A time mixer-based regional power grid short-time voltage disturbance dominant cause identification method

Abstract

The present application belongs to the technical field of intelligent monitoring of power system, and discloses a regional power grid short-time voltage disturbance leading cause identification method based on TimeMixer. Firstly, the multi-channel time sequence data is unified through the maximum common sampling period, the voltage change boundary time is accurately identified, and the change segment division and position marking are performed on the disturbance chain time sequence data. Then, the change segment features are extracted from three time scales of coarse, medium and fine by using the TimeMixer++ multi-scale hybrid encoder, and the intervention effect of each leading segment on the voltage disturbance is evaluated by using the counterfactual conditional modeling method. Finally, the leading cause object is determined based on the intervention effect score and the time sequence constraint. The present application realizes the accurate quantification of the voltage disturbance causal relationship and the automatic identification of the leading cause, breaks through the limitation of traditional experience judgment, and provides reliable technical support for power grid fault analysis, predictive maintenance and operation optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for power systems, and more specifically, to a method for identifying the dominant causes of short-term voltage disturbances in regional power grids based on TimeMixer. Background Technology

[0002] In actual power grid operation, under scenarios involving cascading operation of multiple devices, traditional analysis methods rely excessively on experience-based judgments and simple rules, making it difficult to penetrate complex event chains and accurately pinpoint the true source of disturbances. This leads to attribution bias in numerous fault analysis reports. Especially in large substation disturbance events, when protection devices, circuit breakers, and control equipment respond simultaneously, the boundary between the cause and the response becomes blurred. The diverse and heterogeneous data collected by power grid monitoring systems exhibit significant asynchrony and inconsistency. The sampling frequencies of devices such as PMUs, fault recorders, and SCADA systems differ significantly. Traditional interpolation methods introduce human distortion when processing these timing differences, resulting in misalignment of critical time points. Existing technologies lack quantitative assessment methods for the causal relationship between electrical quantity fluctuations and equipment actions, making it impossible to further determine the intensity of influence based on temporal sequence. This makes the analysis results highly dependent on the analyst's personal experience, and different experts often have significantly different judgments on the same event. Power grid disturbance signals exhibit significant multi-scale characteristics, ranging from millisecond-level harmonic fluctuations to second-level power fluctuations. Traditional feature extraction methods struggle to simultaneously capture these different time-scale feature patterns, particularly in complex disturbances involving oscillations, transients, and steady-state processes. Power system analysis suffers from a severe lack of counterfactual reasoning capabilities, failing to effectively simulate hypothetical scenarios of "system response when specific equipment does not operate," resulting in an inability to quantitatively differentiate the contribution of each event. This is especially true when multiple protection systems operate simultaneously, making it difficult to determine which protection action is the dominant cause. Furthermore, the power grid disturbance analysis process is highly fragmented, often requiring multiple systems and manual intervention from data acquisition and feature extraction to result analysis. This not only leads to lengthy analysis cycles but also frequent human bias, severely limiting analytical capabilities when facing large-scale disturbances. Current assessment systems also tend to focus on a single indicator of voltage deviation amplitude, neglecting dimensions such as disturbance duration and propagation range. This makes it difficult to comprehensively reflect the severity of disturbances, resulting in disturbances with the same amplitude but vastly different impact ranges receiving the same rating, hindering the development of differentiated response strategies.

[0003] In view of this, the present invention proposes a method for identifying the dominant causes of short-term voltage disturbances in regional power grids based on TimeMixer to solve the above problems. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a method for identifying the dominant causes of short-time voltage disturbances in regional power grids based on TimeMixer, comprising: Acquire multi-channel time-series data and device action records of a single short-time voltage disturbance in the regional power grid. Calculate the maximum common sampling period according to the sampling frequency of each channel. Resample the multi-channel time-series data and device action records to the unified sampling time corresponding to the maximum common sampling period to form disturbance chain time-series data. For the voltage channel numerical sequence of the disturbance chain time series data, calculate the absolute value of the difference between adjacent sampling points, identify the sampling time when the absolute value of the difference first exceeds the preset change threshold, and record it as the voltage change boundary time; The signed time difference of each sampling moment is calculated based on the voltage change boundary moment to form a boundary moment reference sequence. The device action record is converted into an action indication sequence aligned with the unified sampling moment and written into the disturbance chain timing data. Based on the boundary time reference sequence, the perturbation chain time series data is divided into change segments for each channel. The chain position markers are assigned according to the positional relationship of the time range of each change segment relative to the voltage change boundary time. The change segments with chain position markers are input into the TimeMixer++ multi-scale hybrid encoder to obtain the leader segment features to be evaluated and form a list of leader segment features to be evaluated. The feature list of the lead segments to be evaluated is traversed one by one. The starting sampling point values ​​of the data corresponding to the time period of each lead segment to be evaluated are extended along the time axis to form a suppression reference value sequence. The time period data is replaced by the suppression reference value sequence and then combined with the data of the other time periods to form counterfactual conditional data. The counterfactual voltage sequence is obtained by inputting it into the consistency generation model. The amplitude difference index and duration difference index are calculated within the preset scoring window before and after the voltage change boundary time. The intervention effect score is formed by combining the results and associating it with the lead segment identifier to form an intervention effect score list. Based on the intervention effect score list, compare the intervention effect score with the dominant threshold, take the leader fragment identifier with the largest intervention effect score among those that meet the dominant threshold, and combine it with the corresponding intra-chain position mark to form the dominant leader fragment; The end time of the dominant leader fragment is extracted and compared with the voltage change boundary time. If the end time is earlier than the voltage change boundary time, the dominant leader fragment is identified as the dominant cause object. The output includes the dominant cause object identifier, the time range of the dominant leader fragment, and the intervention effect score.

[0005] The technical effects and advantages of the present invention, a method for identifying the dominant causes of short-time voltage disturbances in regional power grids based on TimeMixer, are as follows: This invention enhances the analysis capabilities of short-term voltage disturbances in regional power grids. By accurately identifying the dominant causal mechanisms, it solves the attribution problem in complex disturbance scenarios, shifting power grid fault location from experience-based judgment to quantitative analysis. The time alignment and fusion technology of multi-channel heterogeneous data eliminates time-series analysis errors caused by inconsistent sampling, ensuring high data fidelity. The quantitative assessment framework for causal relationships overcomes the limitation of relying solely on temporal sequence, establishing an objective measurement system for event impact. The ability to capture multi-scale time-series features enhances the system's accuracy in identifying electrical phenomena at different time scales, performing particularly well in complex disturbances. The counterfactual conditional analysis method enables accurate modeling of hypothetical scenarios, solving the attribution problem in situations where multiple protection systems operate simultaneously. The automated analysis process eliminates human intervention, improving analysis efficiency and ensuring result consistency. The multi-dimensional comprehensive evaluation system fully considers the amplitude and temporal characteristics of disturbances, providing a precise basis for differentiated processing strategies. Attached Figure Description

[0006] Figure 1 This is a schematic diagram of a method for identifying the dominant causes of short-term voltage disturbances in a regional power grid based on TimeMixer, according to the present invention. Figure 2 This is a schematic diagram illustrating the identification of voltage change boundary moments according to the present invention; Figure 3 This is a schematic diagram of the encoder network architecture of the present invention; Figure 4 This is a schematic diagram illustrating the quantitative evaluation of the indicators in this invention; Figure 5 This is a schematic diagram of the data storage and transfer architecture of the system of the present invention. Detailed Implementation

[0007] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0008] This application provides a method for identifying the dominant causes of short-term voltage disturbances in regional power grids based on TimeMixer. The entities executing this method include, but are not limited to, power grid monitoring systems, power fault analysis platforms, smart grid diagnostic systems, and power disturbance tracing platforms, which can be considered as general computing nodes in this application.

[0009] Please see Figure 1 In this embodiment of the invention, a specific implementation process of a method for identifying the dominant cause of short-term voltage disturbances in a regional power grid based on TimeMixer includes: This system acquires multi-channel time-series data and device operation records of a single short-term voltage disturbance in the regional power grid. The maximum common sampling period (MCS) is calculated based on the sampling frequency of each channel. The multi-channel time-series data and device operation records are then resampled to a unified sampling time corresponding to the MCS, forming disturbance chain time-series data. Multi-channel time-series data of short-term voltage disturbances typically includes time-series records of electrical quantities such as voltage, current, and power, as well as operation records of protection devices, switching equipment, and other devices. These data originate from multiple acquisition points in the power grid monitoring system, and the sampling frequencies may be inconsistent. By calculating the MCS, the system can unify data from different frequencies onto the same time axis, ensuring time alignment and data integrity for subsequent analysis. The resampling process utilizes numerical interpolation and averaging algorithms to map all channel data onto a unified time axis, forming standardized disturbance chain time-series data, providing a foundation for subsequent identification of voltage change boundary moments.

[0010] For the voltage channel numerical sequence of the disturbance chain time series data, the absolute value of the difference between adjacent sampling points is calculated. The sampling moment when the absolute value of the difference first exceeds a preset change threshold is identified and recorded as the voltage change boundary moment. The voltage change boundary moment is a key time marker for identifying the start point of the disturbance. By performing differential analysis on the voltage channel, the moment of voltage abrupt change can be accurately captured. The formula for calculating the absolute value of the difference is: ; in, For a moment The absolute value of the voltage difference, and They are time points and The voltage value. When First time exceeding the preset change threshold At that time, the corresponding sampling time is marked as the voltage change boundary time. This moment serves as the time anchor for perturbation chain analysis, providing a benchmark for subsequent determinations of chronological relationships. (See reference...) Figure 2 The diagram illustrates the identification of voltage change thresholds: the upper layer is the original voltage waveform, and the lower layer is the differential sequence, marking the threshold values ​​that have exceeded the threshold. The moment And the positive and negative signed time difference based on this.

[0011] The signed time difference at each sampling time is calculated based on the voltage change boundary moment, forming a boundary moment reference sequence. The device action record is then converted into an action indication sequence aligned with the unified sampling time and written into the disturbance chain timing data. The boundary moment reference sequence is a crucial step in establishing a relative time coordinate system, enabling the differentiation between events before and after the disturbance. The formula for calculating the signed time difference is: ; in, Sampling time relative to the voltage change boundary moment The signed time difference is used, with positive values ​​indicating sampling points after the boundary time and negative values ​​indicating sampling points before the boundary time. Alignment and transformation of device action records ensure that all device action information is mapped to a unified time axis, facilitating subsequent analysis of the temporal relationship between action sequences and voltage disturbances. Action indication sequences are represented using binary values, set to 1 at the moment of device action and during its duration, and set to 0 at other times, thus fully recording the temporal characteristics of device state changes.

[0012] The perturbation chain time series data is segmented into change segments based on the boundary time reference sequence. Intra-chain position markers are assigned according to the positional relationship of each change segment's time range relative to the voltage change boundary time. These segmented change segments with intra-chain position markers are input into the TimeMixer++ multi-scale hybrid encoder to obtain the features of the leader segments to be evaluated, forming a list of leader segment features to be evaluated. Change segment segmentation is a fundamental step in identifying discrete events in the perturbation chain. Meaningful data segments are determined by analyzing the change characteristics of each channel's data. The segmentation process determines the preceding and following change segments for each channel based on the boundaries between continuous unchanged segments and significantly changed segments of the channel values, combined with the boundary time reference sequence. The intra-chain position markers ("preceding" or "following") indicate the time position of the segment relative to the voltage change boundary time, providing an important basis for causal relationship determination.

[0013] The TimeMixer++ multi-scale hybrid encoder is the core technological innovation of this method. It captures the feature patterns of changing segments at different time scales through a multi-scale temporal convolutional network. The encoder takes the sequence of changing segments processed to the same length as input and extracts low-frequency trend components, mid-frequency fluctuation components, and high-frequency detail components at three time scales: coarse, medium, and fine. Then, it generates fixed-dimensional feature vectors through channel mixing mapping, forming a feature list of leading segments to be evaluated.

[0014] It's worth noting that the TimeMixer++ multi-scale hybrid encoder is an innovative temporal feature extraction architecture that integrates the attention mechanism of the Transformer with the multi-scale processing capabilities of the Temporal Convolutional Network (TCN). Compared to traditional temporal models such as RNNs, LSTMs, or single-scale Transformers, TimeMixer++ introduces parallel multi-scale analysis channels in the temporal dimension. This architecture employs three parallel temporal convolutional branches, each configured with different kernel sizes and strides: a coarse-scale branch (large kernels, such as 64-128) captures long-term trends; a meso-scale branch (medium kernels, such as 16-32) extracts periodic fluctuations; and a fine-scale branch (small kernels, such as 4-8) identifies transient details. This design borrows from the multi-resolution concept in wavelet analysis but achieves adaptive feature extraction through learnable convolutional kernels. TimeMixer++ also innovatively introduces a cross-scale attention mechanism, enabling features at different scales to interact with each other, overcoming the limitation of feature independence in traditional multi-scale CNNs. This architecture is particularly well-suited for time-series analysis of power systems because grid disturbances simultaneously include millisecond-level transient changes (protection actions), second-level fluctuations (power oscillations), and minute-level trends (load changes). TimeMixer++ can capture and fuse these features across different time scales.

[0015] For reference Figure 3 The diagram shows the encoder network architecture: it illustrates how the input segment goes through parallel convolutions of three branches: coarse scale (trend), medium scale (fluctuation), and fine scale (detail), as well as the final feature concatenation and channel mixing mapping layer.

[0016] The feature list of the lead segments to be evaluated is traversed one by one. The starting sampling point values ​​of the data corresponding to the time period of each lead segment to be evaluated are extended along the time axis to form a suppression reference value sequence. The time period data is replaced by the suppression reference value sequence and then combined with the data of the other time periods to form counterfactual conditional data. The counterfactual voltage sequence is obtained by inputting it into the consistency generation model. The amplitude difference index and duration difference index are calculated within the preset scoring window before and after the voltage change boundary time. The intervention effect score is formed by combining the results and associating it with the identification of the lead segment to be evaluated to form an intervention effect score list.

[0017] For reference Figure 4 The diagram illustrates the quantitative evaluation of the indicators: the scoring window range is marked on the waveform, which intuitively shows the area difference (amplitude difference) and recovery time difference (duration difference) between the real waveform and the counterfactual waveform.

[0018] Counterfactual conditional analysis is a key method for assessing the causal impact of change segments. It quantifies the contribution of these segments to voltage disturbances by constructing hypothetical scenarios of "if the change had not occurred." A suppressed reference value sequence is used to simulate a suppressed change by extending the steady-state value at the segment's initial sampling point along the time axis. Counterfactual conditional data replaces the original change segment while retaining real data from other time periods, creating a counterfactual scenario. A consistency generation model receives the counterfactual conditions and predicts the voltage sequence evolution under those conditions. By comparing the differences between the counterfactual and actual voltage sequences within a scoring window, amplitude difference and duration difference indices are calculated to comprehensively assess the intervention effect of the change segment on voltage disturbances. The formula for calculating the intervention effect score is: ;

[0019] in, To score the intervention effect, For amplitude difference index, As a duration difference indicator, and These are the weighting coefficients, and The intervention effect scoring checklist records the identifiers and corresponding scores of each lead segment to be evaluated, providing a quantitative basis for identifying the dominant trigger.

[0020] The intervention effect scores are compared with the dominant threshold based on the intervention effect score list. The leader segment with the highest intervention effect score among those satisfying the dominant threshold is identified and combined with its corresponding position marker within the chain to form the dominant leader segment. Dominant threshold screening is a crucial step in identifying key influencing factors. By setting a lower limit for the intervention effect score, candidate inducements with significant influence are screened out. When multiple segments simultaneously satisfy the dominant threshold, the one with the highest intervention effect score is selected first. If there are cases with the same score, the end time and start time are further compared to ensure the selection of a unique dominant leader segment. This multi-level screening strategy ensures the uniqueness and optimality of inducement identification and avoids decision ambiguity caused by multiple solutions.

[0021] The process extracts the end time of the dominant leader segment and compares it with the voltage change boundary time. If the end time is earlier than the voltage change boundary time, the dominant leader segment is identified as the dominant cause. The output includes the dominant cause identification, the time range of the dominant leader segment, and the intervention effect score. Verifying the temporal order is crucial to ensuring the rationality of causal logic; only leader segments whose end time is earlier than the voltage change boundary time can be confirmed as valid cause objects. If the dominant leader segment does not meet the temporal order constraint, it is recursively reselected from the remaining candidates until a cause object that meets the conditions is found or it is confirmed that no valid cause exists. The final output of the dominant cause identification result includes key information such as the cause object identification, time range, and intervention effect score, providing power grid operation and maintenance personnel with accurate location and quantitative assessment of disturbance sources.

[0022] In this embodiment of the invention, the maximum common sampling period is calculated according to the sampling frequency of each channel, and the multi-channel time-series data and device action records are resampled to a unified sampling time corresponding to the maximum common sampling period, including: The sampling frequency of each channel's time-series data is extracted, and the greatest common divisor (GCD) of all channel sampling frequencies is calculated. The sampling period corresponding to the GCD is used as the unified sampling period. Calculating the GCD of sampling frequencies is fundamental to achieving data time alignment, ensuring that data from different frequencies can be compared at a common time point. The calculation process first extracts the original sampling frequency of each channel. Then, the Euclidean algorithm is used to find the greatest common divisor of all frequencies. Unified sampling period It is the reciprocal of the greatest common divisor, that is: ;

[0023] This sampling period selection based on the greatest common divisor ensures that key information in the original data is not lost during resampling, while minimizing data storage and computing resource requirements.

[0024] For channels with sampling frequencies higher than the uniform sampling frequency, the average of the original sampling points between adjacent uniform sampling times is taken as the channel value at that uniform sampling time. High-frequency data downsampling is an effective method to preserve signal characteristics while reducing data redundancy. Regarding the sampling frequency... Higher than the uniform sampling frequency The channel, each uniform sampling time The values ​​are calculated through a time window, The arithmetic mean of the original sampling points is used to obtain the signal. This averaging method effectively reduces the impact of high-frequency noise while preserving the main characteristics of the signal, ensuring that the downsampled data can still accurately reflect the changing trend of the original signal.

[0025] For channels with sampling frequencies lower than the uniform sampling frequency, linear interpolation is used between adjacent original sampling points to complete the channel values ​​at the uniform sampling time. Low-frequency data interpolation is a key step in solving the sampling sparsity problem, ensuring data continuity by reasonably estimating the values ​​at unsampled times. Regarding the sampling frequency... Below the uniform sampling frequency The channel, in adjacent original sampling points and Uniform sampling time between The corresponding values ​​are calculated using linear interpolation: ;

[0026] in, Indicates time The channel value.

[0027] In practical applications, the linear interpolation calculation process can be illustrated through the following specific example. The sampling frequency of phase A voltage channel in a 220kV substation is 2Hz, while the uniform sampling frequency is 5Hz. In the original data, Voltage value measured in seconds , Voltage value measured in seconds We now need to interpolate and calculate the unified sampling time. Voltage value at seconds .

[0028] Substitute the specific parameters: ; Through linear interpolation, we obtained Voltage value at a second This value will be used as a data point on a unified sampling time axis for subsequent analysis. In a real system, this type of interpolation calculation will be automatically performed on all channels and time points that require upsampling, ensuring that all data are aligned to a unified time axis, laying the foundation for accurately identifying voltage change boundary moments.

[0029] Linear interpolation methods achieve a good balance between computational efficiency and accuracy, and are suitable for the smooth variation characteristics of most electrical parameters.

[0030] For device action records, the time of action occurrence is mapped to the most recent uniform sampling time. At this uniform sampling time and its subsequent durations, the action indicator value is set to 1, and all others are set to 0, forming an action indicator sequence. Binarization of the device action records is a crucial step in transforming discrete events into continuous time-series data. (Action occurrence time) First, map to the most recent uniform sampling time. Then and all sampling times during the duration of the subsequent actions. Setting the action indicator value to 1 indicates that the device is active; setting it to 0 at other times indicates that the device is not active. This binarization method integrates the temporal characteristics of device state changes with other continuous analog quantities into the same data structure, providing a unified interface for subsequent comprehensive analysis.

[0031] In this embodiment of the invention, the perturbation chain time series data is segmented into change segments for each channel according to the boundary time reference sequence, including:

[0032] Identify the boundary sampling point where the signed time difference changes from negative to positive from the boundary time reference sequence. Use the sampling time before this boundary sampling point as the end time parameter of the segment before the boundary, and the sampling time of this boundary sampling point as the start time parameter of the segment after the boundary. Time axis segmentation is the fundamental step in constructing the segments before and after the boundary, accurately dividing the data intervals before and after the perturbation through the boundary time. First, search in the boundary time reference sequence... The inflection point from negative to non-negative values, corresponding to the sampling point index. The voltage change boundary moments are marked. The termination time parameter of the segment before the boundary is set to... The start time parameter of the segment after the boundary is set to This boundary determination method ensures seamless connection between segments before and after the boundary, avoids data overlap or omission, and provides an accurate time range definition for subsequent extraction of changed segments.

[0033] For the time series data of each channel, the end time of the continuous unchanged segment in the channel's numerical sequence is used as the start time parameter of the segment before the boundary, and the start time of the continuous unchanged segment in the channel's numerical sequence is used as the end time parameter of the segment after the boundary. Identifying the changing segments within a channel is a crucial step in extracting effective information. The time boundaries of the changing segments are determined by detecting the steady state and changing intervals of the numerical sequence. First, change detection is performed on the numerical sequence of each channel to identify continuous unchanged segments (continuous intervals where the numerical change is below a preset threshold) and changing segments. Within the interval before the boundary, the end time of the last continuous unchanged segment is used as the starting time parameter. The parameter representing the start time of the segment before the boundary; and the parameter representing the start time of the first continuous, unchanged segment within the interval after the boundary. This serves as the termination time parameter for the segment after delimitation. This segment boundary determination method based on change detection can adaptively capture the change characteristics of different channels and extract the most informative data intervals.

[0034] Each channel is segmented according to start and end time parameters to obtain pre-boundary and post-boundary change segments. Pre-boundary change segments are assigned positions within the chain and marked as "previous," while post-boundary change segments are assigned positions within the chain and marked as "postvous." Segmentation and labeling of change segments are the final steps in constructing structured data, transforming time-series data into labeled discrete segments. For each channel, based on the aforementioned determined start and end time parameters, data within the corresponding time interval is extracted from the time-series data to form the pre-boundary change segment. and the changes after the boundary Simultaneously, each segment is assigned an intra-chain position marker, with segments before the boundary marked as "previous" and segments after the boundary marked as "subsequent." These markers directly relate to the segment's position in the causal chain, providing crucial temporal information for subsequent causal analysis. Through this segmentation and marking method, continuous time-series data is transformed into a structured set of changing segments, each carrying rich information such as time range, numerical sequence, and position markers.

[0035] In this embodiment of the invention, the changed segment with in-chain position markers is input into the TimeMixer++ multi-scale hybrid encoder to obtain the features of the leader segment to be evaluated, including: Each variation segment is uniformly padded to a preset maximum length according to the time axis. The padding value is taken from the starting sampling point value of each variation segment, forming a sequence of variation segments of equal length. Length uniformity is a prerequisite for batch processing; padding techniques are used to transform segments of different lengths into fixed-dimensional inputs. The processing first determines the preset maximum length. This is typically set to the 90th-95th percentile of the length of all varied segments, and then for each varied segment... Adjust the length: For lengths exceeding For segments that are too short, shorten them to the target length by downsampling or truncation; for segments that are too short... The segment, using its starting sampling point value A padding expansion is performed. The padding strategy selects the starting sampling point value instead of zero or the mean, which can maintain the continuity and steady-state characteristics of the sequence and reduce the interference of artificial boundary effects on feature extraction. The altered sequence after equal-length processing retains the temporal dynamic characteristics of the original data, while also meeting the requirement of deep learning models for input dimension consistency.

[0036] The TimeMixer++ multi-scale hybrid encoder performs temporal convolution operations on equal-length variable-length sequences at three time scales: coarse, medium, and fine, extracting low-frequency trend components, mid-frequency fluctuation components, and high-frequency detail components, respectively. Multi-scale feature extraction is a core technology for capturing temporal patterns, identifying signal features in different frequency domains through convolution operations at different scales. The TimeMixer++ encoder employs a hierarchical temporal convolutional network architecture, processing the input sequence in parallel across the three time scales. Coarse-scale layers use larger convolutional kernels (such as 64 or 128) and large strides to capture low-frequency trend components and reflect the overall direction of signal change. The mesoscale layer uses a medium convolution kernel (such as 16 or 32) and a medium stride to extract the mid-frequency fluctuation component, reflecting the periodic variation pattern of the signal. Fine-scale layers use small convolutional kernels (such as 4 or 8) and small strides to capture high-frequency detail components and reflect the instantaneous change characteristics of the signal; The convolutional operations at each scale layer employ a causal convolution design, ensuring that feature extraction depends only on data from the current and historical moments, thus meeting the causal requirements of time-series analysis. The advantage of this multi-scale design lies in its ability to simultaneously capture changing features at different time scales, comprehensively characterizing the time-frequency properties of the signal and improving feature representation capabilities.

[0037] The low-frequency trend component, mid-frequency fluctuation component, and high-frequency detail component are concatenated along the channel dimension and then compressed into a fixed-dimensional feature vector using a channel-mixed linear mapping. This vector serves as the leading feature for the segment to be evaluated. Feature fusion and compression are key steps in generating a compact representation, forming a comprehensive description by effectively integrating multi-scale features. The fusion process first concatenates feature components from three scales along the channel dimension to form a multi-channel feature tensor. Then, the concatenated features are processed through a channel-mixing linear mapping layer to achieve cross-scale feature interaction and information fusion. The channel-mixing mapping employs an attention mechanism design, which can adaptively adjust the importance weights of features at different scales. The final output feature vector dimension is typically 128-512, which preserves the rich expression of multi-scale information while optimizing computational efficiency. This feature encoding method based on TimeMixer++ has significant advantages in capturing complex temporal patterns and is particularly suitable for signal analysis with multi-scale dynamic characteristics, such as power grid disturbances.

[0038] In this embodiment of the invention, counterfactual conditional data is input into a consistency generation model to obtain a counterfactual voltage sequence, including: The consistency generation model uses multi-channel time-series data prior to the voltage change boundary in the counterfactual conditional data as conditional input to extract a conditional encoding vector. Conditional encoding is a fundamental step in generating counterfactual voltage sequences, providing the model with a generation context by encoding preceding data. The encoding process employs a bidirectional LSTM (Long Short-Term Memory) network structure to sequentially process the multi-channel time-series data prior to the boundary in the counterfactual conditional data, capturing the temporal dependencies within and between channels. The hidden state vector of the last time step is transformed nonlinearly to form the conditional encoding vector. This serves as the initial state of the generative model. The conditional encoding vector contains information about the operational state before the perturbation, providing the necessary historical context for subsequent counterfactual sequence generation.

[0039] Using the conditional encoding vector as the initial hidden state, an autoregressive method generates voltage sequence sampling points for the time period following the voltage change boundary. At each generation step, the generated sampling points are concatenated with the corresponding non-voltage channel values ​​from the counterfactual conditional data as the input for the next step. Autoregressive generation is the core method for constructing counterfactual sequences, achieving continuous sequence generation through stepwise prediction. The generation process is based on a GRU (Gated Recurrent Unit) network design, using the conditional encoding vector... As the initial hidden state Then, from the voltage change boundary moment Begin by gradually generating voltage values ​​for subsequent time periods. At each time step... The model input includes the voltage value generated in the previous step. and the current non-voltage channel value Output the predicted voltage value at the current moment. and the updated hidden state The advantage of autoregressive generation is that it can maintain temporal continuity, and each prediction step takes into account the cumulative effect of historical information, which is consistent with the physical characteristics of electrical systems.

[0040] The voltage sequence sampling points generated by autoregression are concatenated with the real voltage sequence before the voltage change boundary moments in the counterfactual conditional data to form a complete counterfactual voltage sequence on the time axis. Sequence concatenation is the final step in forming a complete counterfactual scenario, constructing a complete time axis by seamlessly combining historical real data and predicted data. The concatenation process includes the boundary moments... Previous true voltage sequence Predicted voltage sequence after the boundary time generated by autoregression Connect to form a complete counterfactual voltage sequence. To ensure a smooth transition at the splicing point, a smoothing connection function is applied near the boundary time to reduce artificial boundary effects. The complete counterfactual voltage sequence represents the expected evolution of the voltage under the condition that "if this change segment is suppressed," and the difference from the true voltage sequence directly reflects the degree of causal influence of this change segment on the voltage disturbance.

[0041] In this embodiment of the invention, the intervention effect score is formed by combining the amplitude difference index and the duration difference index, including: Within a preset scoring window, the absolute value of the difference between the counterfactual voltage sequence and the true voltage sequence in the perturbation chain time series is calculated for each sampling point. The mean of the absolute values ​​of the differences within the preset scoring window is taken as the amplitude difference index. The amplitude difference index is a direct measure of the degree of voltage deviation, and the strength of the intervention effect is assessed by calculating the voltage difference between the counterfactual and real scenarios. The calculation process first determines the range of the scoring window. ,in and These represent the window widths before and after the boundary moment; then, the absolute value of the voltage difference is calculated point by point within the window, and the average value is taken as the amplitude difference index. The calculation formula is: ; in, For amplitude difference index, For the true voltage sequence at time... The value, For the counterfactual voltage sequence at time... The amplitude difference index directly reflects the degree of influence of the intervention on the voltage magnitude and is an important indicator for assessing the severity of the disturbance.

[0042] In the real voltage sequence, the number of consecutive sampling points where the voltage deviation exceeds a preset change threshold is counted, and multiplied by a uniform sampling period to obtain the duration of the voltage deviation. The duration of the counterfactual voltage deviation is calculated in the same way in the counterfactual voltage sequence, and the absolute value of the difference between the two is taken as the duration difference index. The duration difference index is a key measure for quantifying the temporal characteristics of disturbances, assessing the temporal impact of intervention effects by comparing changes in disturbance duration. The calculation process first determines the voltage deviation judgment standard, usually set as a certain percentage of the nominal voltage value (e.g., ±5%); then, within the scoring window, the interval of consecutive sampling points exceeding the deviation threshold in the real voltage sequence is identified, and the total duration is calculated. Similarly, the deviation duration in the counterfactual voltage sequence is calculated. Finally, the absolute value of the difference between the two is calculated as the duration difference index. The calculation formula is: ; in, As a duration difference indicator, and These represent the duration of deviation from the true and counterfactual voltage sequences, respectively. The duration difference index reflects the impact of the intervention on the persistence of the disturbance and is an important reference for assessing resilience.

[0043] The amplitude difference index and the duration difference index are weighted and summed according to preset weights to obtain the intervention effect score. Comprehensive scoring is a key step in the fusion of multi-dimensional indicators, balancing the relative importance of different indicators through weight allocation. The scoring process first normalizes the amplitude difference index and the duration difference index to eliminate dimensional differences; then, it is weighted and summed according to preset weight coefficients to obtain the final intervention effect score. The weight coefficients are set according to the specific application scenario and power grid characteristics, typically with a higher weight for the amplitude difference index. Within the range of 0.6-0.8, the weight of the duration difference index... The range of 0.2-0.4 reflects the reality that amplitude changes are often more critical than durations in power grid disturbance analysis. The intervention effect score provides a single comprehensive measure, facilitating direct comparison and ranking of the influence of different change segments, and providing a quantitative basis for determining the dominant causes.

[0044] In this embodiment of the invention, the leader fragment identifier that satisfies the dominant threshold with the highest intervention effect score is selected and combined with the corresponding intra-chain position marker to form the dominant leader fragment, including: If multiple intervention effect scores have the same maximum value among the lead segments to be evaluated that meet the dominance threshold, the end times of the corresponding time ranges of the lead segments to be evaluated are further compared, and the lead segment with the earliest end time is taken as the unique dominance identifier. Time priority comparison is a key strategy to resolve cases of tied scores, further distinguishing the priority of candidate segments by their temporal order. The comparison process first filters out intervention effect scores that reach the dominance threshold. The candidate set is then used to compare the end times of each segment among the highest-scoring and identical candidates. The earliest ending segment is selected as the dominant candidate. This priority principle based on the ending time reflects the temporal constraints in causal relationships; events that end earlier are more likely to be the cause of subsequent perturbations and have higher causal priority.

[0045] If a tie still exists after the above comparison, the start times of the corresponding lead segment time range to be evaluated are compared, and the lead segment identifier with the earliest start time is selected as the unique dominant identifier. Start time comparison is the final decision-making step in time priority determination, ensuring that a unique dominant candidate is selected under all conditions. If the end times are the same, the start times of the segments are further compared. The earliest segment is selected as the unique dominant identifier. This multi-level comparison mechanism establishes strict priority ranking rules, ensuring the uniqueness and certainty of the dominant leader segment identification result and avoiding decision ambiguity in the case of multiple solutions.

[0046] The unique dominant identifier, the corresponding time range of the leader segment to be evaluated, and the intra-chain position marker are combined to form the dominant leader segment. Dominant leader segment assembly is the final step in integrating the identification results, organizing the scattered information into a structured output. The assembly process combines the unique dominant identifier, the corresponding time range of the segment... This, along with the in-chain position marker (usually "preceding"), forms a complete description of the dominant initiating fragment. This description contains key information such as the identity of the dominant initiating factor, its occurrence time, and its positional relationship, providing the necessary data structure for subsequent time sequence constraint verification. In this embodiment of the invention, when the end time is earlier than the voltage change boundary time, the dominant leader segment is identified as the dominant cause object, and the dominant cause identification result is output, including: If the termination time is not earlier than the voltage change threshold, the corresponding identifier of the dominant leader fragment is excluded from the feature list of leader fragments to be evaluated. The intervention effect score comparison and dominant threshold screening are then re-executed among the remaining leader fragments to be evaluated, iteratively obtaining new dominant leader fragments until the condition that the termination time is earlier than the voltage change threshold is met. The temporal constraint verification is the core step in ensuring the rationality of causal logic. Repeated screening ensures that the finally identified trigger occurs temporally before the disturbance. The verification process first checks the termination time of the dominant leader fragment. Is it earlier than the voltage change boundary moment? If the conditions are not met, the current dominant identifier is excluded from the candidate list, and the scoring comparison and threshold screening are re-executed among the remaining candidates to obtain a new dominant leading fragment. Then, the time constraint conditions are checked again, and this process is iterated until a fragment that meets the conditions is found or all candidates are exhausted. This iterative verification method ensures that the finally identified dominant cause strictly adheres to the temporal constraints of causality, avoiding logical contradictions and incorrect attributions.

[0047] When no remaining leader fragments satisfying the evaluation criteria are found in the feature list during the iteration process, an empty dominant cause identification result is output with a "no preceding dominant cause" flag. This empty result processing is a fallback strategy to handle situations where no valid cause exists, ensuring clear output under all conditions. When all candidates are excluded during iteration, or the intervention effect scores of the remaining candidates do not reach the dominant threshold, it is determined that the current voltage disturbance lacks a clear preceding dominant cause and may be caused by external or combined factors. In this case, a special empty result is output with a "no preceding dominant cause" flag, clearly indicating the special nature of the analysis result. This fallback mechanism enhances robustness and usability, providing meaningful analysis results even in complex or atypical scenarios.

[0048] After satisfying the sequential constraints, the time range of the dominant lead fragment and the intervention effect score are extracted from the feature list of the lead fragment to be evaluated and the intervention effect score list corresponding to the dominant cause object identifier. These are then combined according to the field correspondence to form the dominant cause identification result and output. Result assembly is the final step in the analysis process, integrating scattered information into a structured output report. The assembly process extracts the time range of the corresponding fragment from the feature list of the lead fragment to be evaluated based on the dominant cause object identifier. Extract the corresponding intervention effect score from the intervention effect score list. Then, this information is organized according to a predefined structural template to form a complete dominant cause identification result. The final output includes the dominant cause object identifier (usually the device name or channel ID), the time range of occurrence, the intervention effect score, and necessary supplementary information, providing power grid operation and maintenance personnel with clear and intuitive disturbance attribution conclusions, supporting subsequent fault diagnosis and prevention measure formulation.

[0049] To support the efficient execution and result reuse of the above methods, this system also constructs a hierarchical data storage system, such as... Figure 5 As shown, it specifically includes: Raw Data Area: This area is used for persistent storage of multi-channel time-series data and device action records directly acquired from systems such as PMU, fault recorder, and SCADA. It is managed using a time-series database (such as InfluxDB) or a distributed file system, storing raw sampled values ​​and their metadata, including timestamps and channel identifiers, with high compression ratios, providing basic input for data preprocessing.

[0050] Data processing area: Serves as the working area during method execution, dynamically storing intermediate results at each stage. It mainly includes: The perturbation chain time series data formed after resampling and alignment.

[0051] The voltage change boundary moment, boundary moment reference sequence, and action indication sequence are generated by voltage change detection.

[0052] The list of features to be evaluated, generated after extraction by the TimeMixer++ encoder.

[0053] An intervention effect score list is obtained through counterfactual generation and evaluation calculation.

[0054] Data in this area is typically stored in temporary tables in relational databases (such as MySQL) or in memory caches (such as Redis) to support high-frequency read / write operations and fast access.

[0055] Results Data Area: Used to permanently store the final analysis results. Each record corresponds to a short-term voltage disturbance event, and stores the dominant causative object identifier (e.g., device ID), the time range of the dominant leading segment, the intervention effect score, and the associated intra-chain position marker in a structured manner. This area provides historical case data support for power grid fault analysis, predictive maintenance, and operation optimization, and can be managed using a relational database.

[0056] This invention achieves accurate identification of the dominant cause of voltage disturbances through multi-channel time-series data preprocessing, voltage change boundary moment identification, multi-scale feature encoding, counterfactual condition analysis, and intervention effect evaluation. The counterfactual condition modeling method of this invention can quantitatively assess the causal contribution of each change segment to the disturbance and effectively distinguish between the cause and response relationship.

[0057] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0058] It should be noted that all formulas in this manual are calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0059] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for identifying the dominant causes of short-time voltage disturbances in regional power grids based on TimeMixer, characterized in that, include: Acquire multi-channel time-series data and device action records of a single short-time voltage disturbance in the regional power grid. Calculate the maximum common sampling period according to the sampling frequency of each channel. Resample the multi-channel time-series data and device action records to the unified sampling time corresponding to the maximum common sampling period to form disturbance chain time-series data. The absolute value of the difference between adjacent sampling points is calculated for the voltage channel numerical sequence of the disturbance chain time series data, and the voltage change boundary moment is identified; The signed time difference of each sampling moment is calculated based on the voltage change boundary moment to form a boundary moment reference sequence. The device action record is converted into an action indication sequence aligned with the unified sampling moment and written into the disturbance chain timing data. Based on the boundary time reference sequence, the perturbation chain time series data is segmented into change segments for each channel. The chain position markers are assigned according to the positional relationship of the time range of each change segment relative to the voltage change boundary time. The change segments with chain position markers are input into the TimeMixer++ multi-scale hybrid encoder to obtain the leader segment features to be evaluated and form a leader segment feature list to be evaluated. The feature list of the lead segments to be evaluated is traversed one by one. The starting sampling point values ​​of the time period data corresponding to each lead segment to be evaluated are extended along the time axis to form a suppression reference value sequence. The time period data is replaced by the suppression reference value sequence and then combined with the data of other time periods to form counterfactual conditional data. The counterfactual voltage sequence is obtained by inputting it into the consistency generation model. The amplitude difference index and duration difference index are calculated within a preset scoring window before and after the voltage change boundary time. The intervention effect score is formed by combining the intervention effect score and associating it with the lead segment identifier to be evaluated to form an intervention effect score list. Based on the intervention effect score list, compare the intervention effect score with the dominant threshold, take the leader fragment identifier with the largest intervention effect score among those that meet the dominant threshold, and combine it with the corresponding intra-chain position mark to form the dominant leader fragment; The end time of the dominant leader segment is extracted and compared with the voltage change boundary time. If the end time is earlier than the voltage change boundary time, the dominant leader segment is identified as the dominant cause object, and the dominant cause identification result is output.

2. The method for identifying the dominant causes of short-term voltage disturbances in regional power grids based on TimeMixer according to claim 1, characterized in that, The step of calculating the maximum common sampling period based on the sampling frequency of each channel, and resampling the multi-channel time-series data and device action records to the unified sampling time corresponding to the maximum common sampling period, includes: Extract the sampling frequency of the time series data of each channel, calculate the greatest common divisor of the sampling frequencies of all channels, and use the sampling period corresponding to the greatest common divisor as the unified sampling period; For channels with a sampling frequency higher than the uniform sampling frequency, the average value of the original sampling points between adjacent uniform sampling times is taken as the channel value at that uniform sampling time. For channels with sampling frequencies lower than the uniform sampling frequency, the channel values ​​at the uniform sampling time are supplemented by linear interpolation between adjacent original sampling points; For device action records, the time of action occurrence is mapped to the most recent unified sampling time. At the unified sampling time and the subsequent continuous time, the action indication value is set to 1, and the rest are set to 0, forming an action indication sequence.

3. The method for identifying the dominant causes of short-term voltage disturbances in regional power grids based on TimeMixer according to claim 1, characterized in that, The step of segmenting the perturbation chain time series data into change segments for each channel based on the boundary time reference sequence includes: Identify the boundary sampling point where the signed time difference turns from negative to positive from the boundary time reference sequence, and use the sampling time before the boundary sampling point as the end time parameter of the segment before the boundary, and use the sampling time of the boundary sampling point as the start time parameter of the segment after the boundary. For each channel's time series data, the end time of the continuous, unchanged segment in the channel's numerical sequence is used as the start time parameter of the segment before the boundary, and the start time of the continuous, unchanged segment in the channel's numerical sequence is used as the end time parameter of the segment after the boundary. Each channel is divided according to the start time parameter and the end time parameter to obtain the change segment before the boundary and the change segment after the boundary. The change segment before the boundary is assigned a position in the chain and marked as the front position, and the change segment after the boundary is assigned a position in the chain and marked as the back position.

4. The method for identifying the dominant causes of short-term voltage disturbances in regional power grids based on TimeMixer according to claim 1, characterized in that, The step of inputting the changed segment with in-chain position markers into the TimeMixer++ multi-scale hybrid encoder to obtain the features of the leader segment to be evaluated includes: Each variation segment is uniformly filled to the preset maximum length according to the time axis length, and the filling value is taken as the value of the starting sampling point of each variation segment to form a sequence of variation segments of equal length. The TimeMixer++ multi-scale hybrid encoder performs temporal convolution operations on the equal-length variable segment sequence at three time scales: coarse, medium, and fine, to extract low-frequency trend components, medium-frequency fluctuation components, and high-frequency detail components, respectively.

5. The method for identifying the dominant causes of short-term voltage disturbances in regional power grids based on TimeMixer according to claim 4, characterized in that, The obtained features of the leader fragment to be evaluated include: The low-frequency trend component, mid-frequency fluctuation component, and high-frequency detail component are concatenated along the channel dimension and then compressed into a fixed-dimensional feature vector through a channel hybrid linear mapping, which serves as the feature of the leading segment to be evaluated.

6. The method for identifying the dominant causes of short-term voltage disturbances in regional power grids based on TimeMixer according to claim 1, characterized in that, The list of intervention effect scores includes: Within the preset scoring window, the absolute value of the difference between the counterfactual voltage sequence and the true voltage sequence in the disturbance chain time series data is calculated for each sampling point, and the mean of the absolute value of the difference within the preset scoring window is taken as the amplitude difference index. In the real voltage sequence, the number of consecutive sampling points where the voltage deviation exceeds a preset change threshold is counted, and the voltage deviation duration is obtained by multiplying it by a uniform sampling period. In the counterfactual voltage sequence, the counterfactual voltage deviation duration is calculated in the same way, and the absolute value of the difference between the two is taken as the duration difference index. The intervention effect score is obtained by weighting and summing the amplitude difference index and the duration difference index according to preset weights.

7. The method for identifying the dominant causes of short-term voltage disturbances in regional power grids based on TimeMixer according to claim 1, characterized in that, The input consistency generation model yields a counterfactual voltage sequence, including: The consistency generation model takes the multi-channel time series data before the voltage change boundary moment in the counterfactual conditional data as conditional input and extracts the conditional encoding vector; Using the conditional encoding vector as the initial hidden state, the voltage sequence sampling points for the period after the voltage change boundary time are generated by autoregression. At each generation step, the generated sampling points are concatenated with the non-voltage channel values ​​of the corresponding time in the counterfactual conditional data and used as the input for the next step. The sampled points of the voltage sequence generated by autoregression are spliced ​​with the real voltage sequence before the voltage change boundary time in the counterfactual conditional data to form a counterfactual voltage sequence on the complete time axis.

8. The method for identifying the dominant causes of short-term voltage disturbances in regional power grids based on TimeMixer according to claim 1, characterized in that, The process of selecting the leader fragment identifier that satisfies the dominant threshold and has the highest intervention effect score, and combining it with the corresponding intra-chain position marker to form the dominant leader fragment, includes: If multiple intervention effect scores have the same maximum value among the lead segments to be evaluated that meet the dominant threshold, then the end times of the corresponding lead segments to be evaluated are further compared, and the lead segment identifier with the earliest end time is taken as the unique dominant identifier. If there are still ties after the above comparison, then the start times of the corresponding time range of the lead segment to be evaluated are compared, and the lead segment identifier with the earliest start time is taken as the unique dominant identifier.

9. The method for identifying the dominant causes of short-time voltage disturbances in regional power grids based on TimeMixer according to claim 8, characterized in that, The constituent dominant leader fragment includes: The unique dominant identifier, the corresponding time range of the leader segment to be evaluated, and the position marker within the chain are combined to form the dominant leader segment.

10. The method for identifying the dominant causes of short-term voltage disturbances in regional power grids based on TimeMixer according to claim 1, characterized in that, When the end time is earlier than the voltage change boundary time, the dominant leader segment is identified as the dominant cause object, and the dominant cause identification result is output, including: If the end time is not earlier than the voltage change boundary time, the corresponding identifier of the dominant leader fragment is excluded from the feature list of the leader fragments to be evaluated. The intervention effect score comparison and dominant threshold screening are re-executed in the remaining leader fragments to be evaluated, and new dominant leader fragments are obtained iteratively until the condition that the end time is earlier than the voltage change boundary time is met. When there are no remaining leader fragments that meet the conditions in the feature list of leader fragments to be evaluated during the iteration process, the empty dominant cause identification result is output and a marker of no preceding dominant cause is attached. After satisfying the sequential constraints, the time range of the dominant leading fragment and the intervention effect score are extracted from the feature list of the leading fragment to be evaluated and the intervention effect score list corresponding to the dominant inducing object identifier. The dominant inducing identification result is formed by combining the fields according to the field correspondence and output.

Images