Load resource data anomaly detection system based on novel power system

Abstract

The invention relates to a load resource data anomaly detection system based on a novel power system, which comprises a historical data management module, a real scene management module and a simulation generation module, the historical data management module is used for marking historical load data in a historical data load information base to generate a historical load sample. The whole system effectively breaks through the technical bottlenecks of insufficient samples and weak model generalization ability through multi-dimensional sample generation, dynamic compensation correction and adversarial training, and provides an efficient and reliable solution for load anomaly detection of a novel power system.

Description

Technical Field

[0001] The present invention relates to the field of power grid data analysis, and more specifically, to a load resource data anomaly detection system based on a new type of power system. Background Art

[0002] At present, with the popularization of new energy technologies and the wide application of new energy equipment in the market, the new type of power system based on new energy has become a crucial topic in the current power grid power supply system. The current mainstream charging piles are divided into 220V single-phase charging piles and 380V three-phase charging piles. The 220V single-phase charging piles are generally used alone, that is, customized according to the load of new energy equipment, while the 380V three-phase charging piles are generally provided for public systems. As a new type of consumption mainstay in the power system, the electricity consumption characteristics and functions of charging piles are still constantly differentiating and being developed. Therefore, it is particularly difficult to analyze and detect anomalies in charging piles through power load data. Because under the premise of more extensive updated demand for load data, there is no very detailed model covering all characteristics. When training this model, the problems faced are that the currently accumulated sample quantity is insufficient, and traditional samples have the drawbacks of single function and fewer sample dimensions. The training models based on traditional samples are generally prone to misjudging new type of charging piles. And for new type of charging piles to form sufficient sample data, a large number of actual measurements and tests are required, with high difficulty. In this way, the technical judgment of whether it is abnormal by traditional sample training and classification will be missing due to the time lag. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a load resource data anomaly detection system based on a new type of power system.

[0004] To solve the above technical problems, the technical solution of the present invention is: a load resource data anomaly detection system based on a new type of power system:

[0005] It includes a historical data management module, a real-scene management module, and a simulation generation module. The historical data management module is configured with a historical data load information database, and the historical data management module is used to mark the historical load data in the historical data load information database to generate historical load samples;

[0006] The real-scene management module is used to collect the load data of the target charging pile in different working states through the real-scene load monitoring interface to generate real-scene load samples;

[0007] The simulation generation module is used to construct a simulation discharge model, and run it in the simulation discharge model by substituting the charging pile simulation parameters and load simulation parameters to obtain the corresponding simulation load samples;

[0008] The sample screening module is configured with a sample confidence algorithm for calculating the confidence of the load sample, and classifying the load sample into one type of load sample, two type of load sample and three type of load sample according to the confidence;

[0009] The first sample generation module includes an environmental deviation association unit and a first sample generation unit. The environmental deviation association unit generates environmental features of historical load samples through a preset environmental analysis strategy, calculates load deviations of real-life load samples and historical load samples with the same classification features, and associates the load deviations with the environmental features to generate an environmental association sub-model. The first sample generation unit retrieves the load deviation from the environmental association sub-model according to the input real-life simulated environmental features, and processes the real-life load samples according to the load deviations to generate environmental compensation load samples.

[0010] The second sample generation module includes a hardware deviation association unit and a second sample generation unit, wherein the hardware deviation association unit calculates the simulated load samples and the real-scene load samples with the same classification characteristics to obtain hardware deviation characteristics, and performs cluster analysis on the collected hardware deviation characteristics to obtain each hardware deviation index cluster; the second sample generation unit matches the corresponding hardware deviation index cluster through the simulated load sample to retrieve the corresponding hardware deviation characteristics, and processes the corresponding simulated load sample through the hardware deviation characteristics to generate a hardware compensation load sample;

[0011] The training execution module includes inputting a type of load sample, an environmental compensation load sample, and a hardware compensation load sample into an abnormal training model, and inputting a type of load sample into a first sample generation module or a second sample generation module.

[0012] Further: it also includes an adversarial generation module, the adversarial generation module is configured with an abnormal waveform database, the abnormal waveform database stores abnormal waveform features, the abnormal waveform features have a matching index set, the matching index set includes a number of matching sub-features, each matching sub-feature can be indexed to a corresponding abnormal waveform feature, each abnormal waveform feature corresponds to a matching abnormal sub-value, the adversarial generation module is configured with an adversarial generation strategy, the adversarial generation strategy is used to extract matching sub-features in three types of load samples and call a number of corresponding abnormal waveform features, replace the abnormal waveform features with the load samples through abnormal load constraints to generate abnormal load samples, and the abnormal load constraints are that the total value of the replacement anomaly of the abnormal load sample falls within a preset replacement threshold range;

[0013] The training execution module brings the abnormal load samples into the abnormal training model.

[0014] Furthermore, it further includes a data acquisition subsystem, which is used to collect the load data of each charging pile to generate historical load samples; the historical data management module is configured with a data re-acquisition unit and a data correction unit, and the data correction unit is used to extract the discredited features in the historical load samples judged as class II load samples. The data re-acquisition unit generates corresponding re-acquisition instructions according to the discredited features and sends the re-acquisition instructions to the data acquisition subsystem.

[0015] Furthermore, the acquisition subsystem is configured with a virtual inspection load, which is used to access the charging pile and execute the corresponding re-acquisition instruction. The re-acquisition instruction includes load parameters, and the load parameters are used to configure the corresponding virtual inspection load.

[0016] Furthermore, the real-scene management module is configured with a function mirroring unit. The function mirroring unit generates corresponding function classification judgment models by controlling the target charging pile to work in different function states, analyzes the historical load samples through the function classification judgment models to generate corresponding charging pile working sequences, and controls the working states of the target charging pile according to the charging pile working sequences to generate real-scene load samples corresponding to the historical load samples.

[0017] Furthermore, the environmental analysis strategy includes

[0018] Step S1: Obtain the location information of the charging pile corresponding to the historical load sample;

[0019] Step S2: Retrieve the charging pile numbers with location relevance from the pre-constructed charging pile distribution topology model according to the location information;

[0020] Step S3: Obtain the corresponding relevant environmental information according to the charging pile numbers;

[0021] Step S4: Identify reliable environmental features in the relevant environmental information through a preset predation deduction model to generate the environmental feature data.

[0022] Furthermore, the hardware deviation unit is configured with a deviation classification sub-strategy, and the deviation classification sub-strategy includes

[0023] Step A1: Subtract the simulated load waveform and the real-scene load waveform under the same working conditions to obtain a hardware difference waveform;

[0024] Step A2: Classify the hardware difference waveform through a preset feature splitting strategy to obtain the hardware difference features corresponding to each hardware type item;

[0025] Step A3: Repeat Step A1 until all the simulated load waveforms are collected;

[0026] Step A4: Mark the hardware intervention degree for the hardware difference features of each hardware type item through a preset deviation calculation algorithm;

[0027] Step A5: Perform mean processing on the hardware difference features through the hardware intervention degree marking to obtain the corresponding difference mean features;

[0028] Step A6: Combine different hardware type items to obtain a hardware deviation cluster;

[0029] Step A7: In the hardware deviation cluster, perform clustering analysis on the difference mean features through a clustering analysis algorithm to obtain the corresponding hardware deviation index cluster.

[0030] Furthermore: The first sample generation module is configured with a model construction strategy, and the model construction strategy includes

[0031] Step B1: Split the environmental features into a link of several environmental sub-features, obtain environmental feature response items, and each environmental feature response item corresponds to a response sensitivity value configured for the environmental sub-feature;

[0032] Step B2: Construct an environmental feature network with the environmental sub-features as nodes;

[0033] Step B3: Calculate the comprehensive sensitivity value of each environmental sub-feature through a preset response sensitivity algorithm;

[0034] Step B4: Determine whether any adjacent environmental sub-features meet the transposition condition through a preset transposition extension algorithm. If the transposition condition is met, swap the positions of the two environmental sub-features in the environmental feature network until all environmental sub-features no longer meet the transposition condition.

[0035] Furthermore: The second sample generation unit is configured with a sample random strategy, and the sample random strategy includes generating corresponding random parameter items and corresponding random variable ranges according to the hardware deviation features, generating random variable values for each random parameter item according to the random variable ranges, generating a random alternative load waveform according to the random variable values, substituting the random alternative load waveform into the simulation load sample to generate the hardware compensation load sample, and correcting the corresponding hardware deviation features.

[0036] The technical effects of the present invention are mainly reflected in the following aspects: The load resource data anomaly detection system based on the new power system significantly improves the accuracy and adaptability of abnormal detection of charging pile load data through the collaborative design of multiple modules. First of all, the combination of the historical data management module and the real scene management module realizes the complementarity of historical samples and real-time data, solving the problem of single-dimensional traditional samples. The simulation generation module effectively makes up for the deficiency of measured data of new charging piles by constructing a simulation discharge model. The confidence classification mechanism of the sample screening module can dynamically optimize the sample quality. Combining the dual compensation strategies of environmental deviation correlation and hardware deviation correlation, more realistic load samples are generated through environmental feature compensation and hardware deviation correction respectively, thus improving the adaptability of the model to complex working conditions. The introduction of the adversarial generation module enhances the recognition robustness of the model to unknown abnormal patterns by simulating abnormal waveform features, solving the misjudgment problem caused by the lag of traditional methods. The virtual inspection load and re-sampling mechanism of the data acquisition subsystem ensure the authenticity and integrity of the data, and the collaborative work of the functional mirror unit and the environmental analysis strategy further improves the pertinence of sample generation. Through multi-dimensional sample generation, dynamic compensation and correction, and adversarial training, the whole system effectively breaks through the technical bottlenecks of insufficient samples and weak model generalization ability, providing an efficient and reliable solution for load anomaly detection in the new power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 : Schematic diagram of the architecture of a load resource data anomaly detection system based on the new power system of the present invention;

[0038] Figure 2 : Schematic diagram of the architecture of the historical data management module of the present invention;

[0039] Figure 3 : Schematic diagram of the architecture of the real scene management module of the present invention;

[0040] Figure 4 : Schematic diagram of the architecture of the first sample generation module of the present invention;

[0041] Figure 5 : Schematic diagram of the architecture of the second sample generation module of the present invention.

[0042] Reference numerals: 100, historical data management module; 101, historical data load information database; 102, data re - acquisition unit; 103, data correction unit; 200, real - scene management module; 201, real - scene load monitoring interface; 202, function mirroring unit; 300, simulation generation module; 400, sample screening module; 500, first sample generation module; 510, environmental deviation correlation unit; 520, first sample generation unit; 600, second sample generation module; 610, hardware deviation correlation unit; 620, second sample generation unit; 700, training execution module; 800, adversarial generation module; 801, abnormal waveform database; 900, data acquisition subsystem. Detailed implementation manners

[0043] The following further details the specific implementation manners of the present invention in conjunction with the accompanying drawings, so that the technical solutions of the present invention are easier to understand and master.

[0044] A load resource data anomaly detection system based on a new - type power system: including a historical data management module 100, a real - scene management module 200, and a simulation generation module 300. The historical data management module 100 is configured with a historical data load information database 101. The historical data management module 100 is used to mark the historical load data in the historical data load information database 101 to generate historical load samples. The historical data management module 100 serves as the basic data support unit of the entire anomaly detection system. Its core function is to construct and maintain a structured historical load information database. By deeply mining and feature - marking the existing charging pile operation data, it provides a representative sample set for subsequent model training. The construction process of this module first needs to integrate historical load data from different data sources, including voltage and current waveforms collected from the grid side, working state parameters reported by the charging pile controller, and auxiliary information such as temperature and humidity recorded by environmental monitoring devices. These data are stored in a time - series form at a fixed sampling frequency (for example, 100 sampling points per second), forming an original data set containing multi - dimensional features such as time stamps, active power, reactive power, power factor, and environmental temperature. In particular, the historical load data adopts a hierarchical storage structure. The bottom layer is the unprocessed original binary stream data, the middle layer is converted into a structured table form through data parsing and feature engineering, and the top layer is further aggregated into statistical features with different time granularities such as hourly and daily according to the application scenario requirements.

[0045] When generating historical load samples, the module adopts an event-driven marking strategy. First, the continuous time series is segmented into sample units of fixed length through the sliding window technique (for example, each sample contains 10 seconds of continuous data), and then each sample is labeled according to a preset rule base. For example, when it is detected that the charging current drops suddenly by more than 80% within 3 seconds and the voltage fluctuates by more than ±15%, this sample will be marked as "abnormal charging interruption"; if the active power has been stable within the range of 95%-105% of the rated power for 5 consecutive minutes, it will be marked as "normal charging". It should be noted that the module introduces a dynamic weight allocation mechanism, which assigns different confidence coefficients to each sample according to the proximity of the sample generation time. For example, the weight coefficient of data in the past three months is 1.2, while the weight coefficient of data over one year old drops to 0.8, so as to balance the contradiction between data timeliness and model generalization ability. In addition, the module also supports the generation of samples based on the combination of working conditions. For example, the high-temperature environment data and low-temperature environment data in the fast charging mode are aggregated respectively to form composite samples with environmental-condition coupling characteristics, providing basic data support for the subsequent environmental deviation compensation model. This data processing method not only retains the temporal characteristics of the original data.

[0046] The real - scene management module 200 is used to collect the load data of the target charging pile in different working states through the real - scene load monitoring interface 201 to generate real - scene load samples; as the core data collection unit of the anomaly detection system, the core function of the real - scene management module 200 is to directly access the target charging pile through the real - time monitoring interface, and dynamically collect the load data of the device in different working states in a unified test environment to build a real and effective sample library. The construction process of this module first requires physically connecting the charging pile under test to a dedicated monitoring system to ensure that key load parameters such as voltage, current, and power factor during the operation of the device can be obtained in real time, and at the same time, real - time environmental data such as temperature, humidity, and grid fluctuations in the environment where the device is located are synchronously recorded. The built - in function mirroring unit 202 of the module can accurately control the automatic switching of the charging pile in different function modes such as fast charging, slow charging, and reserved charging by parsing the working sequence model constructed from historical load samples, and continuously collect complete load waveform data in each mode, thereby generating real - scene load samples covering a variety of working conditions. In particular, the module adopts a dynamic compensation mechanism. By real - time monitoring the changes in environmental parameters and adjusting the data collection strategy, it effectively eliminates the influence of external interference on data quality, ensuring that the generated samples can truly reflect the operating characteristics of the device under specific working conditions. It should be noted that by controlling the coordinated operation of different types of charging piles in the same test environment, this module can not only obtain the load characteristics of a single device, but also simulate the interaction data in complex scenarios such as multiple piles in parallel, providing more comprehensive real - scene data support for model training. This centralized data collection method not only ensures the controllability of sample generation, but also enhances the diversity of samples through multi - dimensional working condition simulation, thus significantly enhancing the adaptability and accuracy of the anomaly detection model to the actual operation scenario.

[0047] The real-scene management module 200 is configured with a function mirroring unit 202. The function mirroring unit 202 controls the target charging pile to work in different functional states to generate corresponding function classification judgment models, analyzes historical load samples through the function classification judgment models to generate corresponding charging pile working sequences, and controls the working state of the target charging pile according to the charging pile working sequences to generate real-scene load samples corresponding to the historical load samples. As the core component of the real-scene management module 200, the core role of the function mirroring unit 202 is to generate a working condition sequence highly consistent with historical data in real-scene tests by accurately simulating the charging pile working mode in historical load samples, so as to isolate the influence of environmental factors on load characteristics under controlled variable conditions. The construction process of this unit first needs to extract the load characteristics in different functional states from the historical data management module 100, such as the current step-up during fast charging mode, the constant voltage characteristic during slow charging mode, etc., and generate a function classification judgment model by constructing a hidden Markov model (HMM) based on time series. This model analyzes the transition probabilities between different working modes of the charging pile in historical data, such as the probability of fast charging to standby is 0.3, the probability of standby to reserved charging is 0.5, etc., establishes a state transition matrix, and combines the observed probability distributions of each state (such as voltage fluctuation range, power change rate, etc.) to realize the dynamic prediction of the charging pile working sequence.

[0048] Specifically, the training process of the function classification judgment model adopts an incremental learning strategy. When new historical load samples are input, the state transition matrix is updated by calculating the Kullback-Leibler divergence between the new and old samples to ensure that the model can adapt to the evolution of the functional characteristics of new types of charging piles. For example, when it is found that a certain batch of new types of charging piles has a unique trickle charging stage after fast charging, the model will automatically identify this new state and adjust the original state transition probability. When generating the charging pile working sequence, the unit first constructs a Markov chain according to the time distribution characteristics of each functional mode in historical data (such as the fast charging proportion is 60% in the evening on weekdays, and the slow charging proportion is 45% at noon on weekends), and then generates a specific working condition sequence through Monte Carlo sampling. For example, a working sequence including fast charging (18:00 - 19:30), standby (19:30 - 20:00), and reserved charging (20:00 - 22:00) is generated, and the time parameters of this sequence strictly match the statistical laws in historical data.

[0049] It should be noted that when the functional mirror unit 202 controls the physical charging pile to execute the work sequence, a dual closed-loop feedback mechanism is adopted: the outer loop triggers mode switching according to the preset time nodes, and the inner loop dynamically adjusts the charging strategy by real-time monitoring parameters such as voltage and current to ensure the consistency of the load waveform with the historical samples. For example, when the grid voltage fluctuation in the physical test causes the charging current to deviate from the historical sample curve, the unit will automatically adjust the PWM duty cycle to restore the current waveform to the target trajectory. This control strategy not only ensures the strict consistency of other variables except environmental parameters, but also eliminates the influence of hardware differences on the experimental results through the real-time compensation mechanism, thus providing pure input data for the environmental deviation correlation model.

[0050] The simulation generation module 300 is used to construct a simulation discharge model and obtain corresponding simulation load samples by substituting the simulated parameters of the charging pile and the simulated parameters of the load into the simulation discharge model. As the core component of the load resource data anomaly detection system based on the new power system, the main function of the simulation generation module 300 is to simulate the dynamic electrical behavior of the charging pile and its load by constructing a high-precision simulation discharge model, so as to generate simulation load samples covering multiple scenarios and working conditions, providing data support for the subsequent training of the anomaly detection model. The construction process of this module needs to comprehensively consider the physical characteristics, control strategies of the charging pile and charging equipment, and the influence of the actual operating environment. The core steps are as follows: First, the basic framework of the simulation discharge model needs to be established. This model is based on circuit theory and simulates the power conversion process of the charging pile through the combination of equivalent circuit elements (such as resistors, inductors, capacitors) and control modules (such as PWM modulators, current controllers). At the same time, the equivalent model of the load device (such as the battery charging curve, internal resistance change model) is introduced to reflect the actual load characteristics.

[0051] When constructing a simulation discharge model, special attention should be paid to the selection and setting of charging pile simulation parameters and load simulation parameters. Charging pile simulation parameters are mainly used to characterize the electrical characteristics of charging piles, such as rated voltage (e.g., 220V / 380V), rated current (e.g., 16A / 32A), power factor, conversion efficiency, AC side impedance characteristics, and charging control strategy parameters (such as constant current charging threshold, constant voltage charging cut-off voltage), etc. Taking a three-phase charging pile as an example, its simulation parameters may include three-phase voltage unbalance degree, zero-sequence current component, PWM modulation frequency, etc. These parameters directly affect the ability of the simulation model to reproduce the harmonic characteristics and power fluctuations during the actual operation of the charging pile. Load simulation parameters are used to describe the electrical behavior of charging devices. Typical parameters include battery type (such as lithium-ion battery, lead-acid battery), battery capacity (such as 50Ah / 100Ah), state of charge (SOC), equivalent internal resistance (such as milliohm-level dynamic internal resistance), polarization capacitance, temperature coefficient, and charging protocol parameters (such as communication instruction response time in CCS / CHAdeMO protocol). For example, for the charging process of electric vehicles, load simulation parameters need to include the open-circuit voltage curve of the battery pack, charging acceptance ability in different SOC intervals, and the limitation strategy of the thermal management system on the charging current.

[0052] The construction of the simulation discharge model needs to go through two key links: parameter calibration and dynamic verification. First, the model parameters are initially calibrated through the historical load samples in the historical data management module 100. For example, the equivalent circuit parameters are inferred by using the actually measured charging voltage and current waveforms. Secondly, an adaptive algorithm is introduced to dynamically correct the model. For example, the temperature compensation coefficient of the battery internal resistance model is adjusted based on the real-time collected on-site load samples (from the on-site management module 200) to ensure that the model can adapt to the charging characteristic changes under different environmental temperatures. During the simulation operation stage, by inputting different combinations of charging pile simulation parameters and load simulation parameters, simulation load samples covering various scenarios such as normal charging, abnormal charging (such as overvoltage, overcurrent), and special working conditions (such as low-temperature charging, fast charging) can be generated. These samples are compared and analyzed with the on-site load samples through the hardware deviation correlation unit 610, and then the hardware deviation characteristics are extracted for subsequent sample compensation processing.

[0053] It should be noted that when constructing the simulation discharge model, both model accuracy and computational efficiency need to be considered. The multi-time scale hybrid modeling technology is adopted, that is, the switching behavior of power electronic devices is simulated on the microsecond time scale, and the electrochemical process of the battery is simulated on the second time scale. At the same time, the reduced-order model technology is used to reduce the computational complexity. In addition, the model also needs to consider the dynamic influence of the grid side. For example, factors such as grid voltage fluctuations and harmonic pollution are taken as external disturbances and input to enhance the authenticity and diversity of the simulation load samples, so as to improve the adaptability of the anomaly detection model to complex grid environments.

[0054] The sample screening module 400 is configured with a sample confidence algorithm for calculating the confidence of load samples, and classifies the load samples into first-class load samples, second-class load samples, and third-class load samples according to the confidence.

[0055] As the core component of the load resource data anomaly detection system, the sample screening module 400 functions to achieve hierarchical management of multi-source load samples (historical, real-scene, simulation) through a confidence evaluation mechanism, providing high-confidence data support for subsequent model training. The construction process of this module integrates data feature analysis, statistical inference, and a dynamic feedback mechanism. First, multi-dimensional feature extraction needs to be performed on the input original load samples, including time-domain features (such as voltage amplitude, current effective value), frequency-domain features (such as harmonic content, frequency deviation), operating condition features (such as charging stage, power factor), and metadata features (such as equipment number, geographical location). Subsequently, these features are comprehensively evaluated through the sample confidence algorithm, and finally, a confidence score is output and classification is completed. For the design of the sample confidence algorithm, multi-dimensional variables need to be comprehensively considered to fully reflect the credibility of the samples. The key to algorithm design lies in quantifying the contribution degree of each dimension to the confidence and achieving adaptive adjustment through dynamic weight allocation. The following are the specific implementation methods for each dimension: Similarity of similar samples (weight α1): This dimension measures the data consistency by calculating the cosine similarity between the target sample and similar historical samples. The higher the similarity, the more the sample conforms to the typical characteristics. The calculation formula for the correlation coefficient is: where x i is the abscissa of the feature vector of the i-th sample similar to the target sample, y i is the ordinate of the feature vector of the i-th sample similar to the target sample, is the feature mean of the abscissa of the feature vector, is the feature mean of the ordinate of the feature vector, cr is the preset feature coefficient weight, and n is the total number of similar samples in the similar sample library.

[0056] Dissimilarity of similar samples (weight α2): This dimension uses the Euclidean distance to measure the difference degree between the target sample and other types of samples. The larger the distance, the more unique the sample is. The calculation formula for the dissimilarity index is: where x j is the target sample feature, y j is the feature mean of other types of samples. m is the total number of the similar sample library.

[0057] Regularity of load data (weight α3): Analyze the periodic characteristics of the load curve through Fourier transform and calculate the proportion of the main frequency component. The regularity index is defined as:

[0058] where P dominant is the main frequency power spectral density, Pk is the power spectral density of the k-th harmonic, and r is the number of load cycles. This parameter reflects whether the sample conforms to the typical working mode of the charging device.

[0059] Abnormal repair frequency (weight α4): This parameter is directly related to historical repair records, and counts the number of abnormal repairs Nfault of the target charging pile within a specific time period. The higher the repair frequency, the worse the operating stability of the device and the lower the credibility of the sample. The data is sourced from the fault logs of the power grid operation and maintenance management system.

[0060] Usage frequency (weight α5): Calculate the number of charging starts fuse per unit time of the charging pile. Too low usage frequency may lead to unrepresentative samples. This parameter is obtained through the operation logs of the acquisition subsystem and needs to be normalized in combination with the device type (such as public charging piles and private charging piles).

[0061] Instruction response timeliness (weight α6): Measure the response delay time tresp of the charging pile to control instructions, including operations such as charging mode switching and power adjustment. Samples with a response time exceeding the threshold (such as 500ms) will be deducted points. The data is collected in real time through the function mirror unit of the actual scene management module.

[0062] Switching action response (weight α7): Analyze the waveform smoothness when the charging pile switches between different working modes (such as constant current charging to constant voltage charging), and use the dynamic time warping (DTW) algorithm to calculate the difference Dswitch between the actual switching waveform and the standard waveform. The greater the difference, the more unstable the execution of the device control strategy.

[0063] Management category attribution (weight α8): Different weights are assigned according to the management category to which the charging pile belongs (such as Class A key monitoring equipment, Class B regular equipment). This parameter is predefined by the power grid management department and reflects the importance and monitoring priority of the device.

[0064] The comprehensive calculation of confidence uses a weighted linear combination model: Among them, the weight α i is determined by training through the analytic hierarchy process (AHP) combined with historical data to ensure the rationality of the contributions of each dimension.

[0065] Metrici is the preset weight ratio parameter of the i-th weight item. Samples with a confidence score higher than the threshold T1 are classified into one category, those between T1 and T2 are in the second category, and those lower than T2 are in the third category. The threshold setting needs to be optimized through cross-validation experiments, taking into account both data utilization and model robustness.

[0066] The first sample generation module includes an environmental deviation correlation unit and a first sample generation unit. The environmental deviation correlation unit generates the environmental characteristics of historical load samples through a preset environmental analysis strategy. The first sample generation module plays an important role in the load resource data anomaly detection system based on the new power system. It is mainly composed of an environmental deviation correlation unit and a first sample generation unit. Its function is to generate environmentally compensated load samples to improve the accuracy of abnormal detection of charging pile load data.

[0067] From the construction process, first, the environmental deviation correlation unit 510 comes into play. It generates the environmental characteristics of historical load samples according to a preset environmental analysis strategy. This environmental analysis strategy includes a series of steps. First, it obtains the location information of the charging pile corresponding to the historical load sample, then retrieves the charging pile numbers with location correlation from the pre-constructed charging pile distribution topology model according to this location information, then obtains the corresponding relevant environmental information according to these numbers, and finally identifies the reliable environmental characteristics in the relevant environmental information through a preset predation deduction model to generate environmental characteristic data. After that, the environmental deviation correlation unit 510 calculates the load deviation between the real-scene load sample and the historical load sample with the same classification characteristics, and correlates the load deviation with the environmental characteristics to generate an environmental correlation sub-model.

[0068] Subsequently, the first sample generation unit 520 retrieves the load deviation from the environmental correlation sub-model according to the input real-scene simulation environmental characteristics, and then processes the real-scene load sample according to this load deviation to finally generate an environmentally compensated load sample. It should be noted that in the environmental analysis strategy, the accurate acquisition of location information and the precise construction of the charging pile distribution topology model are crucial, because this will directly affect the accuracy of the subsequent relevant environmental information obtained, and thus affect the reliability of the environmental characteristic data. At the same time, in the first sample generation module 500, the quality of the environmental correlation sub-model generated by the environmental deviation correlation unit 510 will also affect the effect of the first sample generation unit 520 in generating environmentally compensated load samples. Only by ensuring the accuracy and reliability of each link can the first sample generation module 500 play a better role in the entire anomaly detection system.

[0069] The said environmental analysis strategy includes

[0070] Step S1: Obtain the location information of the charging pile corresponding to the historical load sample;

[0071] Step S2: Retrieve the charging pile numbers with location correlation from the pre-constructed charging pile distribution topology model according to the location information;

[0072] Step S3: Obtain the corresponding relevant environmental information according to the charging pile numbers;

[0073] Step S4: Identify reliable environmental features in the relevant environmental information through a preset predation deduction model to generate the environmental feature data. The specific implementation process of the environmental analysis strategy is as follows: First, it is necessary to obtain the charging pile location information corresponding to the historical load samples. These location information may come from the GPS positioning module of the charging pile itself or the geographic information system database, and its role is to provide a spatial positioning reference for subsequent analysis to facilitate the association of environmental variables in different geographical locations. Next, according to the location information, retrieve the charging pile numbers with location relevance from the pre-constructed charging pile distribution topology model. This distribution model is established through grid division or spatial clustering algorithms and can identify adjacent charging piles within a certain radius of the target charging pile. The environmental features of these adjacent charging piles may be similar or related to those of the target charging pile. Subsequently, obtain the corresponding relevant environmental information according to these charging pile numbers. These information may include meteorological data (such as temperature, humidity, wind speed), geographical information (such as altitude, terrain), social environment data (such as traffic flow, population density), etc., and need to be extracted from multi-source data such as external meteorological databases, remote sensing image libraries, and urban information platforms. Finally, identify reliable environmental features in the relevant environmental information through a preset predation deduction model. This model is constructed based on the sparrow predation algorithm, and its core idea is to simulate the behavior of a sparrow group quickly locating and screening high-quality food sources in a complex environment. The specific implementation process includes initializing the population individuals, calculating the fitness value of each individual (i.e., the correlation coefficient between the environmental feature and the load data), updating the position through the discoverer-joiner mechanism (the discoverer is responsible for exploring new areas, and the joiner follows the discoverer or conducts local searches), introducing anti-predation behavior (when detecting danger, some individuals quickly move to a safe area), and finally, through multiple iterations of optimization, screen out environmental features that have a significant correlation with the load data and high stability. It should be noted that reasonable parameters, such as population size, maximum number of iterations, early warning probability, etc., need to be set when constructing the predation deduction model. At the same time, the non-linear relationship between different environmental features should be considered to avoid feature selection bias caused by overfitting.

[0074] The core of the environmental analysis strategy lies in dynamically extracting reliable features of the environment where the charging pile is located through multi-dimensional data association and intelligent algorithms. Its process starts with obtaining the location information of historical load samples (step S1). This location information not only includes the physical coordinates of the charging pile but also its node identifier in the power grid topology, so as to accurately map its electrical connection relationship and spatial distribution characteristics when constructing the distribution topology model later. The charging pile distribution topology model retrieved in step S2 is a dynamic network based on a graph database. This model is constructed by fusing geographical information system (GIS) and power grid node parameters, and can reflect the spatial density of the charging pile cluster, the capacity constraints of the power supply lines, and the interactive influence of adjacent nodes. When the location of the target charging pile is input, the model will traverse the associated charging pile numbers that have electrical coupling or physical proximity to it through the breadth-first search algorithm, forming a set of devices with location relevance. Step S3 further extracts the real-time operation data (such as power fluctuation patterns) and environmental monitoring data (such as temperature and humidity, electromagnetic interference intensity) corresponding to these associated charging piles from the external environment database. These heterogeneous data are aligned by time stamps to form a multi-dimensional feature matrix, providing the original input for subsequent environmental feature mining.

[0075] In step S4, the predation deduction model, as the core algorithm for environmental feature extraction, draws its design inspiration from the efficient search and resource location capabilities demonstrated by the sparrow population during the foraging process. This model first abstracts the environmental information in the external database into a high-dimensional data space, where each data point corresponds to the environmental state under a specific time segment. Sparrow individuals are initially initialized as randomly distributed search agents, and their movement trajectories follow the "exploration - exploitation" dynamic balance mechanism: in the exploration stage, the agents perform long-distance jumps in the data space through the Lévy flight strategy to cover unknown areas and quickly lock in the sub-spaces where significant environmental features may exist; in the exploitation stage, the agents switch to a local refinement search mode, using an adaptive step size to perform gradient descent analysis on the data points within the target area to identify environmental parameters that have a strong correlation with load fluctuations. For example, when the charging pile group in a certain area shows a collective steep drop in load under high-temperature conditions, the agents will mark the correlation features between temperature and heat dissipation efficiency through cooperative pheromones, while suppressing the interference of irrelevant variables (such as light intensity). The model iteratively optimizes the position weights of the agents and finally outputs a set of environmental feature vectors with clear physical meanings, such as "coupling coefficient of surrounding transformer load rate - environmental temperature" and "influence factor of humidity gradient on insulation resistance". These features can not only explain the abnormal fluctuations of historical load data but also provide a deviation correction basis for the generation of subsequent environmental compensation load samples. The entire process is accelerated by a parallel computing framework to ensure the purification of features from massive environmental information within seconds, thus effectively overcoming the defects of low feature extraction efficiency and dimensional redundancy of traditional methods in complex scenarios.

[0076] And calculate the load deviation between the real-scene load samples and the historical load samples with the same classification features, associate the load deviation with the environmental features to generate an environmental association sub-model. The first sample generation unit 520 retrieves the load deviation from the environmental association sub-model according to the input real-scene simulation environmental features, and processes the real-scene load samples according to the load deviation to generate environmentally compensated load samples; when calculating the load deviation between the real-scene load samples and the historical load samples with the same classification features, it is first necessary to establish a multi-dimensional classification label system to define core parameters such as charging pile type, working mode, time period characteristics, and environmental temperature range, etc. These parameters are used as classification features. Subsequently, the time series phase difference between the two is aligned through the dynamic time warping algorithm. At the same time, a geographical grid coding system is introduced to map the sample coordinates into a spatial unit with a side length of 500 meters to achieve accurate physical location matching. During this process, sliding window mean filtering is performed on the basic load characteristics such as the mean active power and the standard deviation of reactive power. At the same time, for dynamic characteristics such as harmonic distortion rate and power fluctuation slope, wavelet packet decomposition technology is used to extract the harmonic components from the 2nd to the 15th order. For response characteristics such as load mutation recovery time and voltage sag tolerance, the step response analysis method is used to capture their dynamic characteristics, and then a deviation quantification model including weighted Euclidean distance is constructed, where the weight coefficients of each dimension are dynamically determined by the entropy weight method, and the contribution degrees of the harmonic distortion rate and the power fluctuation slope are particularly strengthened to cope with the non-linear characteristics of new charging piles. On this basis, it is necessary to establish an environmental sensitivity matrix to quantify the partial derivative effects of external variables such as temperature, humidity, and grid voltage on power parameters, and use an improved sparrow search algorithm to optimize the environmental compensation coefficient. 50 search agents are initialized for 200 iterative calculations, and the cumulative difference between the compensated deviation and the reference value is minimized as the fitness function, and finally the optimal compensation parameter combination is output. At the same time, the closed-loop verification mechanism inputs the samples with a matching degree higher than 95% into the simulation platform for cross-verification. If the compensation error exceeds 3%, the weight coefficient adjustment process is triggered, and the dynamic threshold mechanism corrects the abnormal boundary conditions in real time according to the grid load rate. When the real-time deviation breaks through the threshold, environmentally compensated samples are automatically generated. This multi-level deviation calculation framework not only integrates technologies such as spatio-temporal alignment, feature decoupling, and dynamic optimization, but also can continuously update the environmental association model through the online calibration function of virtual inspection loads, so as to effectively improve the analysis accuracy of the load characteristics of new charging piles under the condition of limited sample data.

[0077] The first sample generation module 500 includes a configured model construction strategy, and the model construction strategy includes

[0078] Step B1: A link that splits the environmental features into several environmental sub-features, obtains environmental feature response items, and each environmental feature response item configures a corresponding response sensitivity value for the environmental sub-feature;

[0079] Step B2: Construct an environmental feature network with environmental sub-features as nodes; that is, environmental features with the same environmental sub-features will be associated in the environmental feature network. Step B3: Calculate the comprehensive sensitivity value of each environmental sub-feature through a preset response-sensitive algorithm; algorithm steps and calculation formulas. First, it is necessary to collect environmental sub-feature data and connection information of the environmental feature network. Assume there are n environmental sub-features, and the set of environmental sub-features is represented by E = {e1, e2, …, en}. At the same time, construct an adjacency matrix A to represent the connection relationship of the environmental feature network, where A ij represents the connection strength between environmental sub-feature e i and e j . If there is no connection between e i and e j , then A ij = 0. In addition, each environmental sub-feature e i has an initial response sensitivity value s i0 , which is configured in the first step of the model construction strategy. The direct influence score d i of each environmental sub-feature can be determined by its initial response sensitivity value s i0 . The calculation formula is as follows:

[0080] d i = s i0

[0081] The indirect influence score of an environmental sub-feature is propagated by the influence of other environmental sub-features connected to it. For environmental sub-feature e i , the calculation formula for its indirect influence score i i is:

[0082]

[0083] where s j is the comprehensive sensitivity value of environmental sub-feature e j (in the first iteration, s j = s j0 ).

[0084] To balance the direct influence and indirect influence, weights need to be assigned to them respectively. Let the direct influence weight be α and the indirect influence weight be β, and α + β = 1. Usually, these two weights can be adjusted according to the actual situation. For example, α = 0.6 and β = 0.4.

[0085] Taking into account the direct influence score and indirect influence score comprehensively, the calculation formula for the comprehensive sensitivity value s i of environmental sub-feature ei is:

[0086] s i = α × d i+β×i i

[0087] To make the comprehensive sensitivity value more accurate, multiple iterations can be performed for updating. In each iteration, the comprehensive sensitivity value s obtained from the previous iteration is used j to calculate the indirect impact score i of the current iteration i . When the change in the comprehensive sensitivity values of all environmental sub-features between two adjacent iterations is less than a preset threshold ∈, it is considered that the algorithm converges and the iteration stops.

[0088] Step B4: Determine whether any two adjacent environmental sub-features satisfy the transposition condition through a preset transposition extension algorithm. If the transposition condition is satisfied, the positions of the two environmental sub-features in the environmental feature network are swapped until all environmental sub-features no longer satisfy the transposition condition. The model construction strategy is mainly to construct an environmental feature network that can accurately reflect the relationship between environmental features and load samples, so as to provide strong support for subsequent analysis. First, the environmental features are split into a link of several environmental sub-features. Environmental sub-features are the refinement and decomposition of environmental features, and they are the basic components of environmental features. For example, the large environmental feature of environmental temperature can be split into sub-features such as daily average temperature, maximum temperature, and minimum temperature. This can study the impact of environmental factors in different aspects on load samples more carefully. At the same time, obtain the environmental feature response items, which represent the association relationship between environmental features and load samples, and configure corresponding response sensitivity values for each environmental feature response item corresponding to environmental sub-features. This value reflects the sensitivity degree of the environmental sub-feature to the impact of load samples. The larger the value, the more obvious the impact.

[0089] Next, construct an environmental feature network with these environmental sub-features as nodes. This is like building a framework to connect various environmental sub-features. Through this network, the mutual relationship and connection method between environmental sub-features can be intuitively seen, which is convenient for subsequent analysis of their comprehensive impact on load samples.

[0090] Then, calculate the comprehensive sensitivity value of each environmental sub-feature through a preset response sensitivity algorithm. The response sensitivity algorithm is a calculation method that comprehensively considers various factors. It not only considers the attributes of environmental sub-features themselves, but also considers their positions in the environmental feature network, the connection strength with other environmental sub-features, etc. For example, if an environmental sub-feature is closely connected to multiple other important sub-features, its comprehensive sensitivity value may be relatively high. The algorithm obtains a value that can comprehensively reflect the impact degree of environmental sub-features on load samples through a series of calculation rules and weight assignments.

[0091] Finally, the preset transposition extension algorithm is used to determine whether any adjacent environmental sub - features meet the transposition conditions. The transposition extension algorithm is designed based on in - depth analysis of the relationships between environmental sub - features and their impacts on load samples. The transposition conditions usually involve factors such as the correlation between environmental sub - features and the synergy of their impacts on load samples. If the positions of two adjacent environmental sub - features are swapped under certain conditions and can more reasonably reflect their relationships with load samples, or can make the environmental feature network more stable and efficient, then the transposition operation is performed. By continuously making such judgments and transpositions until all environmental sub - features no longer meet the transposition conditions, the environmental feature network reaches a relatively optimal state, which can more accurately reflect the internal connection between environmental features and load samples. This enables the constructed network to quickly search for the target content to complete the judgment.

[0092] The second sample generation module 600 includes a hardware deviation correlation unit 610 and a second sample generation unit 620. The hardware deviation correlation unit 610 calculates the simulation load samples and real - scene load samples with the same classification features to obtain hardware deviation features, and performs clustering analysis on the collected hardware deviation features to obtain each hardware deviation index cluster; the second sample generation unit 620 matches the simulation load samples with the corresponding hardware deviation index clusters to retrieve the corresponding hardware deviation features. The hardware deviation unit is configured with a deviation classification sub - strategy, and the deviation classification sub - strategy includes

[0093] Step A1: Subtract the simulation load waveform and the real - scene load waveform under the same working conditions to obtain a hardware difference waveform;

[0094] Step A2: Classify the hardware difference waveform through a preset feature splitting strategy to obtain the hardware difference features corresponding to each hardware type item;

[0095] Step A3: Repeat Step A1 until all simulation load waveforms are collected;

[0096] Step A4: Mark the hardware intervention degree for the hardware difference features of each hardware type item through a preset deviation calculation algorithm;

[0097] Step A5: Perform mean - value processing on the hardware difference features through the hardware intervention degree marking to obtain the corresponding difference mean features;

[0098] Step A6: Combine different hardware type items to obtain a hardware deviation cluster set;

[0099] Step A7: In the hardware deviation cluster set, perform clustering analysis on the difference mean features through a clustering analysis algorithm to obtain the corresponding hardware deviation index clusters. The design goal of the second sample generation unit 620 is to compensate and correct the simulation load samples through the hardware deviation features to improve the accuracy and generalization ability of the anomaly detection model. The core process of this unit includes four stages: acquisition, classification, calculation, and compensation application of hardware deviation features. The process of obtaining hardware deviation features begins with the analysis of the differences between the simulation load waveform and the actual scene load waveform under the same working conditions. By taking the difference between the two point by point to generate a hardware difference waveform, this difference waveform reflects the inherent deviation between the charging pile hardware device and the simulation model during actual operation. For example, the internal resistance difference or temperature drift of the charging module may cause a systematic shift in the voltage waveform, and such differences will be recorded as hardware difference features.

[0100] Next, the deviation classification sub-strategy disassembles the hardware difference waveform in multiple dimensions: First, decompose the difference waveform into frequency domain features (such as harmonic components), time domain features (such as rising edge slope), and statistical features (such as variance, peak value) through a preset feature splitting strategy. Different hardware type items (such as power supply module, communication module, control unit) will correspond to different feature combination patterns. For example, the hardware deviation of the power supply module may be concentrated in the voltage ripple in the low frequency band, while the deviation of the communication module may be manifested as high frequency noise interference. Subsequently, perform quantitative analysis on the difference features of each hardware type item through a deviation calculation algorithm. This algorithm will combine the physical characteristics and historical operation data of the hardware device to assign a hardware intervention degree mark to each feature. The higher the intervention degree mark value, the greater the impact of the feature on the load data. For example, by calculating the correlation coefficient between a certain feature and the load data and combining the device failure history records, determine the intervention degree weight of the feature.

[0101] After collecting the difference features of all samples, the system will perform mean processing on the difference features of each hardware type item to generate difference mean features. This process eliminates the influence of random noise through methods such as sliding window or exponential smoothing to extract stable hardware deviation patterns. For example, take the average value of the voltage difference waveforms of 100 charging modules under the same working conditions to obtain the typical deviation features of this module. Subsequently, combine the difference mean features of different hardware type items to form a hardware deviation cluster set, and each cluster set represents a specific hardware deviation combination pattern. For example, the low frequency ripple deviation of the power supply module and the delay deviation of the control unit may jointly form a hardware deviation cluster set.

[0102] Finally, classify the hardware deviation clusters through a clustering analysis algorithm (such as DBSCAN or K-means) to generate hardware deviation index clusters. During the clustering process, the Euclidean distance or cosine similarity between features is considered. Similar hardware deviation patterns are grouped into the same cluster, and a unique index identifier is assigned to each cluster. For example, through analysis, it is found that the hardware deviation features of a certain cluster mainly concentrate on the power supply module and the temperature sensor, and the system will mark it as index cluster A. These index clusters provide a key basis for subsequent hardware compensation. When processing the simulation load samples, the second sample generation unit 620 will match the corresponding index cluster according to the real-time collected hardware operation parameters, retrieve the corresponding hardware deviation features to compensate and correct the simulation data, so as to generate hardware compensation load samples closer to the actual scenario. It should be noted that during the calculation process of the hardware deviation features, the clustering model needs to be updated regularly to adapt to the deviation changes brought by newly emerging hardware models or aging devices.

[0103] The second sample generation unit 620 is configured with a sample randomization strategy, which includes generating corresponding random parameter items and corresponding random variable ranges according to the hardware deviation features, generating random variable values for each random parameter item according to the random variable ranges, generating random alternative load waveforms according to the random variable values, substituting the random alternative load waveforms into the simulation load samples to generate the hardware compensation load samples, and correcting the corresponding hardware deviation features.

[0104] Process the corresponding simulation load samples through the hardware deviation features to generate hardware compensation load samples; the core of the sample randomization strategy is to enhance the diversity of the hardware compensation load samples by introducing a randomization mechanism, thereby improving the adaptability of the anomaly detection model to complex hardware deviation scenarios. First, the strategy generates corresponding random parameter items according to the hardware deviation features. These parameter items cover the key performance indicators of the hardware devices, such as the voltage deviation of the power supply module, the current ripple coefficient, and the drift range of the temperature sensor. Each parameter item is directly associated with the feature values in the hardware deviation index cluster. Next, the system determines the random variable ranges for each random parameter item according to the historical data statistical distribution and the physical characteristics of the hardware devices. For example, by analyzing the historical difference waveforms of a certain type of charging module, it is found that its voltage deviation fluctuates between -0.5V and +0.8V. Therefore, the random range of this parameter item is set to [-0.6V, +1.0V] to cover possible extreme cases.

[0105] Subsequently, the system uses methods such as Gaussian mixture model or Latin hypercube sampling to generate random variable values within a preset range, ensuring that the generated values not only conform to the probability distribution characteristics but also have sufficient dispersion. For example, for the temperature drift parameter, the system may increase the generation probability of positive drift values during high-temperature periods in summer based on prior knowledge of seasonal changes, and vice versa in winter. After generating the random variable values, the system maps these values to the corresponding positions of the simulation load samples, and generates a random alternative load waveform by superposition or replacement. For example, a randomly generated voltage deviation value is superimposed on the simulation voltage waveform, and the continuity and physical meaning of the waveform are maintained through Fourier transform.

[0106] Finally, the system corrects the hardware deviation characteristics of the generated random alternative load waveform. By recalculating the frequency domain, time domain, and statistical characteristics of the waveform, it ensures that the compensated samples not only contain the original hardware deviations but also incorporate the dynamic change characteristics brought by randomization. For example, after replacing a certain section of the current waveform, the system reextracts the harmonic components and ripple coefficients of the waveform and updates the corresponding hardware deviation index cluster feature library. It should be noted that during the randomization process, the physical constraint relationships between parameters need to be maintained, such as the correlation between voltage deviation and current deviation, to avoid generating invalid samples that violate Ohm's law. In addition, the system also regularly evaluates the effectiveness of the random strategy through a cross-validation mechanism, and dynamically adjusts the random variable range and generation method according to the model training effect, ensuring that the compensated samples can cover the hardware differences in the real scenario without introducing too much noise interference.

[0107] The training execution module 700 includes inputting a first type of load sample, an environment compensation load sample, and a hardware compensation load sample into an anomaly training model, and inputting a second type of load sample into the first sample generation module 500 or the second sample generation module 600.

[0108] It further includes an adversarial generation module 800. The adversarial generation module 800 is configured with an abnormal waveform database 801. The abnormal waveform library stores abnormal waveform characteristics, and the abnormal waveform characteristics have a matching index set. The matching index set includes several matching sub-characteristics, each of which can be indexed to the corresponding abnormal waveform characteristic, and each abnormal waveform characteristic corresponds to a matching abnormal sub-value. The adversarial generation module 800 is configured with an adversarial generation strategy. The adversarial generation strategy is used to extract the matching sub-characteristics in the third type of load sample and retrieve several corresponding abnormal waveform characteristics, and replace the abnormal waveform characteristics with the load sample through abnormal load constraints to generate an abnormal load sample. The abnormal load constraint is that the alternative abnormal total value of the abnormal load sample falls within a preset alternative threshold range.

[0109] The training execution module 700 brings the abnormal load samples into the abnormal training model. The core function of the adversarial generation module 800 is to enhance the robustness of the anomaly detection model by introducing adversarial samples, and its workflow includes three key links: feature matching, abnormal waveform replacement, and surrogate abnormal total value constraint. First, the system extracts matching sub-features from three types of load samples. These sub-features are key feature points separated from the load waveform through a preset feature extraction algorithm (such as short-time Fourier transform or wavelet decomposition), such as the timestamp of voltage sag, the amplitude of current harmonics, etc. Each matching sub-feature corresponds to one or more matching index sets in the abnormal waveform database 801, and these index sets are feature template libraries pre-constructed according to historical fault cases and equipment operation characteristics.

[0110] Next, the system retrieves the corresponding abnormal waveform features from the abnormal waveform library according to the matching sub-features. For example, when it is detected that the voltage fluctuation frequency of a certain section of the load waveform matches the contactor fault feature of a certain type of charging pile in the historical database, the voltage distortion waveform corresponding to this fault will be extracted. At this time, the adversarial generation strategy needs to replace these abnormal waveform features into the original load sample through abnormal load constraint to generate adversarial abnormal load samples. The replacement process uses waveform interpolation or feature superposition technology. For example, on the premise of maintaining the overall trend of the original waveform, the voltage waveform in a specific time period is replaced with an abnormal waveform containing high-frequency oscillation.

[0111] The calculation of the surrogate abnormal total value is the core of this process. It is obtained by accumulating the matching abnormal sub-values corresponding to all replaced abnormal waveform features. Each abnormal waveform feature has been assigned a matching abnormal sub-value when stored in the database, and this value reflects the contribution weight of this abnormal feature to the abnormal degree of the load data. For example, the voltage distortion corresponding to the contactor fault may be assigned a higher abnormal sub-value (such as 0.8), while the deviation corresponding to the temperature sensor drift may be assigned a lower value (such as 0.3). The calculation formula for the surrogate abnormal total value is: where p is the total number of matching abnormal sub-values with a matching relationship. The replacement weight is a coefficient dynamically adjusted according to the coverage range of the abnormal waveform feature in the sample. For example, when a certain section of the abnormal waveform replaces 20% of the entire sample, its replacement weight is 0.2. The system calculates the surrogate abnormal total value in real time and ensures that it falls within the preset surrogate threshold range (such as 0.5 - 1.5). If it exceeds the range, it will automatically adjust the number of replaced abnormal waveforms or the feature combination, such as reducing the replacement ratio of high-abnormal-sub-value features or increasing the coverage range of low-abnormal-sub-value features to maintain the controllability of the overall abnormal degree.

[0112] It should be noted that the calculation of the alternative abnormal total value needs to meet the dual constraints of time and amplitude simultaneously. For example, some abnormal waveform features have high requirements for temporal continuity, and waveform mutations that are not coherent should be avoided during replacement. In addition, the system will continuously optimize the feature annotation and matching strategy of the abnormal waveform library through an adversarial training mechanism. For example, according to the misjudgment of the generated samples by the model during the training process, the weight allocation for matching abnormal sub-values will be dynamically adjusted to improve the effectiveness of the adversarial generated samples.

[0113] It also includes a data acquisition subsystem 900, which is used to collect the load data of each charging pile to generate historical load samples; the historical data management module 100 is configured with a data re-acquisition unit 102 and a data correction unit 103. The data correction unit 103 is used to extract the discreditable features in the historical load samples determined to be type-two load samples. The data re-acquisition unit 102 generates corresponding re-acquisition instructions according to the discreditable features and sends the re-acquisition instructions to the data acquisition subsystem. And corresponding re-acquisition instructions are generated according to the discreditable features and sent to the data acquisition subsystem 900. The acquisition subsystem is configured with a virtual inspection load, which is used to access the charging pile and execute the corresponding re-acquisition instructions. The re-acquisition instructions include load parameters, and the load parameters are used to configure the corresponding virtual inspection load. The core of the data acquisition subsystem 900 is to replace manual work, simulate charging by means of an inspection intelligent vehicle, collect the load during this period remotely, obtain samples with higher reliability through on-site collection, and at the same time, abnormal judgment can be made on the local charging pile.

[0114] Of course, the above are only typical examples of the present invention. In addition, the present invention can also have many other specific implementation manners. Any technical solutions formed by equivalent replacement or equivalent transformation fall within the scope of protection required by the present invention.

Claims

1. A load resource data anomaly detection system based on a new power system, characterized in that: It includes a historical data management module, a real-scene management module, and a simulation generation module. The historical data management module is configured with a historical data load information library. The historical data management module is used to mark the historical load data in the historical data load information library to generate historical load samples; The real-scene management module is used to collect the load data of the target charging pile in different working states through the real-scene load monitoring interface to generate real-scene load samples; The simulation generation module is used to construct a simulation discharge model, and obtain corresponding simulation load samples by substituting the charging pile simulation parameters and load simulation parameters into the simulation discharge model and running; The sample screening module is configured with a sample confidence algorithm to calculate the confidence of the load samples, and divide the load samples into first-class load samples, second-class load samples, and third-class load samples according to the confidence; The first sample generation module includes an environment deviation correlation unit and a first sample generation unit. The environment deviation correlation unit generates the environmental characteristics of the historical load samples through a preset environment analysis strategy, calculates the load deviation between the real-scene load samples and the historical load samples with the same classification characteristics, associates the load deviation with the environmental characteristics to generate an environment correlation sub-model. The first sample generation unit retrieves the load deviation from the environment correlation sub-model according to the input real-scene simulation environmental characteristics, and processes the real-scene load samples according to the load deviation to generate environment-compensated load samples; The second sample generation module includes a hardware deviation correlation unit and a second sample generation unit. The hardware deviation correlation unit calculates the hardware deviation characteristics of the simulation load samples and the real-scene load samples with the same classification characteristics, and performs clustering analysis on the collected hardware deviation characteristics to obtain each hardware deviation index cluster; The second sample generation unit retrieves the corresponding hardware deviation characteristics by matching the simulation load samples with the corresponding hardware deviation index clusters, and processes the corresponding simulation load samples through the hardware deviation characteristics to generate hardware-compensated load samples; The training execution module includes inputting the first-class load samples, environment-compensated load samples, and hardware-compensated load samples into the anomaly training model, and inputting the second-class load samples into the first sample generation module or the second sample generation module.

2. The load resource data anomaly detection system based on a new power system according to claim 1, characterized in that: It also includes an adversarial generation module. The adversarial generation module is configured with an abnormal waveform database. The abnormal waveform library stores abnormal waveform characteristics. The abnormal waveform characteristics have a matching index set. The matching index set includes several matching sub-characteristics. Each matching sub-characteristic can be indexed to the corresponding abnormal waveform characteristic. Each abnormal waveform characteristic corresponds to a matching abnormal sub-value. The adversarial generation module is configured with an adversarial generation strategy. The adversarial generation strategy is used to extract the matching sub-characteristics in the third-class load samples and retrieve several corresponding abnormal waveform characteristics, and replace the abnormal waveform characteristics into the load samples through abnormal load constraints to generate abnormal load samples. The abnormal load constraint is that the alternative abnormal total value of the abnormal load samples falls within a preset alternative threshold range; The training execution module brings the abnormal load samples into the abnormal training model.

3. The load resource data anomaly detection system based on a new power system according to claim 1, wherein: It further includes a data acquisition subsystem for acquiring the load data of each charging pile to generate historical load samples; the historical data management module is configured with a data re-acquisition unit and a data correction unit. The data correction unit is used to extract the credit loss features in the historical load samples judged as class II load samples. The data re-acquisition unit generates corresponding re-acquisition instructions according to the credit loss features and sends the re-acquisition instructions to the data acquisition subsystem.

4. The load resource data anomaly detection system based on a new power system according to claim 3, characterized in that: The acquisition subsystem is configured with a virtual inspection load for accessing the charging pile and executing the corresponding re-acquisition instructions. The re-acquisition instructions include load parameters for configuring the corresponding virtual inspection load.

5. The abnormal detection system for load resource data based on a new power system according to claim 1, wherein: The actual scene management module is configured with a function mirroring unit. The function mirroring unit generates a corresponding function classification judgment model by controlling the target charging pile to work in different function states, analyzes the historical load samples through the function classification judgment model to generate a corresponding charging pile working sequence, and controls the working state of the target charging pile according to the charging pile working sequence to generate actual scene load samples corresponding to the historical load samples.

6. The load resource data anomaly detection system based on a new power system according to claim 1, wherein: The environmental analysis strategy includes Step S1: Obtain the location information of the charging pile corresponding to the historical load sample; Step S2: Retrieve the charging pile numbers with location relevance from the pre-constructed charging pile distribution topology model according to the location information; Step S3: Obtain the corresponding relevant environmental information according to the charging pile numbers; Step S4: Identify the reliable environmental features in the relevant environmental information through a preset predation deduction model to generate the environmental feature data.

7. The load resource data anomaly detection system based on a new power system according to claim 1, wherein: The hardware deviation unit is configured with a deviation classification sub-strategy, and the deviation classification sub-strategy includes Step A1: Subtract the simulated load waveform and the actual scene load waveform under the same working conditions to obtain a hardware difference waveform; Step A2: Classify the hardware difference waveform through a preset feature splitting strategy to obtain the hardware difference features corresponding to each hardware type item; Step A3: Repeat Step A1 until all the simulated load waveforms are collected; Step A4: Mark the hardware intervention degree for the hardware difference features of each hardware type item through a preset deviation calculation algorithm; Step A5: Perform an averaging process on the hardware difference features through the hardware intervention degree marking to obtain the corresponding difference mean features; Step A6: Combine different hardware type items to obtain a hardware deviation cluster set; Step A7: In the hardware deviation cluster set, perform cluster analysis on the difference mean features through a cluster analysis algorithm to obtain the corresponding hardware deviation index cluster.

8. The load resource data anomaly detection system based on the new power system according to claim 6, wherein: The first sample generation module includes a configured model construction strategy, and the model construction strategy includes Step B1: Split the environmental features into links of several environmental sub-features, obtain environmental feature response items, and configure corresponding response sensitivity values for each environmental feature response item corresponding to the environmental sub-features; Step B2: Construct an environmental feature network with the environmental sub-features as nodes; Step B3: Calculate the comprehensive sensitivity value of each environmental sub-feature through a preset response sensitivity algorithm; Step B4: Determine whether any two adjacent environmental sub-features meet the transposition condition through a preset transposition and extension algorithm. If the transposition condition is met, swap the positions of the two environmental sub-features in the environmental feature network until all environmental sub-features no longer meet the transposition condition.

9. The abnormal detection system for load resource data based on a new power system according to claim 1, wherein: The second sample generation unit is configured with a sample random strategy, which includes generating corresponding random parameter items and corresponding random variable ranges according to the hardware deviation features, generating random variable values for each random parameter item according to the random variable ranges, generating random alternative load waveforms according to the random variable values, substituting the random alternative load waveforms into the simulation load samples to generate the hardware compensation load samples, and correcting the corresponding hardware deviation features.

Images