucial piece of information necessary for the source-grid-load-storage coordinated control model to perform optimal scheduling.

[0039] Finally, to ensure the system can continuously generate high-quality features, the integrated agent also includes a fitness monitor. This monitor is connected to the output of the feature co-reorganization engine, and its input periodically acquires the electricity load feature tensor generated by the engine. This includes the expected input data distribution of the current target analysis model. The fitness monitor calculates the mutual information entropy between the two. To quantify the fit of the feature tensor to the model, where This represents the target output of the model. Mutual information entropy can be calculated using kernel density estimation or a nonparametric estimation method based on k-nearest neighbors. A higher mutual information entropy value indicates that the feature tensor contains more information for predicting the model output. The richer the information, the better. When the calculated mutual information entropy is lower than a preset fitness threshold... This indicates that the currently generated features may lack sufficient information or deviate from the model's requirements. At this point, the fitness monitor sends a feedback signal to the dynamic cooperative scheduler, triggering the scheduler to re-evaluate and adjust its strategy. This could involve reselecting the cooperative agent combination or allocating new resource weights to each agent to generate more informative features. This closed-loop feedback adjustment process continues until the mutual information entropy recovers above the threshold. This technique enables the system to self-optimize, adapting to slow changes in the analysis model or data environment and maintaining a high level of feature extraction quality over the long term.

[0040] Of course, the above are just typical examples of the present invention. In addition, the present invention may have many other specific embodiments. All technical solutions formed by equivalent substitution or equivalent transformation fall within the scope of protection claimed by the present invention.

Claims

1. A power load feature extraction system based on a multi-source information intelligent agent, characterized in that, include: Multiple information processing agents, each uniquely corresponding to a preset type of information port, are used to receive raw electricity-related data collected by their corresponding port and extract the basic feature vector that forms the source domain. At least one integrated intelligent agent is communicatively connected to multiple information processing intelligent agents; The integrated intelligent agent includes: The task-aware interface is used to obtain feature requirement information for at least one target analysis model. A dynamic collaborative scheduler is connected to the task awareness interface and multiple information processing agents respectively. It is used to dynamically select a set of agents to participate in the collaboration from the multiple information processing agents according to the feature requirement information, and allocate real-time computing resources and data sampling strategies to each selected agent. The feature-coordinated recombination engine, connected to the dynamic coordinated scheduler, is used to receive the basic feature vectors output by the selected agent according to the data sampling strategy, and to perform cross-domain coordinated fusion of these feature vectors from different source domains based on the feature requirement information to generate an electricity load feature tensor adapted to the target analysis model.

2. The power load feature extraction system based on a multi-source information intelligent agent as described in claim 1, characterized in that: The plurality of said information processing agents include at least two of the following types: The low-frequency steady-state intelligent agent is connected to the smart meter's data acquisition port and is used to extract the daily load curve, peak-valley difference, and load rate statistical features as basic feature vectors. An environmental context intelligent agent connects to a meteorological data interface, a calendar interface, or an electricity price system interface to extract temperature, humidity, holiday markers, and electricity price signal context features as basic feature vectors. A high-frequency transient intelligent agent is connected to the power grid transient waveform acquisition port to extract harmonic, interharmonic, and transient event waveform features as basic feature vectors. The user behavior intelligent agent connects to the user-side intelligent terminal or demand response system to extract user response behavior and electricity consumption habit behavior features as basic feature vectors. The distributed power intelligent agent connects to the distributed power monitoring port and is used to extract the output fluctuation and power prediction characteristics of photovoltaic and wind power as basic feature vectors.

3. The power load feature extraction system based on a multi-source information intelligent agent as described in claim 1, characterized in that: The dynamic cooperative scheduler includes an event-triggered alignment module; The event triggering alignment module is configured to: monitor the power grid operation status in real time; when a preset type of event is detected, take the time of the event as the time reference, force the data sampling strategies of multiple information processing agents to be synchronously adjusted to a preset time window sampling mode centered on the event, and align the basic feature vectors output by each agent within the time window in the feature collaborative recombination engine on the time axis. The preset event types include at least one of power grid failure, voltage sag, and load surge.

4. The power load feature extraction system based on a multi-source information intelligent agent as described in claim 1, characterized in that: The information processing intelligent agent has a built-in data quality perception unit. The data quality sensing unit is used to evaluate the integrity or signal-to-noise ratio of the input data at its corresponding port in real time, and generate a corresponding confidence coefficient to be added to the output basic feature vector. When allocating computing resources, the dynamic collaborative scheduler dynamically adjusts the sampling weight of the corresponding agent based on the confidence coefficient. For agents with a confidence coefficient lower than a preset threshold, its data sampling frequency is reduced or its participation in collaboration is suspended. The dynamic collaborative scheduler is also used to pass the confidence coefficient to the feature collaborative recombination engine, so that it can be used as the basis for feature weighting during cross-domain collaborative fusion.

5. The power load feature extraction system based on a multi-source information intelligent agent as described in claim 1, characterized in that: The plurality of information processing agents include high-frequency transient agents deployed at edge acquisition terminals; The high-frequency transient agent has a built-in waveform fingerprint encoder, which is used to compress the original transient waveform into a low-dimensional waveform fingerprint vector in real time at the edge. The complete transient waveform features are only uploaded to the integrated agent when the waveform fingerprint vector matches the templates in the pre-stored fault template library with a preset threshold. The dynamic collaborative scheduler determines whether to include the high-frequency transient agent in the current collaborative task based on the matching degree information, and allocates a higher sampling frequency to it when it is included.

6. The power load feature extraction system based on a multi-source information intelligent agent as described in claim 1, characterized in that: When the feature demand information obtained by the task perception interface comes from the load prediction model, the dynamic collaborative scheduler selects a first information processing agent corresponding to the meteorological port and a second information processing agent corresponding to the historical load port. The feature collaborative reorganization engine adopts a temporal attention interaction mechanism, using the meteorological feature vector output by the first information processing agent as the query sequence and the historical load feature vector output by the second information processing agent as the key value sequence to calculate the meteorological-load cross-attention map for future periods, which is then output to the load prediction model as the electricity load feature tensor.

7. The power load feature extraction system based on a multi-source information intelligent agent as described in claim 1, characterized in that: When the task perception interface simultaneously acquires feature requirement information from the power quality analysis model and the non-intrusive load decomposition model, the feature collaborative recombination engine performs frequency band segmentation processing on the basic feature vector output by the same high-frequency transient agent. The frequency band segmentation process includes: decomposing the basic feature vector into high-frequency detail components and low-frequency envelope components through wavelet packet decomposition; routing the high-frequency detail components to the output channel corresponding to the power quality analysis model; and routing the low-frequency envelope components to the output channel corresponding to the non-intrusive load decomposition model, thereby generating two different power load feature tensors.

8. The power load feature extraction system based on a multi-source information intelligent agent as described in claim 1, characterized in that: The dynamic collaborative scheduler also includes a resource monitoring unit; The resource monitoring unit is used to obtain the CPU and memory load of the computing node where the integrated intelligent agent is located in real time; When the load exceeds a preset threshold, the dynamic collaborative scheduler automatically shuts down the calls to one or more information processing agents with the highest data sampling frequency, generates simplified feature tensors based only on the remaining agents, and resumes full collaboration after the load returns to normal. The simplified feature tensor is generated by reducing the feature dimension or by using a lightweight fusion algorithm.

9. The power load feature extraction system based on a multi-source information intelligent agent as described in claim 1, characterized in that: It also includes a third information processing intelligent agent corresponding to the distributed power information port, used to extract the output fluctuation characteristics of photovoltaic or wind power; When the task perception interface obtains the feature demand information from the source-grid-load-storage coordinated control model, the feature collaborative reorganization engine performs correlation analysis on the output fluctuation characteristics and load characteristics of the third information processing agent, generates source-load matching degree characteristics, and incorporates them into the power load feature tensor. The correlation analysis includes calculating the Pearson correlation coefficient or mutual information between the output fluctuation characteristics and the load characteristics over time.

10. The power load feature extraction system based on a multi-source information intelligent agent as described in claim 1, characterized in that: The integrated intelligent agent also includes an adaptation monitor, which is connected to the output of the feature collaborative recombination engine; The fit monitor is used to periodically calculate the mutual information entropy between the generated electricity load characteristic tensor and the input of the target analysis model. When the mutual information entropy is lower than a preset threshold, the adaptability monitor triggers the dynamic collaborative scheduler to readjust the selection strategy and resource allocation weights for information processing agents until the mutual information entropy recovers to above the threshold.

Images