Generator set load test method based on multi-source data fusion and deep learning

Abstract

The application provides a generator set load test method based on multi-source data fusion and deep learning, relates to the technical field of computer systems based on specific calculation models, and comprises the following steps: collecting generator set load test data, forming standardized time sequence load test data, determining a load deviation judgment result, forming a correction benchmark for current test action analysis when the load deviation exists, performing pre-correction, locking the current dynamic stage in a manner that combines rule judgment and deep learning stage recognition, extracting stage characteristics according to the current dynamic stage, generating a current state result based on the stage characteristics, generating a test promotion decision according to the load deviation judgment result and the current state result, and performing correction and correction verification according to the test promotion decision. The application solves the problems of analysis benchmark deviation caused by load characteristic changes in the continuous load test process and inaccurate test stage determination caused by mixed multi-time scale electrical parameters.

Description

Technical Field

[0001] This invention relates to the field of computer system technology based on specific computational models, and in particular to a generator set load testing method based on multi-source data fusion and deep learning. Background Technology

[0002] With the widespread application of generator sets in power systems, accurate assessment of their performance and load-carrying capacity is of great significance for safe operation and maintenance. Load-carrying tests of generator sets are typically conducted during factory acceptance, on-site acceptance, and operation and maintenance inspections. By continuously collecting and analyzing key parameters such as voltage, current, active power, and frequency under different load levels, the load-carrying capacity, dynamic response, and steady-state performance of the generator set are verified.

[0003] In practical applications, high and low voltage load cells are often used to simulate actual load conditions, enabling continuous load testing of generator sets at different load levels. The high and low voltage load cells apply load through internal resistors or inductors and controllable circuit switching, covering different operating conditions from low load to full load. During testing, the system needs to collect various electrical parameters in real time and analyze the test results to evaluate the unit's operating status at each load level.

[0004] During continuous load testing of generator sets using high- and low-voltage load cells, loading, holding, and switching operations are typically performed sequentially according to a preset load level. Electrical parameters such as voltage, current, active power, and frequency are collected at each test stage to evaluate the operating status under the current test action. However, in actual testing, the load characteristics of the load cell are not ideally constant. Factors such as changes in the temperature rise of its internal resistance units, changes in circuit connection status, and switching between high-voltage and low-voltage operating conditions can cause the actual load conditions corresponding to the same test action to change at different time periods. This causes the analysis benchmark established based on the preset target load level to gradually deviate from the actual load condition, affecting the accuracy of subsequent test judgments.

[0005] On the other hand, during the response process of a generator set after load switching, electrical parameters such as voltage and current typically change rapidly, while electrical parameters such as active power and frequency change relatively slowly. In existing technologies, if multiple parameter data within the same time window are analyzed uniformly without distinguishing the differences in their change processes, data from different change stages can easily overlap, making it difficult to accurately determine the operating status of the current testing phase. Summary of the Invention

[0006] In view of this, the embodiments of the present invention provide a generator set load testing method based on multi-source data fusion and deep learning, which solves the problems of deviation of analysis benchmark due to load characteristic changes and inaccurate judgment during the test stage caused by the mixing of electrical parameters at multiple time scales during continuous load testing.

[0007] This invention provides a generator set load testing method based on multi-source data fusion and deep learning, the method comprising: Collect generator set load test data under the current test action, and perform time alignment and outlier removal to form standardized time-series load test data.

[0008] Based on standardized time-series load test data, the load offset judgment result is determined, and when the load offset exists, a correction benchmark is formed for the analysis of the current test action, and the subsequent test actions are pre-corrected.

[0009] Based on standardized time-series load test data, a combination of rule-based judgment and deep learning stage identification is used to pinpoint the current dynamic stage.

[0010] Extract stage features based on the current dynamic stage, and generate the current state result based on the stage features.

[0011] Based on the load offset judgment result and the current state result, a test advancement decision is generated, and correction and correction verification are performed according to the test advancement decision.

[0012] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. This invention provides a generator load testing method based on multi-source data fusion and deep learning. It addresses the issue of drift in the actual load conditions of the same test action during continuous load testing due to changes in temperature rise, load connection status, and high / low voltage switching. The method connects standardized time-series load test data, load offset judgment, dynamic stage locking, stage feature extraction, current state result generation, test advancement decision-making, and correction verification into a single test closed loop. This allows the test to promptly correct the analysis basis when the current action deviates from the original target conditions, and to continue testing when the advancement conditions are met, and to perform observation or pause when the advancement conditions are not met. This reduces the possibility of deviations from previous test actions being transmitted to and amplified in subsequent test actions during continuous load testing, and improves the comparability between actions in the entire test cycle and the reliability of the final test results.

[0013] 2. This invention, by establishing a correction benchmark for analyzing the current test action when load offset exists, addresses the issue that the target interval corresponding to the original target load level no longer accurately reflects the actual load level of the current action. It regenerates a correction power interval and a correction current interval using the current average active power and current obtained from the current load analysis data segment. This correction interval replaces the original target interval in subsequent stage locking, stage feature extraction, and state determination for the current test action, preventing subsequent judgments from mistakenly identifying actual load level changes as unit malfunctions. After this processing, even if the current action deviates from the preset target level due to load box temperature rise, analysis can still be performed around the actual load center of the current action, reducing stage and state misjudgments caused by inaccurate comparison benchmarks, and ensuring that the internal analysis results of the current action are consistent with the actual load state.

[0014] 3. This invention uses a combination of rule-based judgment and deep learning-based stage identification based on standardized time-series load test data to pinpoint the current dynamic stage. It addresses the problem of overlapping rapid changes in voltage and current with relatively slow changes in active power and frequency within the same timeframe, which makes it difficult for fixed-window analysis to distinguish between the impact, recovery, and steady-state stages. First, it calculates the fast variable deviation sequence and the slow variable deviation sequence based on the current analysis benchmark. Then, it uses whether the deviation value within a preset sampling window exceeds a stability threshold and the relationship between the mean deviation values ​​before and after the window to form a rule-based stage identification result. Simultaneously, it introduces a deep learning stage identification model trained on historical load samples for parallel identification. The current dynamic stage is determined only when the two identification results are consistent; otherwise, it is re-identified after extended observation and resampling. This accurately separates the processes where fast variables have not yet converged and slow variables are still recovering. It also avoids one-sided errors caused by relying solely on threshold rules or model output, ensuring that the subsequently extracted stage features correspond one-to-one with the actual response process, thus enhancing the stability and repeatability of the stage identification results.

[0015] 4. This invention generates test progression decisions based on load offset judgment results and current state results. It addresses the problems in existing load testing where decisions on whether to continue loading, extend observation, or pause progression rely primarily on manual experience, and where different operators have inconsistent processing rhythms under the same test state. This invention transforms the existence of load offset in the current action and whether the current dynamic stage meets requirements into executable progression categories, which are then used to trigger subsequent processing such as continuing progression, slowing down progression, maintaining observation, or pausing progression. When the current action has already experienced load offset and the state still does not meet requirements, the test will not continue at the original pace, carrying the deviation into the next action. When the current action has offset but the stage state has met requirements, the load change intensity of subsequent actions can be reduced by slowing down progression. When the current action has no load offset but the current stage has not yet met requirements, supplementary sampling data can be obtained by maintaining observation. When the current action has neither load offset nor meets requirements, the next action can proceed. This establishes a clear correspondence between the test progression rhythm and the actual operating state of the current action, reducing test distortion and decreased test efficiency caused by premature progression or excessive waiting. Attached Figure Description

[0016] Figure 1 This is a flowchart of a generator set load testing method based on multi-source data fusion and deep learning provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the dynamic response curve of a generator set under sudden load changes; Figure 3 This is a schematic diagram of the training accuracy curve of the recognition model in the deep learning stage; Figure 4 This is a schematic diagram of the training loss curve of the recognition model in the deep learning stage. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on the invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0018] In the following description, some embodiments are referred to, which describe a subset of all possible embodiments. However, it is understood that some embodiments may be the same subset or different subset of all possible embodiments and may be combined with each other without conflict.

[0019] Embodiment 1 of the present invention: This embodiment provides a generator set load testing method based on multi-source data fusion and deep learning. For example... Figure 1The flowchart shown is for a generator load testing method based on multi-source data fusion and deep learning. The processing flow of this method can include the following steps: collecting generator load test data under the current test action, performing time alignment and outlier removal to form standardized time-series load test data, determining the load offset judgment result, and when the load offset exists, forming a correction benchmark for the analysis of the current test action and pre-correcting subsequent test actions; locking the current dynamic stage based on the standardized time-series load test data using a combination of rule judgment and deep learning stage recognition; extracting stage features based on the current dynamic stage; generating the current state result based on the stage features; generating a test advancement decision based on the load offset judgment result and the current state result; and performing correction and correction verification based on the test advancement decision.

[0020] In this embodiment, the load test of the generator set is executed sequentially according to a pre-set test action table. The test action table records at least several sequentially arranged test actions, each corresponding to a target load level, target load state, planned loading step size, and planned holding time. At the start of the test, the generator set is first put into an initial operating state, and a test action table corresponding to the current test task is established; then, each test action is executed sequentially from front to back according to the test action table. When executing each test action, the load cell is first controlled to complete the corresponding load connection or switching according to the planned loading step size corresponding to the test action, so that the generator set enters the target load level corresponding to the test action; after the load is connected, the current load level is maintained for a preset holding time, and during this period, test data such as voltage, current, active power, frequency, load state, and cell temperature are continuously collected; then, based on the collected data, the current test action is judged for load offset, dynamic stage locking, stage feature extraction, current state result generation, and test progress decision judgment. When the current test action is determined to meet the advancement conditions, the next test action in the test action list is initiated; when the current test action is determined not to meet the advancement conditions, the current test action is continuously observed, the next test action is pre-corrected, or the execution of subsequent test actions is suspended. Through the above method, the entire load test is carried out step by step in a closed-loop process of action execution, data acquisition, status analysis, advancement decision-making, and execution of the next action.

[0021] Under the above testing process, the following details the data processing, status judgment, and test progress control process for a single current test action.

[0022] In this embodiment, for the multi-source electrical parameter data generated during the continuous load testing of the generator set, it is necessary to first unify the processing of various test data to ensure that data from different sources and sampling frequencies are comparable under the same time reference, and to eliminate the impact of missing data and abnormal fluctuations on subsequent analysis. Therefore, during the execution of the current test action, the collected test data of voltage, current, active power, frequency, load status, and enclosure temperature are first time-aligned, and missing values ​​and outliers in the data are corrected. Finally, standardized time-series load test data corresponding to the current test action is formed, which serves as the basis for subsequent load status analysis and test process adjustment.

[0023] Specifically, within each execution cycle corresponding to the current test action, the following processing is performed: The start time of the current test action is determined before it begins. In different embodiments, the start time can be determined by the time the control system issues the current test action command, the time when the load state changes, or the earlier of these two times and is recorded by the system. In this embodiment, the time the control system issues the current test action command is used as the start time, and the data processing starting point for this current test action is established around this starting point. This sampling time axis is used to uniformly mark the data positions of each sampling moment after the start of the current test action, so that subsequently acquired voltage, current, active power, frequency, load state, and enclosure temperature can all be mapped to the same set of consecutive sampling moments.

[0024] After establishing the sampling time axis, data for each test item corresponding to the current test action is collected according to this sampling time axis. These test items include voltage, current, active power, frequency, load status, and enclosure temperature. Voltage, current, active power, and frequency reflect the electrical response of the generator set under the current test action; load status reflects the connection status and action execution of the high and low voltage load boxes; and enclosure temperature reflects the thermal state changes of the load box during the current test action. During data collection, each test item forms a set of raw time-series data, and each raw data entry includes at least the corresponding sampling time. Since the sampling sources for different test items may differ—for example, voltage and current from the electrical parameter acquisition unit, load status from the control execution unit, and enclosure temperature from the temperature acquisition unit—the sampling times of these raw time-series data are usually not entirely consistent. Therefore, the raw time-series data corresponding to each test item needs to be resampled using a preset sampling period. The preset sampling period can be pre-set according to the control accuracy and data processing accuracy of the test system, for example, set to a fixed millisecond or second-level sampling interval.

[0025] During resampling, a set of consecutive target sampling times is first generated according to the action start time and preset sampling period. Then, the original time series data of each test item is mapped to this set of target sampling times. If an original sample value exists at a certain target sampling time, that original sample value is directly used as the data item corresponding to that target sampling time. If an original sample value does not exist at a certain target sampling time, the sample value with the closest time distance to that target sampling time is selected from the original sample values ​​adjacent to that target sampling time. In this way, the data of each test item are arranged on the same set of target sampling times to obtain aligned time series data.

[0026] After obtaining the aligned timing data, it is necessary to further check whether there are missing values ​​in the data items corresponding to each sampling time. A missing value refers to a test item that did not generate valid data at a certain sampling time, or the generated data was marked as invalid by the system. When determining missing values, each test item data at each sampling time is traversed one by one. If the data item is empty or does not meet the preset data format requirements, then the data item is determined to be a missing value. For missing values, the system searches for the nearest previous valid data item for the same test item before the current sampling time, and then searches for the nearest next valid data item for the same test item after the current sampling time. If both valid data items exist, interpolation is performed on the two valid data items based on the time difference between the current sampling time and the sampling time of the previous valid data item, and the time difference between the current sampling time and the sampling time of the next valid data item, to obtain the completed value for the current sampling time. If only the previous valid data item exists, its value is copied as the completed value for the current sampling time. If only the next valid data item exists, its value is copied as the completed value for the current sampling time. If neither of the adjacent valid data items exists, the data item corresponding to that sampling time is left empty, and that sampling time is marked as an unavailable sampling point in subsequent calculations.

[0027] After completing the missing value completion, outliers in the aligned time series data are identified and processed. Outliers refer to abrupt changes in data values ​​between adjacent sampling times that do not conform to the normal variation pattern of the test item. When identifying outliers, based on the completed aligned time series data, adjacent difference calculations are performed for each test item and each data item corresponding to each sampling time. Specifically, for the data item corresponding to the current sampling time, the values ​​of the test item at the previous and next sampling times are read. The absolute values ​​of the difference between the current sample value and the previous sample value are calculated, and the absolute values ​​of the difference between the next sample value and the current sample value are calculated. The larger of the two values ​​is taken as the data jump variable corresponding to the current sampling time. Then, this data jump variable is compared with the pre-set jump variable threshold for the current test item. The above jump variable thresholds are set separately for different test items: voltage corresponds to a voltage jump variable threshold, current corresponds to a current jump variable threshold, and active power, frequency, load state, and enclosure temperature each correspond to their respective jump variable thresholds. When the current data jump variable is greater than the preset jump variable threshold corresponding to the current test item, it indicates that the data change amplitude at the current sampling time exceeds the normal change range allowed for the test item. The data item corresponding to that sampling time is then marked as an anomaly, and the anomaly is replaced to prevent it from directly affecting subsequent load offset judgment and dynamic stage locking. During replacement, the preceding and following valid data items are read first. If both valid data items exist, the values ​​of the preceding and following valid data items are averaged to obtain an intermediate replacement value, which replaces the original value of the current anomaly. If only the preceding valid data item exists, its value is used to replace the current anomaly. If only the following valid data item exists, its value is used to replace the current anomaly. If only the following valid data item exists, its value is used to replace the current anomaly. If neither of the preceding nor following valid data items exists, the anomaly remains empty, and the data item corresponding to that sampling time is marked as an abnormal null value that cannot participate in subsequent calculations.

[0028] In this embodiment, for discrete test item load status, if an anomaly occurs, the effective load status value of the time adjacent to the current anomaly is read; if the load status before and after is consistent, the consistent status is directly used to replace the current anomaly; if the load status before and after is inconsistent, the current anomaly is kept in an unconfirmed state, and the sampling point is not used as the basis for status judgment in subsequent processing.

[0029] After completing missing value completion and outlier replacement, the data from each test item at the same sampling time need to be recombined in chronological order. Specifically, at each sampling time, the voltage, current, active power, frequency, load status, and enclosure temperature data corresponding to that sampling time are read and written into the same time-series record according to a unified field order. Then, all time-series records are arranged in ascending order of sampling time to form standardized time-series load test data corresponding to the current test action. In the data formed in this way, each sampling time corresponds to a complete record, and each record contains the data of each test item at the same time, thus providing a consistent data foundation for subsequent load offset judgment, correction benchmark formation, dynamic stage locking, stage feature extraction, and state result determination based on a unified time reference.

[0030] In this embodiment, after forming the standardized time-series load test data corresponding to the current test action, it is further necessary to determine whether the actual load under the current test action is still consistent with the preset target load level. If it is found that the actual load has deviated from the target load level, a new comparison benchmark applicable to the subsequent analysis of the current test action is re-established. At the same time, the execution parameters of the subsequent test actions are adjusted in advance to reduce the probability of distortion of the subsequent analysis benchmark due to changes in the thermal state of the load box, changes in the load access state, or deviation of the actual load level from the original target during continuous load testing. This ensures that the data on which the subsequent dynamic stage locking, stage feature extraction, and state determination are based have a consistent comparison basis.

[0031] Specifically, the system first reads the target load level, target load status, and action start time corresponding to the current test action, set by the tester before the test begins. The target load level characterizes the preset load level corresponding to the current test action in the test action table, the target load status characterizes the access state that the load bank should achieve under the current test action, and the action start time defines the starting point for data analysis of the current test action. After reading the above information, data items within a preset analysis duration after the action start time are extracted from the standardized time-series load test data to form the current load analysis data segment.

[0032] After obtaining the current load analysis data segment, the current data corresponding to each sampling moment within the current load analysis data segment is read sequentially. These current data are accumulated item by item to obtain the total current. Then, the total current is divided by the number of current sampling points within the current load analysis data segment to obtain the current average current. In the same way, the active power data corresponding to each sampling moment within the current load analysis data segment is read sequentially, and the current average active power is calculated. In this way, the instantaneous fluctuations of individual sampling points are smoothed out, so that subsequent judgments are based on the overall load level within the preset observation period.

[0033] Based on the target load level corresponding to the current test action, the target active power range and target current range corresponding to the current test action are read from the pre-established level mapping relationship. The target active power range is used to limit the allowable range of active power values ​​under the target load level, and the target current range is used to limit the allowable range of current values ​​under the target load level.

[0034] When the current average active power is within the target active power range and the current average current is within the target current range, the load strength matching result is recorded as in-range matching; when either of them does not fall within the corresponding range, the load strength matching result is recorded as out-of-range deviation. Thus, it is possible to determine from the overall load level of the current action whether the actual load strength under the current action is still consistent with the expected strength corresponding to the target load level.

[0035] First, extract the chamber temperature corresponding to the start time of the action from the standardized time-series load test data as the initial chamber temperature; then extract the chamber temperature corresponding to the end of the current load analysis data segment as the current chamber temperature. Subtract the initial chamber temperature from the current chamber temperature to obtain the temperature rise increment. If the temperature rise increment is greater than the preset temperature rise threshold, it indicates that the chamber temperature rise during the current action execution has reached a level that may affect the actual load-bearing characteristics of the load chamber, and the temperature rise impact result is recorded as being affected by continuous temperature rise; if the temperature rise increment is not greater than the temperature rise threshold, the temperature rise impact result is recorded as not being affected by continuous temperature rise, thus incorporating the load offset factor caused by heat accumulation during continuous testing into the judgment.

[0036] Load status data is extracted from the current load analysis data segment. The load status values ​​within the current load analysis data segment are read according to the sampling time and compared item by item with the target load status corresponding to the current test action. If the load status remains consistent with the target load status at more than a preset number of sampling times within the preset analysis time, the load status consistency result is recorded as consistent; if the load status deviates from the target load status at more than a preset number of sampling times, the load status consistency result is recorded as inconsistent.

[0037] When the load strength matching result is within the range, the temperature rise effect result is not affected by continuous temperature rise, and the load state consistency result is consistent, it indicates that the load strength, thermal state, and load connection state under the current test action have not deviated from the expected requirements. In this case, the load offset judgment result is determined to be no load offset. If any of the following is true: the load strength matching result is outside the range, the temperature rise effect result is affected by continuous temperature rise, or the load state consistency result is inconsistent, it indicates that the actual load situation corresponding to the current action has deviated from the original target conditions. In this case, the load offset judgment result is determined to be that load offset exists.

[0038] When the load offset judgment result is determined to be that there is a load offset, the current average active power and current average current obtained from the aforementioned statistics are read. Since the current average active power and current average current represent the actual load center level after the current action is actually executed, these two average values ​​are used as the correction power reference and correction current reference, respectively.

[0039] Using the corrected power reference as the center, subtracting the power reference deviation length from the corrected power reference yields the lower limit of the corrected power range. Adding the power reference deviation length to the corrected power reference yields the upper limit of the corrected power range, thus obtaining the corrected power range. Similarly, using the corrected current reference as the center, subtracting the current reference deviation length from the corrected current reference yields the lower limit of the corrected current range. Adding the current reference deviation length to the corrected current reference yields the upper limit of the corrected current range, thus obtaining the corrected current range. Then, the corrected power range, the corrected current range, and the load state corresponding to the current test action are combined as the corrected reference for subsequent analysis of the current test action.

[0040] After the modified benchmark is established, in the subsequent dynamic phase locking, phase feature extraction and state determination process for the current test action, the target range corresponding to the original target load level of the current test action will no longer be used directly. Instead, the modified benchmark will replace the original target range as the new comparison basis.

[0041] It should be noted that when the actual load has deviated from the original target, if the original target range is still used to judge the dynamic stage and state characteristics of the current action, it is easy to misjudge the actual load level formed by the current action as a continuous abnormality. However, after using the corrected benchmark, the subsequent analysis can be carried out around the actual load state formed by the current action, so that the analysis benchmark corresponds to the actual operating state again.

[0042] After establishing a corrected baseline for analyzing the current test action, subsequent test actions are pre-corrected to prevent the offset state of the current action from being propagated to the next action and further amplified.

[0043] The next test action following the current test action is read from the test action table, and the planned loading step size and planned hold time corresponding to the next test action are extracted. If the current average active power is greater than the upper limit of the target active power range, it means that the actual load intensity of the current action has exceeded the original planned requirements. If the next action continues to proceed according to the original planned step size, it may further aggravate the deviation. Therefore, the planned loading step size corresponding to the next test action is reduced and the planned hold time is extended to slow down the test progress and increase the observation time. Specifically, the first average active power deviation value is obtained by subtracting the upper limit of the target active power range from the current average active power. Using the first average active power deviation value as the query key, the planned loading step size reduction value and the planned hold time extension value are retrieved from the pre-set pre-correction mapping table in the database. The adjusted planned loading step size is obtained by subtracting the planned loading step size reduction value from the planned loading step size, and the adjusted planned hold time is obtained by adding the planned hold time extension value to the planned hold time.

[0044] If the current average active power is less than the lower limit of the target active power range, it means that the actual load intensity of the current action is lower than the original plan requirement. At this time, the planned loading step size corresponding to the next test action remains unchanged, and the planned hold time is extended. This allows for more sufficient hold time after the next action is executed to observe whether it has re-entered a reasonable load level. Specifically, the lower limit of the target active power range is subtracted from the current average active power to obtain the second average active power deviation value. The second average active power deviation value is used as the query key value to retrieve the planned hold time extension value from the pre-set pre-correction mapping table in the database. The planned hold time extension value is added to the planned hold time to obtain the adjusted planned hold time.

[0045] It should be noted that the planned loading step size reduction and planned hold time extension values ​​extracted from the pre-correction mapping table are numerical data without positive or negative values.

[0046] If the current average active power is within the target active power range, it means that the actual load level of the current action has not deviated beyond expectations from the perspective of active power. At this time, the planned loading step size and planned holding time corresponding to the next test action remain unchanged.

[0047] After completing the above adjustments, the adjusted planned loading step size and planned hold time are written into the test action table as a pre-correction result for subsequent test actions. Through this process, the offset judgment of the current action not only affects the analysis benchmark correction of the current action, but also affects the adjustment of the propulsion parameters of subsequent actions, thereby forming advance control for the current offset state during continuous testing.

[0048] If the load offset judgment result for the current test action indicates that a load offset exists, then the correction benchmark corresponding to the current test action is read as the current analysis benchmark; if the load offset judgment result for the current test action indicates that there is no load offset, then the target range corresponding to the target load level of the current test action is read as the current analysis benchmark. The current analysis benchmark includes at least the current benchmark voltage range, the current benchmark current range, the current benchmark active power range, and the current benchmark frequency range.

[0049] After establishing the current analysis baseline, voltage, current, active power, and frequency data corresponding to the current test action are extracted from standardized time-series load test data. Following a preset sampling period and window length, data items corresponding to adjacent sampling moments are continuously extracted from the start of the action to form the current dynamic response sequence. Each sampling moment in the current dynamic response sequence corresponds to a set of voltage, current, active power, and frequency data. This dynamic response sequence serves as the basis for subsequent deviation calculations and stage judgments.

[0050] like Figure 2 As shown in the diagram, the dynamic response curve of the generator set under sudden load changes illustrates that during the load switching process corresponding to the current test action, the system frequency and terminal voltage of the generator set will experience transient fluctuations over time, gradually recovering to a stable state after the impact response. Based on the voltage, current, active power, and frequency data continuously collected after the start of the action, a current dynamic response sequence can be formed. By calculating the deviation of the data corresponding to each sampling time in this dynamic response sequence relative to the current analysis benchmark, a set of fast variable deviation value pairs and a set of slow variable deviation value pairs are further obtained, forming a fast variable deviation sequence and a slow variable deviation sequence, which serve as the basis for subsequent rule-based stage identification and model-based stage identification.

[0051] Read the voltage data corresponding to a specific sampling moment, and then read the lower and upper limits of the current reference voltage interval. When the voltage data is less than the lower limit of the current reference voltage interval, subtract the voltage data from the lower limit to obtain the voltage deviation value at that sampling moment. When the voltage data is greater than the upper limit of the current reference voltage interval, subtract the upper limit to obtain the voltage deviation value at that sampling moment. When the voltage data is between the lower and upper limits of the current reference voltage interval, record the voltage deviation value at that sampling moment as zero. Then divide the voltage deviation value by the width of the current reference voltage interval to obtain the standardized voltage deviation value at that sampling moment.

[0052] Read the current data corresponding to a specific sampling moment, and then read the lower and upper limits of the current reference current interval. When the current data is less than the lower limit of the current reference current interval, subtract the current data from the lower limit to obtain the current deviation value at that sampling moment; when the current data is greater than the upper limit of the current reference current interval, subtract the upper limit to obtain the current deviation value at that sampling moment; when the current data is between the lower and upper limits of the current reference current interval, record the current deviation value at that sampling moment as zero. Then divide the current deviation value by the width of the current reference current interval to obtain the current standardized deviation value at that sampling moment. For the same sampling moment, the larger of the voltage standardized deviation value and the current standardized deviation value is taken as the fast variable deviation value at that sampling moment.

[0053] Similar to the fast variable calculation, slow variable deviations are calculated for active power and frequency. Active power data corresponding to a specific sampling time is read, followed by the lower and upper limits of the current reference active power interval. When the active power data is less than the lower limit of the current reference active power interval, the active power data is subtracted from the lower limit to obtain the power deviation value at that sampling time. When the active power data is greater than the upper limit of the current reference active power interval, the upper limit is subtracted from the active power data to obtain the power deviation value at that sampling time. When the active power data is between the lower and upper limits of the current reference active power interval, the power deviation value at that sampling time is recorded as zero. Then, the power deviation value is divided by the width of the current reference active power interval to obtain the standardized power deviation value at that sampling time.

[0054] Read the frequency data corresponding to a specific sampling time, and then read the lower and upper limits of the current reference frequency interval. When the frequency data is less than the lower limit of the current reference frequency interval, subtract the frequency data from the lower limit to obtain the frequency deviation value at that sampling time. When the frequency data is greater than the upper limit of the current reference frequency interval, subtract the upper limit to obtain the frequency deviation value at that sampling time. When the frequency data is between the lower and upper limits of the current reference frequency interval, record the frequency deviation value at that sampling time as zero. Then divide the frequency deviation value by the width of the current reference frequency interval to obtain the frequency standardized deviation value at that sampling time. For the same sampling time, the larger of the power standardized deviation value and the frequency standardized deviation value is taken as the slow variable deviation value at that sampling time.

[0055] Arrange all fast variable deviation values ​​from front to back according to the sampling time to obtain the fast variable deviation sequence; arrange all slow variable deviation values ​​from front to back according to the sampling time to obtain the slow variable deviation sequence.

[0056] After forming the deviation sequence, statistical analysis is performed on the continuous preset sampling window corresponding to the current dynamic response sequence. The continuous preset sampling window is a set of continuous sampling points truncated from the current dynamic response sequence according to a preset window length. For example, several consecutive sampling times are selected as a window in the current dynamic response sequence, and the window is slid forward by a preset step size as needed. For a given window, the fast variable deviation values ​​in the fast variable deviation sequence are read one by one, and each fast variable deviation value is compared with the preset fast variable stability threshold corresponding to that fast variable. If a fast variable deviation value is greater than the fast variable stability threshold, the sampling point is counted as falling outside the fast variable stability threshold range. After statistical analysis of all sampling points in the window, the number of sampling points in the fast variable deviation sequence falling outside the fast variable stability threshold range is obtained. Then, the same process is performed on the slow variable deviation sequences in the same window, comparing each slow variable deviation value with the preset slow variable stability threshold corresponding to that slow variable, and counting the number of sampling points in the slow variable deviation sequence falling outside the slow variable stability threshold range.

[0057] When the number of sampling points in the fast variable deviation sequence falling outside the fast variable stability threshold range within a certain window is greater than zero, and the number of sampling points in the slow variable deviation sequence falling outside the slow variable stability threshold range is also greater than zero, it indicates that neither the fast nor slow variable has entered a stable range within that window. At this point, the window cannot be directly classified as a slow recovery phase or a steady-state confirmation phase; further differentiation at the fast variable level is needed to determine whether the current window is closer to the shock phase or the fast recovery phase. Therefore, the current sampling window is divided into a first half and a second half based on the number of sampling points. If the number of sampling points is odd, the middle sampling points are included in the second half. Then, the deviation values ​​of each fast variable in the first half are summed to obtain the total fast variable deviations for the first half. This sum is then divided by the number of sampling points included in the first half to obtain the average fast variable deviation for the first half. Similarly, the deviation values ​​of each fast variable in the second half are summed and divided by the number of sampling points included in the second half to obtain the average fast variable deviation for the second half. If the mean deviation of fast variables in the second half of the window is greater than the mean deviation of fast variables in the first half of the window, it indicates that the overall deviation of fast variables in this window has not decreased, but is in a state of expansion or aggravation. At this time, the rule stage identification result is determined to be the impact stage. If the mean deviation of fast variables in the second half of the window is less than or equal to the mean deviation of fast variables in the first half of the window, it indicates that the overall deviation of fast variables in this window has not increased relative to the first half, but is in a process of stabilization or decline. At this time, the rule stage identification result is determined to be the rapid recovery stage.

[0058] When the number of sampling points in the fast variable deviation sequence falling outside the stability threshold range of the fast variable within a certain window is zero, while the number of sampling points in the slow variable deviation sequence falling outside the stability threshold range of the slow variable is greater than zero, it indicates that all fast variables have entered the stable range, while slow variables still have instances of exceeding the stable range. In this case, based directly on the relationship that the fast variables are stable and the slow variables are not, the rule-based stage identification result is determined to be the slow recovery stage.

[0059] When the number of sampling points in the fast variable deviation sequence that fall outside the stability threshold range of the fast variable within a certain window is equal to zero, and the number of sampling points in the slow variable deviation sequence that fall outside the stability threshold range of the slow variable is also equal to zero, it indicates that both the fast and slow variables have entered their respective stable ranges within that window. At this point, the rule-based identification result is determined as the steady-state confirmation stage.

[0060] After obtaining the rule-based stage identification result, it is necessary to further determine the current dynamic stage by combining the output of the deep learning stage identification model. To do this, the current dynamic response sequence is input into a pre-trained deep learning stage identification model, which outputs a stage category. The stage category output by the model uses the same stage definition as the rule-based stage identification result, namely, one of the following: impact stage, fast recovery stage, slow recovery stage, and steady-state confirmation stage.

[0061] The deep learning stage recognition model is trained using historical load test samples before system deployment. Training samples can be derived from labeled generator load test records. For each historical test record, standardized time-series load test data is obtained by processing it in the same way as the current test action. Then, multiple continuous time-series segments are extracted according to a preset sampling window. Each time-series segment is then labeled with its stage category manually or using existing reliable rules, establishing a correspondence between sample inputs and stage labels. These labeled multi-dimensional time-series samples serve as the training set. Continuous sampled values ​​of voltage, current, active power, and frequency are input into the deep learning stage recognition model, which outputs the probability distribution of the corresponding stage category. During training, a loss function is constructed using the difference between the true stage labels of the samples and the stage probabilities output by the model. The model parameters are iteratively updated to gradually reduce the stage recognition error on the training and validation sets until the preset accuracy requirement is met, resulting in a trained deep learning stage recognition model. After training, during actual testing, the model's input is the current dynamic response sequence, and the output is the model's stage recognition result, i.e., the stage category corresponding to the current test action within that window.

[0062] like Figure 3The figure shows a schematic diagram of the training accuracy curve of the deep learning stage recognition model. During the training process, the training accuracy and verification accuracy of the deep learning stage recognition model generally show an upward trend as the number of training rounds increases, and tend to stabilize after reaching the preset number of training rounds. This indicates that the model's ability to distinguish between the impact phase, the fast recovery phase, the slow recovery phase, and the steady-state confirmation phase gradually increases.

[0063] like Figure 4 The figure shows a schematic diagram of the training loss curve of the deep learning stage recognition model. Corresponding to the improvement of training accuracy, the training loss and validation loss show an overall downward trend and gradually converge in the subsequent training process. This indicates that after the model parameters are iteratively updated, they can learn the stage change rules in the historical loaded test samples well, thus providing a model basis for stage recognition of the current dynamic response sequence in the actual testing process.

[0064] When actually locking onto the current dynamic stage, the rule-based stage identification result is compared with the model-based stage identification result. If they are the same, it means that the stage judgment result based on the numerical threshold and the learning result based on historical samples are consistent. At this time, the stage category corresponding to the consistency is determined as the current dynamic stage. If they are different, it means that the stage characteristics of the current test action within the current observation window are not stable enough or the current window length is insufficient to support consistent judgment. At this time, the current dynamic stage is not directly output, but the observation time of the current test action is extended. After extending the observation time, new voltage data, current data, active power data, and frequency data are collected according to the preset sampling period, and the newly added sampling points are added to the standardized time-series load test data of the current test action. Then, based on the updated standardized time-series load test data, the current dynamic response sequence is re-extracted, and the calculation of fast variable deviation and slow variable deviation, rule-based stage identification, and deep learning stage identification are re-executed. The above process is repeated until the rule-based stage identification result is consistent with the model-based stage identification result. At this time, the stage category corresponding to the consistency is taken as the final locked current dynamic stage. This process avoids misjudgments caused by relying solely on rule-based or model-based judgments, ensuring that the dynamic stage locking of the current test action is based on the consistency of rule analysis and model recognition. This provides a stable stage basis for subsequent stage feature extraction and current state result generation.

[0065] The system reads the current dynamic stage, standardized time-series load test data, and current analysis baseline corresponding to the current test action. The current analysis baseline is determined in the preceding processing. When the load offset judgment result indicates the existence of a load offset, the current analysis baseline is the corrected baseline; when the load offset judgment result indicates no load offset, the current analysis baseline is the target range corresponding to the target load level of the current test action. Stage feature extraction does not use a unified feature set. Instead, based on the different current dynamic stages, data reflecting the changing characteristics of that stage is extracted from the standardized time-series load test data, and calculations are performed according to the corresponding calculation method for that stage.

[0066] When the current dynamic phase is the impact phase, voltage and current data within the current sampling window are extracted from the standardized time-series load test data. For each sampling moment in the current sampling window, the voltage data corresponding to that sampling moment is first read, followed by the upper and lower limits of the voltage interval in the current analysis benchmark. The upper and lower limits of the voltage interval are subtracted from the voltage data, respectively, to obtain two voltage differences. The absolute values ​​of these differences are then taken to obtain the deviation values ​​of that sampling moment relative to the two boundaries of the voltage interval. The smaller of the two absolute values ​​is then selected as the voltage deviation value for that sampling moment. This process is repeated for all sampling moments within the current sampling window to obtain a set of voltage deviation values. Then, the values ​​in this set of voltage deviation values ​​are compared one by one, and the maximum value is taken as the peak voltage deviation. Current data is processed in the same way: for each sampling moment within the current sampling window, the upper and lower limits of the current interval in the current analysis benchmark are subtracted from the corresponding current data, resulting in two current differences. The smaller of the absolute values ​​of these differences is selected as the current deviation value for that sampling moment. The maximum value among all current deviation values ​​is then taken as the peak current deviation. Both the peak voltage deviation and the peak current deviation are used as characteristics of the impact phase. This processing method reflects the maximum deviation of voltage and current from the allowable range during the impact phase of the current test action.

[0067] When the current dynamic phase is the fast recovery phase, voltage and current data within the current sampling window are extracted from the standardized time-series load test data, and the current sampling window is divided into a first half and a second half. If the number of sampling points in the current sampling window is even, half of the sampling points are taken from each half; if the number of sampling points is odd, the first half is divided with one fewer sampling point than the second half. For each sampling moment in the first half window, the voltage deviation value at that sampling moment is calculated using the same method as in the impact phase. Then, all voltage deviation values ​​in the first half window are summed up item by item, and the sum is divided by the number of sampling points in the first half window to obtain the average voltage deviation of the first half window. The average voltage deviation of the second half window is calculated in the same way. The current data is also calculated in the same way, with the average current deviation of the first half window and the average current deviation of the second half window calculated separately. These four averages are used as characteristics of the fast recovery phase. After this processing, the changes in the average deviations in the two sub-windows can reflect whether the voltage and current are transitioning from large deviations to small deviations.

[0068] When the current dynamic phase is the slow recovery phase, active power and frequency data within the current sampling window are extracted from the standardized time-series load test data. The current sampling window is also divided into a first half and a second half. For each sampling moment in the first half window, the corresponding active power data is read, and then the upper and lower limits of the active power interval in the current analysis benchmark are subtracted from each, resulting in two power differences. The smaller of the absolute values ​​of these differences is selected as the power deviation value for that sampling moment. All power deviation values ​​within the first half window are summed, and then divided by the number of sampling points in the first half window to obtain the average power deviation for the first half window. The same calculation process is repeated for the sampling points in the second half window to obtain the average power deviation for the second half window. Frequency data is processed in the same way: at each sampling time, the two differences between the frequency data and the upper and lower limits of the frequency interval in the current analysis benchmark are calculated. The smaller of the absolute values ​​is selected as the frequency deviation value at that sampling time. Then, all frequency deviation values ​​in the first and second half windows are summed and averaged respectively to obtain the mean frequency deviation of the first and second half windows. These four means are used as characteristics of the slow recovery phase. This processing can characterize whether active power and frequency are still converging towards the allowable range when the fast variables have already stabilized.

[0069] When the current dynamic phase is the steady-state confirmation phase, voltage, current, active power, and frequency data within the current sampling window are extracted from the standardized time-series load test data. For each type of data, the deviation value relative to the corresponding interval boundary of the current analysis benchmark is calculated at each sampling time using the same method described above. Then, the deviation values ​​of this type within the current sampling window are summed item by item and divided by the number of sampling points to obtain the mean deviation value of this type of data. Subsequently, to obtain the deviation fluctuation value, the mean deviation value of this type of data is subtracted from the deviation value corresponding to each sampling point within the window to obtain the deviation difference. The absolute value of each deviation difference is taken, and these absolute values ​​are summed item by item and divided by the number of sampling points to obtain the deviation fluctuation value of this type of data. Voltage, current, active power, and frequency are all processed in the above steps. The mean deviation value reflects the overall deviation level of the data relative to the analysis benchmark in the steady-state phase, while the deviation fluctuation value reflects the degree of fluctuation of the data around the average deviation in the steady-state phase.

[0070] After extracting the stage features, the current state result is generated based on these features. If the current dynamic stage is an impact stage, the peak voltage deviation and peak current deviation are read and compared with a pre-set voltage allowable deviation threshold and a pre-set current allowable deviation threshold. If the peak voltage deviation is not greater than the voltage allowable deviation threshold and the peak current deviation is not greater than the current allowable deviation threshold, it means that the maximum deviation of voltage and current is within the allowable range in this impact stage window, and the current state result is recorded as meeting the requirements. If either the peak voltage deviation or the peak current deviation is greater than the corresponding allowable deviation threshold, it means that the maximum deviation in the current impact stage still exceeds the allowable range, and the current state result is recorded as not meeting the requirements.

[0071] If the current dynamic phase is the fast recovery phase, then read the average voltage deviation, average voltage deviation, average current deviation, and average current deviation of the first and second half windows. Subtract the average voltage deviation of the first half window from the average voltage deviation of the second half window to obtain the change in the average voltage deviation; subtract the average current deviation of the first half window from the average current deviation of the second half window to obtain the change in the average current deviation. If the change in the average voltage deviation is less than or equal to zero, and the change in the average current deviation is less than or equal to zero, it indicates that the average voltage deviation and average current deviation have not increased within the current sampling window, but have remained the same or decreased. In this case, the current state result is recorded as meeting the requirements. If either of the above two changes is greater than zero, it indicates that the fast variable deviation still has an increasing trend within the current window. In this case, the current state result is recorded as not meeting the requirements.

[0072] If the current dynamic phase is a slow recovery phase, then read the average power deviation of the first half window, the average power deviation of the second half window, the average frequency deviation of the first half window, and the average frequency deviation of the second half window. Subtract the average power deviation of the first half window from the average power deviation of the second half window to obtain the change in the average power deviation; subtract the average frequency deviation of the first half window from the average frequency deviation of the second half window to obtain the change in the average frequency deviation. If the change in the average power deviation is less than or equal to zero, and the change in the average frequency deviation is less than or equal to zero, it indicates that the slow variable deviations corresponding to active power and frequency have not increased, but have remained flat or decreased. In this case, the current state result is recorded as meeting the requirements. If either of the changes is greater than zero, it indicates that the slow variable deviation is still increasing. In this case, the current state result is recorded as not meeting the requirements.

[0073] If the current dynamic stage is the steady-state confirmation stage, the average deviation and fluctuation value of voltage, current, active power, and frequency are read respectively. Then, the average deviation value of each type is compared with the preset average deviation threshold, and the fluctuation value of each type is compared with the preset fluctuation value. If the average deviation value of all data items is not greater than their respective average deviation threshold, and the fluctuation value of all data items is not greater than their respective fluctuation value, it means that within the steady-state confirmation stage window, all data are not only close to the current analysis benchmark as a whole, but the degree of fluctuation is also within the allowable range. At this time, the current state result is recorded as meeting the requirements. If the average deviation value or fluctuation value of any data item exceeds the corresponding threshold, the current state result is recorded as not meeting the requirements.

[0074] After obtaining the current state result, a test progression decision needs to be generated by combining the load offset judgment result corresponding to the current test action. Specifically, first, the load offset judgment result corresponding to the current test action is read, and then the current state result is read. Then, a combined judgment is performed according to the preset decision rules. If the load offset judgment result indicates that there is a load offset, and the current state result indicates that the requirements are not met, it means that under the current test action, there is both an actual load deviation from the target and the allowable state of the current dynamic stage has not been reached. In this case, the test progression decision is to pause the progression. If the load offset judgment result indicates that there is a load offset, and the current state result indicates that the requirements are met, it means that although there is a load offset, the current stage of operation has met the corresponding requirements. In this case, the test progression decision is not to continue at the original pace, but to slow down the progression. If the load offset judgment result indicates that there is no load offset, and the current state result indicates that the requirements are not met, it means that the current overall load is consistent with the original target, but the current stage of operation has not yet met the requirements. In this case, the test progression decision is to maintain observation. If the load offset judgment result is no load offset, and the current state result meets the requirements, it means that the current action has neither resulted in load offset nor does the current stage state meet the requirements. In this case, the test advancement decision is determined to continue. After the judgment is completed, the test advancement decision is written into the test action table as the basis for subsequent execution of correction and correction verification of the current test action.

[0075] In this embodiment, the preset decision rules can be generated in advance before the test is executed. Specifically, the load offset judgment results, current state results, and subsequent processing methods for each test action in the historical load test samples are read first. Combined with the equipment protection requirements, test process constraints, and the corresponding processed test results, the processing methods under different combinations are summarized to form a decision mapping relationship between the load offset judgment results and the current state results.

[0076] When performing calibration and calibration verification, the test progression decision corresponding to the current test action is first read. If the test progression decision is to continue, the next test action following the current test action in the test action table is executed. In this case, no additional hold or pause processing is required for the next test action; instead, the test proceeds directly according to the existing test rhythm, and the calibration action for the current test action is to continue to the next action. If the test progression decision is to slow down, the next test action, which has already been pre-corrected in the preceding processing, is read, and the next test action following the current test action is executed according to the pre-corrected planned load step size and planned hold time. That is, when executing the next test action, it is no longer executed according to the original planned step size and original planned hold time, but instead uses the adjusted smaller load step size and longer hold time as the calibration processing corresponding to the current action. If the test progression decision is to maintain observation, the load level corresponding to the current test action remains unchanged, and the next test action is not entered, while the observation time of the current test action is extended. During the extended observation time, generator load test data continues to be collected according to the preset sampling period, and new standardized time-series load test data is formed, which is used as the calibration processing for the current test action. If the test progress decision is to pause progress, the next test action will be stopped, and the load level corresponding to the current test action will remain unchanged. That is, the load will not be increased, decreased, or switched, as a correction process for the current test action.

[0077] After performing the above correction process, the correction results need to be verified. During verification, the generator load test data is first re-acquired under the current test action after correction, and time alignment and outlier removal are performed as described above to form corrected standardized time-series load test data. Then, based on the corrected standardized time-series load test data, the load offset judgment is re-executed, that is, the data within the preset analysis time after the start time of the action is extracted again, the current average current, current average active power, temperature rise increment, and load state consistency results are recalculated, and the corrected load offset judgment results are regenerated. Afterwards, the current dynamic stage is re-locked based on the corrected standardized time-series load test data, and the corrected stage features are extracted in the same way as described above to further generate the corrected current state results.

[0078] After obtaining the corrected load offset judgment result and the corrected current state result, a correction verification judgment is performed. If the corrected load offset judgment result is no load offset and the corrected current state result meets the requirements, it means that after the correction process implemented by the current test advancement decision, the current test action has returned to a state of no offset and the current stage meets the requirements. At this time, the correction verification result is recorded as correction verification passed, and the current test action is marked as a progressable state in the test action table, allowing entry into the next test action. If the corrected load offset judgment result still indicates a load offset, or the corrected current state result still does not meet the requirements, it means that the current test action has not recovered to the allowable state after the correction process. At this time, the correction verification result is recorded as correction verification failed, and the load level corresponding to the current test action remains unchanged, not allowing direct entry into the next test action. Through this process, the test advancement decision not only affects the actual execution of the current action, but also allows subsequent correction verification to determine whether the decision has achieved the expected effect, thereby ensuring that the advancement of the entire continuous load test process is based on the verified state.

[0079] It should be understood that determining B based on A does not mean determining B solely based on A; it also means determining B based on A and / or other information.

[0080] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0081] The above description is only an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A generator set load testing method based on multi-source data fusion and deep learning, characterized in that, The method includes: Collect generator set load test data under the current test action, and perform time alignment and outlier removal to form standardized time-series load test data; Based on the standardized time-series load test data, the load offset judgment result is determined, and when the load offset exists, a correction benchmark is formed for the analysis of the current test action, and the next test action in the test action table after the current test action is pre-corrected. Based on the standardized time-series load test data, the current dynamic stage is locked using a combination of rule judgment and deep learning stage recognition. Extract stage features based on the current dynamic stage, and generate the current state result based on the stage features; Based on the load offset judgment result and the current state result, a test advancement decision is generated, and correction and correction verification are performed according to the test advancement decision.

2. The generator set load testing method based on multi-source data fusion and deep learning as described in claim 1, characterized in that, The specific process for generating standardized time-series load test data is as follows: Before the current test action begins, read the start time of the action corresponding to the current test action and establish a sampling time axis corresponding to the start time of the action; Under the sampling time axis, each test item corresponding to the current test action is collected, and the original time series data corresponding to each test item is obtained respectively; The test items include voltage, current, active power, frequency, load status, and enclosure temperature; The original time series data corresponding to each test item is resampled at a preset sampling period so that the data items corresponding to each test item are arranged according to the same sampling time, thus obtaining aligned time series data. For each sampling time in the aligned time series data, it is determined whether there is a missing value in the corresponding data item. If there is a missing value, it is filled in according to the adjacent valid data items before and after the sampling time. For each data item in the completed aligned time series data, calculate the difference between the data item and its preceding and following adjacent data items, and compare the difference with the preset jump variable threshold of the corresponding test item; If the difference is greater than the preset jump variable threshold of the corresponding test item, the data item is marked as an anomaly, and the anomaly is replaced according to the adjacent valid data items before and after the anomaly. After completing the missing value completion and outlier replacement, the test data of each item are combined according to the sampling time to form the standardized time-series load test data corresponding to the current test action.

3. The generator set load testing method based on multi-source data fusion and deep learning as described in claim 1, characterized in that, The specific process for determining the load offset judgment result based on the standardized time-series load test data is as follows: Read the target load level, target load status, and start time of the current test action; Extract data items within a preset analysis duration after the start time of the action from the standardized time-series load test data to form the current load analysis data segment; Statistical processing is performed on the current data and active power data in the current load analysis data segment to obtain the current average current and the current average active power. According to the target load level, read the target active power range and target current range corresponding to the current test action. If the current average active power is within the target active power range and the current average current is within the target current range, the load strength matching result is determined to be within the range; otherwise, the load strength matching result is determined to be outside the range. Extract the chamber temperature corresponding to the start time of the action as the starting chamber temperature, extract the chamber temperature corresponding to the end of the current load analysis data segment as the current chamber temperature, and determine the temperature rise increment based on the difference between the current chamber temperature and the starting chamber temperature. If the temperature rise increment is greater than the preset temperature rise threshold, it is determined that the current load state is affected by the continuous temperature rise; otherwise, it is determined that the current load state is not affected by the continuous temperature rise. Extract the load status from the current load analysis data segment. If the load status is consistent with the target load status, the current load status consistency result is determined to be consistent; otherwise, the current load status consistency result is determined to be inconsistent. When the load strength matching result is within the range, the temperature rise effect result is not affected by continuous temperature rise, and the load state consistency result is consistent, the load offset judgment result is determined to be no load offset. When the load strength matching result is outside the range, or the temperature rise effect result is affected by continuous temperature rise, or the load state consistency result is inconsistent, the load offset judgment result is determined to be that there is a load offset.

4. The generator set load testing method based on multi-source data fusion and deep learning as described in claim 1, characterized in that, When a load offset exists, a correction benchmark is formed for the analysis of the current test action. The specific process is as follows: Obtain the current average active power and current average current, and use them as the corrected power reference and corrected current reference, respectively; Based on the preset power reference deviation length and current reference deviation length, a corrected power range is generated with the corrected power reference as the center, and a corrected current range is generated with the corrected current reference as the center. The modified power range, the modified current range, and the load state corresponding to the current test action are combined as a modification benchmark, and the modification benchmark is used to replace the target range corresponding to the target load level of the current test action. This is used for subsequent dynamic stage locking, stage feature extraction, and state determination of the current test action.

5. The generator set load testing method based on multi-source data fusion and deep learning as described in claim 1, characterized in that, The process of pre-correcting the next test action in the test action table after the current test action is as follows: Read the next test action in the test action table that follows the current test action, and extract the planned loading step size and planned hold time corresponding to the next test action; When the current average active power is greater than the upper limit of the target active power range, reduce the planned loading step size and extend the planned hold time; When the current average active power is less than the lower limit of the target active power range, the planned loading step size remains unchanged, and the planned holding time is extended. When the current average active power is within the target active power range, the planned loading step size and planned hold time remain unchanged; The adjusted planned loading step size and planned hold time are written into the test action table as a pre-correction result for subsequent test actions.

6. The generator set load testing method based on multi-source data fusion and deep learning as described in claim 1, characterized in that, The specific locking process for the current dynamic phase is as follows: When the load offset judgment result corresponding to the current test action is that there is a load offset, the correction benchmark corresponding to the current test action is read as the current analysis benchmark. When the load offset judgment result corresponding to the current test action is no load offset, the target range corresponding to the target load level of the current test action is read as the current analysis benchmark. Extract the current reference voltage range, current reference current range, current reference active power range, and current reference frequency range from the current analysis reference. Voltage, current, active power, and frequency data within a continuous time period are extracted from standardized time-series load test data according to a preset sampling period to form the current dynamic response sequence; For each sampling time in the dynamic response sequence, calculate the difference pair between the voltage data and the upper and lower limits of the current reference voltage range, and calculate the difference pair between the current data and the upper and lower limits of the current reference current range to obtain a set of fast variable deviation value pairs; Calculate the difference pairs between the active power data and the upper and lower limits of the current reference active power range, and the difference pairs between the frequency data and the upper and lower limits of the current reference frequency range to obtain a set of slow variable deviation value pairs; Arranged according to sampling time, a fast variable bias sequence and a slow variable bias sequence are formed; Count the number of sampling points in the fast variable deviation sequence that fall outside the stability threshold range of the fast variable, and count the number of sampling points in the slow variable deviation sequence that fall outside the stability threshold range of the slow variable. When the number of sampling points in the fast variable deviation sequence that fall outside the fast variable stability threshold range is greater than zero, and the number of sampling points in the slow variable deviation sequence that fall outside the slow variable stability threshold range is greater than zero, calculate the mean fast variable deviation of the first half window and the mean fast variable deviation of the second half window respectively within the current sampling window. When the mean deviation of the fast variables in the second half of the window is greater than the mean deviation of the fast variables in the first half of the window, the rule-based identification result is determined to be the impact stage; when the mean deviation of the fast variables in the second half of the window is less than or equal to the mean deviation of the fast variables in the first half of the window, the rule-based identification result is determined to be the rapid recovery stage. When the number of sampling points in the fast variable deviation sequence that fall outside the fast variable stability threshold is zero, and the number of sampling points in the slow variable deviation sequence that fall outside the slow variable stability threshold is greater than zero, the identification result of the rule determination stage is the slow recovery stage. When the number of sampling points in the fast variable deviation sequence that fall outside the fast variable stability threshold range is zero, and the number of sampling points in the slow variable deviation sequence that fall outside the slow variable stability threshold range is zero, the identification result of the rule determination stage is the steady-state confirmation stage. Input the current dynamic response sequence into the deep learning stage recognition model to obtain the model stage recognition result; When the rule stage identification result is consistent with the model stage identification result, the corresponding stage category is determined as the current dynamic stage; When the identification results of the rule stage are inconsistent with those of the model stage, the observation time of the current test action is extended and the current dynamic response sequence is reacquired. The rule stage identification and model stage identification are re-executed until the current dynamic stage is locked.

7. The generator set load testing method based on multi-source data fusion and deep learning as described in claim 1, characterized in that, The extraction of stage features based on the current dynamic stage specifically includes: Read the current dynamic stage, standardized timing load test data, and current analysis benchmark corresponding to the current test action; When the current dynamic stage is the impact stage, the voltage and current data within the current sampling window are extracted from the standardized time-series load test data, and the peak voltage deviation and peak current deviation are calculated as characteristics of the impact stage. When the current dynamic phase is the fast recovery phase, the voltage and current data within the current sampling window are extracted from the standardized time-series load test data. The average voltage deviation of the first half window, the average voltage deviation of the second half window, the average current deviation of the first half window, and the average current deviation of the second half window are calculated respectively as characteristics of the fast recovery phase. When the current dynamic phase is the slow recovery phase, the active power data and frequency data within the current sampling window are extracted from the standardized time-series load test data. The average power deviation of the first half window, the average power deviation of the second half window, the average frequency deviation of the first half window, and the average frequency deviation of the second half window are calculated respectively as features of the slow recovery phase. When the current dynamic stage is the steady-state confirmation stage, voltage data, current data, active power data and frequency data within the current sampling window are extracted from the standardized time-series load test data. The mean deviation and fluctuation value of the corresponding data relative to the current analysis benchmark are calculated respectively as characteristics of the steady-state confirmation stage.

8. The generator set load testing method based on multi-source data fusion and deep learning as described in claim 1, characterized in that, The specific process for generating the current state result based on stage features is as follows: Read the current dynamic stage, stage characteristics, and current analysis benchmark corresponding to the current test action; During the impact phase, if the peak voltage deviation and peak current deviation are not greater than the corresponding allowable deviation threshold, the current state result is determined to meet the requirements; otherwise, the current state result is determined to not meet the requirements. During the rapid recovery phase, if the average voltage deviation of the second half window is not greater than the average voltage deviation of the first half window, and the average current deviation of the second half window is not greater than the average current deviation of the first half window, the current state result is determined to meet the requirements; otherwise, the current state result is determined to not meet the requirements. During the slow recovery phase, if the average power deviation of the second half window is not greater than the average power deviation of the first half window, and the average frequency deviation of the second half window is not greater than the average frequency deviation of the first half window, the current state result is determined to meet the requirements; otherwise, the current state result is determined to not meet the requirements. During the steady-state confirmation phase, if the mean deviation and the deviation fluctuation value are not greater than the corresponding threshold, the current state result is determined to meet the requirements; otherwise, the current state result is determined to not meet the requirements.

9. The generator set load testing method based on multi-source data fusion and deep learning as described in claim 1, characterized in that, The process of generating a test advancement decision based on the load offset judgment result and the current state result is as follows: Read the load offset judgment result and current status result corresponding to the current test action; When the load offset judgment result indicates that a load offset exists, and the current state result indicates that the requirements are not met, the test advancement decision is to suspend advancement. When the load offset judgment result indicates that a load offset exists, and the current state result meets the requirements, the test advancement decision is determined to be to slow down the advancement. When the load offset judgment result is no load offset and the current state result is that the requirements are not met, the test progress decision is determined to be to maintain observation; When the load offset judgment result is no load offset and the current state result meets the requirements, the test advancement decision is determined to continue. Write the test advancement decision into the test action table.

10. The generator set load testing method based on multi-source data fusion and deep learning as described in claim 1, characterized in that, The specific process of performing correction and verification based on test advancement decisions is as follows: Read the test progress decision corresponding to the current test action; When the test progression decision is to continue, the test will continue to be executed according to the next test action after the current test action in the test action table, and the correction will be performed accordingly. When the test advance decision is to slow down the advance, the next test action is read after pre-correction, and the action is executed according to the pre-corrected plan loading step size and plan hold time to perform the correction; When the test progression decision is to maintain observation, the load level corresponding to the current test action remains unchanged, and the observation duration of the current test action is extended to re-collect standardized timing load test data in order to perform correction. When the test progress decision is to pause progress, the next test action is stopped, and the load level corresponding to the current test action remains unchanged, thus performing the correction. After performing the corresponding correction, the generator set load test data under the current test action is re-acquired, and time alignment and outlier removal are performed to form the corrected standardized time-series load test data. Based on the corrected standardized time-series load test data, the load offset judgment result is re-determined, the current dynamic stage is locked, the stage features are extracted, and the corrected current state result is generated. When the corrected load offset judgment result is no load offset and the corrected current state result meets the requirements, the correction verification is determined to be successful, the current test action is marked as a progressable state, and the next test action is allowed. If the corrected load offset judgment result indicates that a load offset exists, or if the corrected current state result indicates that the requirements are not met, the correction verification is determined to have failed, and the load level corresponding to the current test action remains unchanged.

Images