A time series data enhancement method and system based on time sequence large model fusion

Abstract

The application relates to the technical field of power system data processing, and discloses a time series data enhancement method and system based on time sequence large model fusion, which comprises the following steps: obtaining time series data to be processed; performing discretization processing on the time series data; obtaining time characteristic sequences after discretization and inputting the time characteristic sequences into a time sequence large model; obtaining a time sequence prediction value sequence corresponding to a current time sequence window output by the time sequence large model; calculating a prediction deviation between prediction value data in the time sequence prediction value sequence and original observation values of corresponding time points in the time series data; if the prediction deviation exceeds a confidence interval, marking the time series data as candidate abnormal data; and matching corresponding completion strategies for the candidate abnormal data according to abnormal types, and completing the candidate abnormal data, so that the problems of insufficient data processing accuracy, poor abnormal detection robustness, low completion accuracy and easy violation of physical constraints in the prior art are solved, and the data quality and processing efficiency are improved.

Description

Technical Field

[0001] This application relates to the field of power system data processing technology, and more specifically, to a time series data augmentation method and system based on time series large model fusion. Background Technology

[0002] In the context of new power system construction, time-series data from the power grid's source and load sides, as well as related meteorological environments, exhibit massive volume, multiple sources, and high volatility, serving as a crucial foundation for power load forecasting, renewable energy output forecasting, and dispatch operation analysis. However, in actual operation, data quality issues such as missing data, abrupt changes, and drift are frequently encountered due to factors like data acquisition equipment failures, communication interruptions, sensor drift, and extreme weather disturbances. Existing technologies for processing this type of data typically employ uniform scale modeling, which struggles to account for the multi-scale fluctuations of power grid data, ranging from minute-level to daily-level. Furthermore, fixed thresholds or static statistical methods fail to adequately consider the non-stationarity of power grid data, resulting in poor robustness in anomaly detection and a high likelihood of false alarms or missed alarms. Moreover, the anomaly detection and missing data completion processes are disconnected, lacking effective feedback optimization mechanisms. The completion strategies often employ fixed, uniform models, making it difficult to capture the complex nonlinear relationships and long-period dependencies in power grid data, leading to low completion accuracy and a tendency to violate the physical constraints of the power system. This over-reliance on human experience makes it difficult for existing technologies to achieve end-to-end automated processing, resulting in insufficient generalization ability and processing efficiency, and failing to provide highly reliable data support for power system forecasting and decision-making.

[0003] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application provides a time series data augmentation method and system based on large-scale time series model fusion, which can solve the quality problems of power grid time series data, including data missing, abrupt changes, and drift, as well as the problems of insufficient precision, poor robustness, lack of feedback optimization mechanism, low completion accuracy and easy violation of physical constraints in the processing of such data by existing technologies.

[0005] In a first aspect, this application provides a time series data augmentation method based on large-scale time series model fusion for processing time series data of power grids. The method includes: Obtain the time series data to be processed, and discretize the time series data to obtain the discretized time feature series; The time feature sequence is input into the time series large model to obtain the time series prediction value sequence output by the time series large model corresponding to the current time series window; Calculate the prediction deviation between the predicted values ​​in the time series and the original observations at the corresponding time points in the time series data; Obtain the historical prediction deviation sequence within the current time series window, construct a confidence interval based on the historical prediction deviation sequence, and mark the time series data as candidate anomalous data if the prediction deviation exceeds the confidence interval; The candidate abnormal data is decomposed into different abnormal types. Based on the abnormal type, a corresponding completion strategy is matched for the candidate abnormal data. The candidate abnormal data is then completed based on the completion strategy to obtain the complete time series sequence.

[0006] Furthermore, after the step of acquiring the time series data to be processed, and before the step of discretizing the time series data, the method includes: Determine whether there is a time series bias in the data that is related to changes in ambient temperature; If the time series deviation is determined to exist, the time series data should be corrected. The corrected time series data is reconstructed using signals across multiple time ranges, and the reconstructed time series data is output as the basis for discretization processing.

[0007] Furthermore, the step of determining whether there is a time series deviation in the time series data that is associated with changes in ambient temperature includes: Monitor the rate of change of ambient temperature; When the rate of change of the ambient temperature exceeds the first preset threshold, the irradiance deviation is predicted based on the preset correlation function between temperature and time deviation. If the difference between the observed irradiance value and the theoretical irradiance value in the time series data is consistent with the irradiance deviation, then it is determined that there is a time series deviation in the time series data that is related to changes in ambient temperature.

[0008] Furthermore, if the time series deviation is determined to exist, the steps for correcting the time series data include: Calculate the correction factor corresponding to the rate of change of the ambient temperature based on the real-time ambient temperature. The time series data is corrected using this correction factor to eliminate the time series bias.

[0009] Furthermore, the steps of reconstructing signals from multiple time ranges into the corrected time series data and outputting the reconstructed time series data as the basis for discretization include: Acquire real-time meteorological observation data within the current time window, including cloud cover change rate and local severe convective weather warning indicators; Based on the real-time meteorological observation data, the process noise covariance and measurement noise covariance parameters of the Kalman filter are adjusted. Specifically, when the cloud cover change rate exceeds the second preset threshold or the local severe convective weather warning sign indicates the existence of a local severe convective weather warning, the process noise covariance is increased while the measurement noise covariance is decreased. The corrected time series data is subjected to Kalman filtering for scaling. Specifically, within a preset minute-level time range, the corrected time series data is subjected to point-by-point prediction and updating using Kalman filtering. Within a preset hour-level time range, the updated time series data is subjected to secondary smoothing to obtain the reconstructed time series data.

[0010] As an optional approach, the steps for constructing confidence intervals based on this historical prediction deviation sequence include: Calculate the center value of this historical prediction deviation sequence; Based on the central value, the dynamic standard deviation is increased and decreased respectively to obtain the upper and lower boundaries of the dynamic confidence interval. The dynamic standard deviation is dynamically adjusted according to external scenario information, which includes weather change warnings, holiday markers, or sudden changes in new energy output. When at least one of the weather change warnings, holiday markers, or sudden changes in new energy output is detected, the dynamic standard deviation is increased by a preset multiple to widen the confidence interval range.

[0011] Furthermore, the steps of performing causal decomposition on the candidate anomalous data and classifying it into different anomalous types include: Extract the multidimensional features of the candidate anomaly data, which include: the absolute and relative values ​​of the prediction bias, the mean and variance of the data before and after the anomaly point, the rate of change of adjacent data points, and the duration of the bias. The extracted multidimensional features are input into a pre-trained anomaly type classifier; The anomaly type classifier outputs the anomaly type determination result and generates a labeled time series sequence with anomaly type labels, where the anomaly type labels include isolated anomalies, mutation anomalies, drift anomalies, and continuous defect anomalies.

[0012] Based on this, according to the anomaly type, the steps of matching the corresponding completion strategy for the candidate anomaly data, and completing the candidate anomaly data according to the completion strategy to obtain the completed time series sequence include: Based on the anomaly type marker in the time sequence, a corresponding completion strategy is matched from the preset completion strategy library. The completion strategy includes the historical context window length and constraint strength used for completion. A preliminary completion value is generated based on the length of the historical context window; Based on the power grid physical constraint rules corresponding to the constraint strength, the preliminary supplementary value is corrected to obtain the corrected final supplementary value. The power grid physical constraint rules include the upper and lower limits of load peak and valley periods, the new energy output not less than zero and not greater than the installed capacity, and the synchronicity constraint between photovoltaic output and solar irradiance. The candidate anomaly data is completed based on the final completion value to obtain the complete time series sequence.

[0013] Preferably, after obtaining the completed time series sequence, the method includes: The complete time series sequence is included in the model training dataset of the large time series model. Combined with the anomaly type and the real time series data within the current time series window, the large time series model is optimized.

[0014] Secondly, this application also discloses a time series data augmentation system based on large-scale time series model fusion, used to perform the aforementioned method, the system comprising: The data acquisition and processing module is used to acquire time series data to be processed, and to discretize the time series data to obtain a discretized time feature sequence. The time series prediction module is used to input the time feature sequence into the time series large model and obtain the time series prediction value sequence output by the time series large model corresponding to the current time series window; The deviation calculation module is used to calculate the prediction deviation between the predicted value data in the time series and the original observation value at the corresponding time point in the time series data; The abnormal data marking module is used to obtain the historical prediction deviation sequence within the current time series window, construct a confidence interval based on the historical prediction deviation sequence, and mark the time series data as candidate abnormal data if the prediction deviation exceeds the confidence interval. The data completion module is used to perform causal decomposition on the candidate abnormal data, classify it into different abnormal types, match the corresponding completion strategy for the candidate abnormal data according to the abnormal type, and complete the candidate abnormal data based on the completion strategy to obtain the complete time series sequence.

[0015] This application discloses a time series data augmentation method based on the fusion of a large time series model. By introducing a large time series model for prediction and constructing a dynamic confidence interval based on prediction bias for anomaly detection, it effectively solves the problems of poor robustness and susceptibility to false alarms or missed alarms in existing technologies. Furthermore, this application performs causal decomposition on candidate anomaly data, classifying it into different anomaly types, and matches customized completion strategies according to the anomaly type. Simultaneously, it corrects the completed values ​​by incorporating power grid physical constraint rules, overcoming the shortcomings of low completion accuracy and easy violation of power system physical constraints in existing technologies. By incorporating the completed time series into the training dataset of the large time series model for optimization, a closed-loop feedback mechanism of data augmentation and model optimization is formed, effectively solving the problem of the lack of an effective feedback optimization mechanism in existing technologies. This significantly improves the overall effect, generalization ability, and processing efficiency of data augmentation, providing highly reliable data support for power system prediction and decision-making. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a time series data augmentation method based on large time series model fusion, provided as an embodiment of this application.

[0017] Figure 2 This is a schematic diagram of the structure of a time series data augmentation system based on time series large model fusion, provided in an embodiment of this application.

[0018] Labeling Explanation: 210, Data Acquisition and Processing Module; 220, Time Series Prediction Module; 230, Deviation Calculation Module; 240, Abnormal Data Marking Module; 250, Data Completion Module. Detailed Implementation

[0019] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0020] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0021] In the context of new power system construction, time-series data from the power grid's source and load sides, as well as related meteorological environments, exhibit massive, multi-source, and highly volatile characteristics, serving as a crucial foundation for power load forecasting, renewable energy output forecasting, and dispatch operation analysis. However, in actual operation, due to various factors such as data acquisition equipment failures, communication interruptions, sensor drift, and extreme weather disturbances, power grid time-series data often suffers from quality issues such as missing data, abrupt changes, and drift. Existing technologies for processing this type of power grid time-series data typically employ uniform-scale modeling, making it difficult to account for the multi-scale fluctuations of power grid data ranging from minute-level to daily-level, resulting in insufficient data processing precision. Furthermore, existing methods often use fixed thresholds or static statistical methods for anomaly detection, failing to fully consider the non-stationarity of power grid data, leading to poor robustness in anomaly detection and a high likelihood of false alarms or missed alarms, impacting subsequent analysis and decision-making. Moreover, the anomaly detection and missing data completion processes in existing technologies are often fragmented, lacking effective feedback optimization mechanisms, resulting in poor overall data augmentation effects. Completion strategies typically employ fixed, uniform models, making it difficult to capture the complex nonlinear relationships and long-period dependencies in power grid data. This results in low completion accuracy and a tendency to violate the physical constraints of the power system. This over-reliance on human experience makes it difficult for existing technologies to achieve end-to-end automated processing, leading to insufficient generalization ability and processing efficiency, and failing to provide highly reliable data support for power system forecasting and decision-making.

[0022] Regarding this, firstly, see... Figure 1 This application proposes a time series data augmentation method based on large-scale time series model fusion for processing power grid time series data. The method includes: S1. Obtain the time series data to be processed, and discretize the time series data to obtain the discretized time feature series; S2. Input the time feature sequence into the time series large model to obtain the time series prediction value sequence output by the time series large model corresponding to the current time series window; S3. Calculate the prediction deviation between the predicted values ​​in the time series forecast value sequence and the original observations at the corresponding time points in the time series data; S4. Obtain the historical prediction deviation sequence within the current time series window, construct a confidence interval based on the historical prediction deviation sequence, and mark the time series data as candidate outlier data if the prediction deviation exceeds the confidence interval. S5. Perform causal decomposition on the candidate abnormal data and divide it into different abnormal types. Based on the abnormal type, match the corresponding completion strategy for the candidate abnormal data. Based on the completion strategy, complete the candidate abnormal data to obtain the complete time series sequence.

[0023] Time series data refers to a series of data points arranged in chronological order, such as load data in a power grid, renewable energy output data, and meteorological data. A large-scale time series model refers to a deep learning model capable of processing and analyzing large-scale time series data, such as a model based on the Transformer architecture. A time series window refers to a continuous subsequence within time series data, used as model input or for analysis. For example, a one-hour time series window might contain 60 minute-level data points.

[0024] Prediction bias refers to the difference between model predictions and actual observations, and is an important indicator of model prediction accuracy. Confidence intervals are used to determine whether the prediction bias is within an acceptable range, thereby identifying outlier data. Candidate outlier data refers to data points initially identified as anomalous and requiring further analysis and processing. Causal decomposition refers to the process of analyzing outlier data to determine its causes; for example, anomalies may be caused by sensor malfunctions, communication interruptions, or extreme weather.

[0025] Imputation strategies refer to methods used to repair or fill in missing or anomalous data; different types of anomalous data may require different imputation strategies. Complete time series data refers to time series data whose quality is improved after anomaly detection and data imputation.

[0026] The time series data augmentation method based on large time series model fusion proposed in this application focuses on refining the processing of power grid time series data to improve data quality and reliability.

[0027] First, the time series data to be processed is acquired and discretized to obtain a discretized time feature sequence. For example, raw minute-level or hourly time series data such as grid load, photovoltaic output, wind power output, and ambient temperature and irradiance can be obtained from data sources such as power grid SCADA systems and meteorological monitoring stations. Several methods can be used to discretize time series data. One method is to convert continuous time series data into discrete values ​​or categories. For example, a 24-hour day can be divided into different time periods (such as peak hours, valley hours, and normal hours), and each time period can be assigned a discrete label. Another method is to extract the periodic features of the time series data. For example, hourly, daily, weekly, monthly, and yearly time periodic features can be extracted and used as discretized time features. Alternatively, Fourier transform or wavelet transform methods can be used to decompose the time series data into components of different frequencies, and specific frequency components can be selected as discretized time features.

[0028] Subsequently, the time feature sequence is input into the time series large-scale model, yielding the time series predicted value sequence output by the model corresponding to the current time series window. The time series large-scale model can employ various advanced deep learning architectures, such as Transformer-based models, recurrent neural networks (RNNs), or long short-term memory networks (LSTMs). These models can capture complex patterns and long-term dependencies in time series data. For example, a Transformer-based time series large-scale model can be constructed, where the input layer receives the discretized time feature sequence, encodes the sequence through a multi-head self-attention mechanism and a feedforward network, and finally outputs the time series predicted value sequence corresponding to the current time series window. The time series predicted value sequence can be a predicted value at a future point in time or a predicted value over a period of time in the future.

[0029] Next, the prediction bias is calculated between the predicted values ​​in the time series and the original observations at the corresponding time points. There are several ways to calculate the prediction bias. For example, absolute error, which is the absolute value of the difference between the predicted value and the original observation, can be used. Relative error, which is the ratio of the absolute error to the original observation, can also be used. Statistical metrics such as mean squared error (MSE) or root mean square error (RMSE) can also be used to measure the prediction bias. For instance, for each time point, the predicted value output by the large time series model is compared with the actual observation value at that time point, and the difference between the two is calculated as the prediction bias.

[0030] Furthermore, the historical prediction deviation sequence within the current time series window is obtained. A confidence interval is constructed based on this historical prediction deviation sequence. If the prediction deviation exceeds the confidence interval, the time series data is marked as candidate outlier data. The historical prediction deviation sequence can contain the set of all prediction deviations over a past period. There are several ways to construct the confidence interval. One method is to use statistical methods, such as calculating the mean and standard deviation of the historical prediction deviation sequence, and then constructing a confidence interval based on the mean plus or minus a certain multiple of the standard deviation, according to the assumption of normal distribution. Another method is to use non-parametric methods, such as constructing the confidence interval based on the quantiles of the historical prediction deviation sequence, for example, using the 5th and 95th percentiles of the historical prediction deviation sequence as the lower and upper limits of the confidence interval. When the current prediction deviation falls outside the constructed confidence interval, the data point is considered outlier data and marked as candidate outlier data.

[0031] Finally, causal decomposition can employ various methods. For example, rule-based causal decomposition can be used; if an anomaly occurs during sensor maintenance, it might be attributed to equipment failure. Machine learning methods can also be used; for instance, a classification model can be trained to take the features of candidate anomaly data (such as anomaly duration, anomaly magnitude, related meteorological data, etc.) as input and output anomaly type. Anomaly types can include isolated anomalies (single data point anomaly), abrupt anomalies (sudden and significant changes in data values), drift anomalies (slow deviations from the normal range in data values), and continuous loss anomalies (multiple consecutive missing data points). Different completion strategies can be matched to different anomaly types. For example, linear interpolation or mean imputation can be used for isolated anomalies; historical similarity-based completion methods can be used for abrupt anomalies; trend prediction or regression models can be used for drift anomalies; and prediction-based completion based on large time-series models can be used for continuous loss anomalies. Completion strategies can include parameters such as the length of the historical context window and the strength of constraints used for completion. Based on the matched completion strategy, the candidate anomaly data is completed, ultimately yielding the completed time-series sequence.

[0032] This application improves prediction accuracy by introducing a large-scale time-series model for prediction, which captures complex nonlinear relationships and long-period dependencies in power grid data. Based on the prediction bias of the large-scale time-series model, combined with dynamic confidence intervals for anomaly detection, it better adapts to the non-stationarity of power grid data, significantly improving the robustness and accuracy of anomaly detection and effectively reducing false alarms and missed alarms. More importantly, this application tightly integrates anomaly detection with the data completion process. By performing causal decomposition on candidate anomaly data and matching customized completion strategies, it achieves accurate repair of different types of anomaly data, effectively avoiding the problems of low completion accuracy and easy violation of physical constraints in traditional methods. This end-to-end automated processing method not only improves the efficiency of data augmentation but also provides highly reliable data support for power system prediction and decision-making.

[0033] Furthermore, after acquiring the time series data to be processed, and before the step of discretizing the time series data, the following steps are included: Determine whether there is a time series bias in the time series data that is related to changes in ambient temperature; If a time series deviation is determined to exist, the time series data is corrected. The corrected time series data is reconstructed using signals across multiple time ranges, and the reconstructed time series data is output as the basis for discretization processing.

[0034] Specifically, determining whether there are time series biases in time series data related to changes in ambient temperature involves analyzing the potential relationship between the characteristics of the time series data and changes in ambient temperature to identify data shifts or anomalies caused by temperature fluctuations. For example, in power grid data, the performance of certain equipment or sensor readings may change linearly or non-linearly with ambient temperature, leading to systematic biases in the data.

[0035] In this context, correcting time series data to address temporal biases involves using appropriate algorithms or models to eliminate or mitigate these biases caused by changes in ambient temperature. In practical applications, this involves reconstructing the corrected time series data across multiple time ranges and outputting the reconstructed data as the basis for discretization. This process involves processing and optimizing the data at different time scales to extract more stable and representative features. Multi-time-range signal reconstruction helps smooth noise, fill in missing values, and reveal data patterns at different time granularities, providing high-quality input for subsequent discretization.

[0036] This application effectively addresses the potential environmental temperature-related time-series bias in raw time-series data by introducing a preprocessing step before discretization. First, by determining whether a time-series bias exists related to environmental temperature changes, data anomalies influenced by external factors can be accurately identified. Once such biases are identified, they are immediately corrected, thus eliminating the interference of temperature fluctuations on data accuracy. Subsequently, signal reconstruction across multiple time ranges further enhances the stability and reliability of the data, enabling discretization to be based on cleaner and more representative data. This avoids error accumulation caused by poor raw data quality, laying a solid foundation for subsequent large-scale time-series model prediction and anomaly detection.

[0037] Through the above technical solutions, this application can significantly improve the quality and reliability of time series data. By identifying and correcting time series deviations associated with changes in ambient temperature, it avoids directly inputting erroneous data into subsequent processing stages, thereby improving the accuracy of discretization processing. Furthermore, signal reconstruction across multiple time ranges optimizes the data at different time scales, enhancing its robustness and feature representation capabilities. This not only provides high-quality input for accurate prediction of large-scale time series models but also makes anomaly detection more sensitive and accurate, effectively avoiding data misjudgment caused by environmental factors, thus improving the overall performance and application value of the entire time series data augmentation method.

[0038] In some preferred embodiments, it is assumed that in a photovoltaic power station on the power grid, the irradiance sensor data is frequently affected by drastic changes in ambient temperature, leading to a deviation between observed and theoretical values. After acquiring the irradiance time series data to be processed, the first step is to determine whether there is a time series deviation in these data related to changes in ambient temperature. Specifically, the rate of change of ambient temperature can be monitored. For example, when the rate of temperature change exceeds a certain preset threshold (e.g., 5 degrees Celsius per hour), the potential amount of irradiance deviation is predicted based on a preset correlation function between temperature and irradiance deviation, such as an empirical formula or machine learning model. If the difference between the observed irradiance and the theoretical irradiance at this time is consistent with the predicted deviation, then a time series deviation is determined to exist.

[0039] Once a time-series bias is identified, a correction factor is calculated based on the real-time ambient temperature. For example, if an increase in temperature causes an overestimation of irradiance reading, the correction factor might be negative, used to adjust the reading downwards. This correction factor is then used to correct the original irradiance time-series data point-by-point to eliminate systematic biases caused by temperature.

[0040] After correction, to further improve the smoothness and reliability of the data, the corrected irradiance data undergoes signal reconstruction across multiple time ranges. For example, Kalman filtering can be used to predict and update the corrected time series data point-by-point over a minute-level time range to smooth high-frequency noise; a second smoothing process is then performed on the updated time series data over an hour-level time range to capture more stable trends. Finally, the corrected and reconstructed irradiance time series data is output as the basis for subsequent discretization processing, ensuring the accuracy and stability of the input data and thus improving the effectiveness of the entire data augmentation process.

[0041] Specifically, the steps to determine whether there is a time series bias in the data that is associated with changes in ambient temperature include: Monitor the rate of change of ambient temperature; When the rate of change of ambient temperature exceeds the first preset threshold, the irradiance deviation is predicted based on the preset correlation function between temperature and time deviation. If the difference between the observed and theoretical irradiance values ​​in the time series data is consistent with the irradiance deviation, then it is determined that there is a time series deviation in the time series data that is related to changes in ambient temperature.

[0042] The rate of change of ambient temperature refers to the magnitude of increase or decrease in ambient temperature per unit time, which can be calculated using continuous ambient temperature measurements. For example, the temperature change per hour or per minute can be calculated. The first preset threshold is an empirical value or a critical value derived from historical data analysis to define whether the change in ambient temperature is significant. Its setting aims to filter out minor daily temperature fluctuations and focus only on drastic temperature changes that may cause significant time-series deviations.

[0043] Furthermore, the pre-defined correlation function between temperature and time-series deviation can be a mathematical model, such as a linear regression model, a multinomial model, or a machine learning-based model. This function, trained on historical data, can describe the quantitative relationship between changes in ambient temperature and potential time-series deviations in power grid time-series data (especially irradiance-related indicators). The irradiance deviation is the theoretical irradiance deviation value caused by temperature changes, predicted based on this correlation function and the current rate of change in ambient temperature.

[0044] Specifically, observed irradiance values ​​refer to real-time irradiance data directly acquired from sensors, while theoretical irradiance values ​​refer to the expected irradiance values ​​calculated based on factors such as geographical location, date, and time under ideal or standard environmental conditions. By comparing the difference between observed and theoretical irradiance values, and then comparing it with the predicted irradiance deviation, it is possible to accurately determine whether the deviation in the current time series data is strongly correlated with changes in ambient temperature.

[0045] Specifically, if the difference between the observed and theoretical irradiance values ​​in the time series data is consistent with the irradiance deviation, it means that the absolute difference between the difference and the irradiance deviation is less than a preset error tolerance threshold, or the difference falls within an interval centered on the irradiance deviation and with a preset error range as the radius. Specifically, a relative error threshold (e.g., the percentage deviation of the difference relative to the irradiance deviation is less than 5%) or an absolute error threshold (e.g., the absolute value of the difference is less than 0.1) can be set. Alternatively, a statistical hypothesis test can be used. When the difference between the difference and the irradiance deviation is not statistically significant, for example, if the p-value of the t-test is greater than 0.05, they are considered consistent. This method accurately identifies time series deviations caused by changes in ambient temperature, avoiding misclassification of other noise or anomalies as temperature-related deviations.

[0046] This application identifies drastic temperature changes that may lead to time-series bias by monitoring the rate of change of ambient temperature and comparing it with a preset first threshold. Based on this, a preset correlation function between temperature and time-series bias is used to quantify and predict the amount of irradiance bias caused by temperature changes. By comparing the observed irradiance bias with the predicted irradiance bias, this scheme effectively distinguishes time-series bias caused by ambient temperature changes from other types of bias. This mechanism ensures that subsequent correction and reconstruction processes can specifically address the effects of temperature, improving the accuracy of data processing.

[0047] Furthermore, if a time series deviation is determined to exist, the steps for correcting the time series data include: Calculate the correction factor corresponding to the rate of change of ambient temperature based on the real-time ambient temperature. Time series data are corrected using correction factors to eliminate time series biases.

[0048] Specifically, real-time ambient temperature refers to the actual ambient temperature value obtained by devices such as temperature sensors at the current moment of data acquisition or processing. The rate of change of ambient temperature can be understood as how quickly the real-time ambient temperature changes over time; for example, it can be obtained by calculating the ratio of the temperature difference at consecutive time points to the time interval. A correction factor is a numerical value used to adjust time series data; its magnitude and direction are directly related to the rate of change of ambient temperature.

[0049] As a preferred implementation, the correction factor can be determined based on a pre-established mapping relationship between the rate of temperature change and the time-series deviation. For example, a function or lookup table can be constructed using historical data analysis, physical model simulation, or machine learning methods. The input is the current rate of change of the ambient temperature, and the corresponding correction factor can be output. The correction factor can be an additive correction or a multiplicative correction coefficient, depending on the nature of the time-series deviation.

[0050] In practical applications, using correction factors to correct time series data involves applying the calculated correction factor to the observed values ​​in the time series data. For example, if the correction factor is an additive correction, it can be directly added to or subtracted from the original observed values; if the correction factor is a multiplicative correction coefficient, it can be multiplied by the original observed values. The purpose is to offset or reduce the time series bias caused by the rate of change of ambient temperature, thereby making the time series data closer to the true values.

[0051] Through the above technical solution, this application can achieve refined and dynamic correction of time-series biases related to changes in ambient temperature in time-series data. By introducing the rate of change in ambient temperature as the determining factor of the correction, the accuracy and real-time performance of the correction are significantly improved. This effectively eliminates transient biases caused by rapid changes in ambient temperature, making the corrected time-series data more realistic and reliable. It provides high-quality input for subsequent large-scale time-series model prediction, anomaly detection, and data completion, thereby improving the robustness and accuracy of the entire data augmentation method.

[0052] First, deploy one or more suitable temperature sensors, such as thermistors, platinum resistance thermometers (RTDs), thermocouples, etc., and connect these sensors to a data acquisition system, such as a microcontroller, data logger, or industrial control system, to continuously collect ambient temperature data at a preset sampling frequency (e.g., per second, per minute). After acquiring continuous temperature data in real time, assume that the temperature collected at time point t_{i-1} is T_{i-1}, and the temperature collected at the next time point t_i is T_i. Then, the time interval Deltat between these two time points is t_i-t_{i-1}, and the temperature change DeltaT is T_i-T_{i-1}.

[0053] The rate of change of ambient temperature (dT / dt) can be approximated as: dT / dt=DeltaT / Deltat=(T_i-T_{i-1}) / (t_i-t_{i-1}) This calculation continues, generating a corresponding instantaneous rate of change for each new temperature reading. In practical applications, to reduce the impact of sensor noise on the rate calculation, preprocessing methods such as moving average or low-pass filtering are sometimes used to smooth the temperature data before calculating the rate of change.

[0054] Once the rate of change of ambient temperature is obtained, a corresponding correction factor can be generated according to pre-defined rules or models. The correction factor is a function that takes the rate of temperature change as input and outputs the correction value. If experience indicates an approximately linear relationship between the rate of temperature change and the required correction, a linear function can be used to define the correction factor. For example, correction factor = k*(dT / dt) + C, where k is a proportionality coefficient and C is a constant offset. These parameters are typically determined through experimental calibration or regression analysis to ensure that the corrected data is as close as possible to the true value.

[0055] In some cases, the effect of the rate of temperature change on data bias may not be linear, such as exhibiting a quadratic, exponential, or other more complex nonlinear relationship. In such cases, a polynomial function, exponential function, or other custom-defined nonlinear function can be used to model the correction factor.

[0056] When the relationship between the rate of temperature change and the correction factor is complex and difficult to express with a simple mathematical formula, a lookup table is pre-established through extensive experiments and tests at different rates of temperature change. This table maps a series of discrete rate of temperature change values ​​to their corresponding correction factor values. During runtime, the appropriate correction factor is obtained by interpolation or direct lookup in the lookup table based on the currently calculated rate of temperature change.

[0057] Furthermore, the steps of reconstructing signals from multiple time ranges into the corrected time series data and outputting the reconstructed time series data as the basis for discretization include: Acquire real-time meteorological observation data within the current time window, including cloud cover change rate and local severe convective weather warning indicators; Based on real-time meteorological observation data, the process noise covariance and measurement noise covariance parameters of the Kalman filter are adjusted. Specifically, when the cloud cover change rate exceeds the second preset threshold or the local severe convective weather warning sign indicates the existence of a local severe convective weather warning, the process noise covariance is increased while the measurement noise covariance is decreased. The corrected time series data is subjected to Kalman filtering for scaling. Specifically, within a preset minute-level time range, the corrected time series data is subjected to point-by-point prediction and updating using Kalman filtering. Within a preset hour-level time range, the updated time series data is subjected to secondary smoothing to obtain the reconstructed time series data.

[0058] Specifically, real-time meteorological observation data refers to meteorological parameters directly related to the power grid operating environment obtained through meteorological monitoring systems or forecasting services within the current time window. These parameters may include cloud cover change rate, precipitation intensity, wind speed, and local severe convective weather warning indicators. Among these, cloud cover change rate reflects the rapid changing trend of solar irradiance, which has a direct impact on photovoltaic power output; precipitation intensity and wind speed affect the operating status of wind power generation and transmission lines; and local severe convective weather warning indicators indicate possible extreme weather events, which are usually accompanied by drastic fluctuations in power grid data.

[0059] The Kalman filter is a highly efficient recursive filter used to estimate the state of dynamic systems. It balances the confidence level between model predictions and actual measurements using process noise covariance and measurement noise covariance parameters. Process noise covariance reflects the uncertainty of the system model or the impact of external disturbances on the system state, while measurement noise covariance reflects the uncertainty or error of sensor measurement data. In this application, these parameters are dynamically adjusted based on real-time meteorological observation data to enable the Kalman filter to adaptively respond to different environmental conditions. For example, when the rate of cloud cover change exceeds a second preset threshold or a local severe convective weather warning indicates the presence of such a warning, increasing the process noise covariance means that the system model's confidence in its own predictions decreases, suggesting that the system state may undergo drastic changes and requires more reliance on new observation data for correction. Conversely, decreasing the measurement noise covariance indicates that under these specific conditions, despite environmental complexity, newly acquired measurement data is given higher weight to quickly capture drastic fluctuations in the actual state.

[0060] Based on the volatility of real-time meteorological observation data and the event warning status, the values ​​of process noise covariance (denoted as Q) and measurement noise covariance (denoted as R) are dynamically adjusted so that the filter has more confidence in model predictions when the environment is stable and more confidence in real-time observation data when the environment changes drastically.

[0061] The specific adjustment strategy is as follows: Pre-define the baseline values ​​for the process noise covariance (Q_base) and the measurement noise covariance (R_base). These baseline values ​​can be obtained through offline calibration, for example, by identifying Kalman filter parameters on historical data under typical stable weather conditions to minimize filtering error.

[0062] Based on real-time meteorological observation data, a dimensionless meteorological disturbance intensity index S is calculated, with a value range of [0,1]. When the cloud cover change rate (unit: % / minute) is lower than the second preset threshold (e.g., 5% / minute) and there is no severe convective weather warning, S is set to 0; when the cloud cover change rate reaches or exceeds the second preset threshold, S increases linearly or exponentially with the cloud cover change rate, with a maximum of 1; when there is a local severe convective weather warning, S is directly set to 1. In addition, other meteorological parameters (such as precipitation intensity and wind speed change rate) can be included as a weighted supplement.

[0063] Based on the disturbance intensity index S, update Q and R according to the following rules: The process noise covariance is adjusted as follows: Q = Q_base × (1 + α × S), where α is a preset positive adjustment coefficient (usually a value greater than 1, such as 2~10). When S increases, Q increases accordingly, indicating that the model prediction becomes unreliable, and the filter should reduce its dependence on the model prediction.

[0064] The measurement noise covariance is adjusted as follows: R = R_base × (1 - β × S), where β is a preset positive adjustment coefficient (usually a value greater than 0 and less than 1, such as 0.5~0.9). As S increases, R decreases accordingly, indicating that the reliability of the real-time observation data is relatively improved, and the filter should place more trust in the current measurement value. It is necessary to ensure that the minimum value of R is not lower than a positive lower limit (e.g., 0.1 times R_base) to avoid filter divergence.

[0065] When Q and R are both diagonal matrices (assuming each state variable is independent) or full matrices, the above scalar multiplication coefficients can be applied to the diagonal elements of the matrix or the entire matrix (i.e., global scaling). For off-diagonal elements, the same scaling factor can be used to maintain the relative structure of the matrix.

[0066] Scaling refers to breaking down the signal reconstruction process into processing steps at different time granularities. Within a preset minute-level time range, Kalman filtering is used to predict and update the corrected time series data point by point. The aim is to capture subtle changes and rapid fluctuations in the data at short time scales, ensuring responsiveness to instantaneous anomalies or rapid changes. Within a preset hour-level time range, secondary smoothing is performed on the updated time series data. The purpose is to further eliminate high-frequency noise and smooth the data, building upon the minute-level processing, to obtain more stable and representative reconstructed data that reflects trends and patterns over longer time scales. Secondary smoothing can employ various methods such as moving averages and exponential smoothing, aiming to improve the overall quality and usability of the data.

[0067] Furthermore, the steps for constructing confidence intervals based on historical prediction bias sequences include: Calculate the center value of the historical prediction deviation sequence; Using the center value as a benchmark, the dynamic standard deviation is increased and decreased respectively to obtain the upper and lower boundaries of the dynamic confidence interval. The dynamic standard deviation is dynamically adjusted according to external scenario information, which includes weather change warnings, holiday markers, or sudden changes in new energy output. When at least one of the weather change warnings, holiday markers, or sudden changes in new energy output is detected, the dynamic standard deviation is increased by a preset multiple to widen the confidence interval range.

[0068] Specifically, the central value of a historical forecast deviation sequence can be understood as the statistical mean or median of the sequence, used to characterize the central tendency of historical forecast deviations. For example, methods such as moving averages or exponentially weighted averages can be used to calculate this central value to reflect the recent trend of deviations. The dynamic standard deviation is a variable parameter used to determine the width of the confidence interval. Unlike the traditional fixed standard deviation, the dynamic standard deviation can be adjusted in real time according to changes in the external environment, allowing the confidence interval to more flexibly adapt to different operating scenarios. External scenario information refers to non-technical factors that affect the normal fluctuations of power grid time series data, including weather change warnings, holiday markers, or sudden changes in renewable energy output.

[0069] Weather abrupt change warnings refer to extreme weather warnings issued by meteorological departments, such as heavy rain, strong winds, and hail. These weather events can cause drastic fluctuations in grid load or renewable energy output. Holiday markers refer to national statutory holidays or specific rest days. These dates are usually accompanied by significant changes in social activity patterns, thus affecting the grid load curve. Sudden changes in renewable energy output refer to large, unexpected changes in the power generation of renewable energy sources such as photovoltaics and wind power within a short period. This may be caused by sudden weather changes such as cloud cover, sudden drops in wind speed, or equipment failure. In practical applications, when at least one of the following events—weather abrupt change warnings, holiday markers, or sudden changes in renewable energy output—is detected, an adjustment mechanism for the dynamic standard deviation is triggered. Specifically, the dynamic standard deviation is increased by a preset multiple, such as 1.5 times, 2 times, or higher, to widen the confidence interval. When the external environment changes drastically, a larger prediction deviation is allowed to be considered normal fluctuation, thus avoiding misjudging normal fluctuations caused by external events as abnormal data.

[0070] Specifically, the steps of causal decomposition of candidate outlier data and classification into different outlier types include: Extract multidimensional features from candidate outlier data. These features include: the absolute and relative values ​​of prediction bias, the mean and variance of data before and after the outlier, the rate of change of adjacent data points, and the duration of the bias. The extracted multidimensional features are input into a pre-trained anomaly type classifier; The anomaly type classifier outputs the anomaly type determination result and generates a labeled time series sequence with anomaly type labels, where the anomaly type labels include isolated anomalies, mutation anomalies, drift anomalies, and continuous defect anomalies.

[0071] Among them, the extraction of multidimensional features aims to quantify the characteristics of candidate anomaly data from different dimensions. Specifically, the absolute and relative values ​​of prediction bias are used to measure the severity of the anomaly; the mean and variance of the data before and after the anomaly point can reflect the local changes in the data distribution when the anomaly occurs; the rate of change of adjacent data points helps to capture the suddenness or gradualness of the anomaly; and the duration of the bias is used to judge the persistence of the anomaly.

[0072] Extracting multidimensional features from candidate anomaly data refers to obtaining various numerical information reflecting the nature and causes of the anomaly. Specifically, this can be achieved in the following ways: the absolute and relative values ​​of the prediction deviation can be obtained by calculating the difference between the predicted and actual observed values ​​and their ratio; the mean and variance of the data before and after the anomaly point can be obtained by statistically calculating the data within a fixed-length time window before and after the anomaly point; where an anomaly point refers to a single data moment or a representative location point in a continuous data segment that deviates from the normal prediction range in the candidate anomaly data. When calculating the mean and variance of the data before and after the anomaly point, one or more anomalies need to be identified from the candidate anomaly data. For example, the moment with the largest absolute value of the prediction deviation in the candidate anomaly data can be taken as the anomaly point; or the time point when the deviation first exceeds a preset threshold can be taken as the anomaly point; or for continuous anomaly segments, their geometric center or deviation peak point can be taken as the anomaly point. The length of the data window before and after the anomaly point can be preset according to the data type (e.g., 5 data points before and 5 data points after the anomaly point). The window should not contain other data points that have already been marked as anomalies to ensure that the statistical features reflect the normal background distribution. The rate of change of adjacent data points can be obtained by calculating the difference or ratio of data values ​​within adjacent time periods; the duration of the deviation can be obtained by recording the length of time the prediction deviation exceeds the confidence interval.

[0073] An anomaly type classifier can be a machine learning model, such as a support vector machine, decision tree, random forest, or neural network. It needs to be trained on a large amount of historical data labeled with known anomaly types to accurately identify and distinguish different types of anomalies. During training, the classifier learns the mapping relationship between multidimensional features and anomaly types.

[0074] Anomaly type labels, including isolated anomalies, abrupt changes, drift anomalies, and continuous loss anomalies, represent common anomaly patterns in power grid time series data. Isolated anomalies typically refer to a single or a few data points suddenly deviating from the normal range; abrupt changes refer to drastic and significant changes in data values ​​within a short period; drift anomalies manifest as data values ​​slowly but continuously deviating from the normal baseline; and continuous loss anomalies refer to data that is continuously missing or exhibits abnormally low / high values ​​over a period of time. Refined classification of these anomalies provides an accurate basis for subsequent targeted data completion.

[0075] As a preferred embodiment, assume that real-time load data of a power grid line is marked as candidate anomaly data. To perform causal decomposition, its multidimensional features are first extracted. Specifically, the absolute deviation between the anomaly load value and the predicted value of the time-series large model is calculated; for example, if the actual load is 100MW and the predicted load is 80MW, the absolute deviation is 20MW. Next, load data for 10 minutes before and after the anomaly point (assuming a data sampling frequency of 1 minute) are selected, and the mean and variance of these 20 data points are calculated to assess the impact of the anomaly on the local load distribution. Simultaneously, the rate of change of the load data from the anomaly point to the previous minute is calculated; for example, if the load in the previous minute was 85MW, the rate of change is (100-85) / 85≈17.6%. Furthermore, the duration of the load deviation is recorded; for example, if the load exceeds the confidence interval for 5 consecutive minutes, the duration is 5.

[0076] These features (20MW absolute deviation, mean over 10 minutes, variance over 10 minutes, 17.6% rate of change, and 5-minute duration) are input into a pre-trained deep neural network-based anomaly type classifier. This classifier was trained using a large amount of historical power grid anomaly data and learned feature patterns for different anomaly types such as isolated anomalies, abrupt changes, drift anomalies, and continuous loss anomalies. For example, if the classifier identifies a combination of features—high rate of change, moderate duration, and significant changes in local data distribution—it might classify it as an abrupt change anomaly.

[0077] Ultimately, the classifier outputs a mutation anomaly determination and appends this label to the anomalous load data point, generating a load time series sequence with a mutation anomaly label. This label sequence will guide the subsequent completion module to match it with a completion strategy specifically designed for mutation anomalies, such as using interpolation or prediction methods based on historically similar mutation patterns, rather than simple linear interpolation.

[0078] Through the above technical solution, this application can accurately decompose the causes and classify the anomalies of candidate abnormal data, thus providing a solid foundation for subsequent differentiated completion strategies. This refined processing method avoids the limitations of using a single completion strategy for all abnormal data, significantly improving the accuracy and rationality of data completion.

[0079] Furthermore, based on the anomaly type, the steps of matching corresponding completion strategies to candidate anomaly data, and completing the candidate anomaly data based on the completion strategies to obtain the completed time series sequence include: Based on the anomaly type markers in the time sequence, the corresponding completion strategy is matched from the preset completion strategy library. The completion strategy includes the historical context window length and constraint strength used for completion. A preliminary complete value is generated based on the historical context window length; the preliminary complete value is corrected according to the power grid physical constraint rules corresponding to the constraint strength to obtain the corrected final complete value. The power grid physical constraint rules include the upper and lower limits of load peak and valley periods, the new energy output being not less than zero and not greater than the installed capacity, and the synchronicity constraint between photovoltaic output and solar irradiance. The candidate outlier data is completed based on the final completion value to obtain the complete time series sequence.

[0080] Specifically, the anomaly type labeling in the time series refers to the classification results output by the anomaly type classifier, such as isolated anomalies, abrupt changes, drift anomalies, and continuous defect anomalies. These labels provide important guidance for the subsequent completion process, enabling the completion strategy to address different types of anomalies in a targeted manner. The pre-defined completion strategy library is a collection of various completion algorithms and parameter configurations, each associated with a specific anomaly type. For example, for isolated anomalies, interpolation-based or local averaging strategies might be used; for continuous defect anomalies, strategies based on sequence generation models or historical similarity pattern matching might be used. These completion strategies specifically include the historical context window length and constraint strength used for completion. The historical context window length defines the range of historical data to be considered when generating the initial completed value, aiming to capture the local trends and periodicity of the time series. The constraint strength indicates the strictness or weight of the power grid physical constraint rules when correcting the initial completed value.

[0081] Specifically, initial imputation values ​​can be generated using various methods, such as linear interpolation, multinomial fitting, moving average, prediction based on historical similarity patterns, or simple machine learning models (e.g., support vector regression, random forest). For example, if the historical context window length is 10 data points, the data trends and patterns of these 10 data points can be used to predict the values ​​of outliers.

[0082] Specifically, the upper and lower limits of load peak and valley periods ensure that the supplemented load data does not exceed the reasonable range of specific time periods; the output of new energy sources being neither less than zero nor greater than the installed capacity ensures that the output data of new energy sources such as photovoltaic and wind power are physically feasible; the synchronicity constraint between photovoltaic output and daily irradiance ensures the consistency between the photovoltaic output data and the trend of environmental irradiance changes. The introduction of these rules makes the supplemented data not only statistically reasonable but also physically feasible. Therefore, based on the final supplemented value, the above-mentioned candidate abnormal data is supplemented to obtain the above complete time series.

[0083] As a preferred embodiment, suppose that the output data of a photovoltaic power station shows continuous abnormalities within a certain period of time.

[0084] First, the anomaly is identified as a continuous defect anomaly by an anomaly type classifier, and a time series with this label is generated.

[0085] Next, based on the markers of consecutive defect anomalies, the corresponding completion strategy is matched from the preset completion strategy library. This strategy stipulates that for consecutive defect anomalies, the historical context window length is set to 2 hours (i.e., data from 1 hour before and after the anomaly point), and the constraint strength is set to strong.

[0086] Then, based on a 2-hour historical context window, linear interpolation is performed using the data within the window to generate preliminary completion values. For example, if the missing data occurs between 10:00 and 10:30, interpolation will be performed using the data trends from 09:00 to 10:00 and 10:30 to 11:30.

[0087] Subsequently, the initial supplementary values ​​are corrected according to the power grid physical constraint rules corresponding to the strong constraint strength. Specifically, the correction process will check: Is the photovoltaic output less than zero? If the initial completion value is negative, then correct it to 0.

[0088] Does the photovoltaic output exceed the installed capacity? If the initial supplementary value exceeds the installed capacity of the photovoltaic power station (e.g., 50MW), then it is corrected to 50MW.

[0089] Synchronization of photovoltaic output with daily irradiance: Obtain daily irradiance data for this period of time, and fine-tune the initial supplementary value according to the trend of irradiance change to ensure that it is consistent with the change of irradiance. For example, when the irradiance rises rapidly, the output should also rise accordingly.

[0090] After the above corrections, the final completed value is obtained.

[0091] Finally, these final completed values ​​are filled into the original continuously missing data points to obtain a complete photovoltaic power output time series that conforms to physical laws.

[0092] Through the above technical solution, this application solves the problems that the matching and application of the completion strategy may have insufficient accuracy, difficulty in capturing complex nonlinear relationships and long-period dependencies, and easy violation of the physical constraints of the power system, resulting in low reliability of the completion results. It improves the accuracy and reliability of data completion and ensures that the completion results comply with the physical constraints of the power grid.

[0093] This application further proposes that, after obtaining the complete time series sequence, it includes: The complete time series sequence is incorporated into the model training dataset of the large time series model. The large time series model is then optimized by combining the anomaly type and the real time series data within the current time series window.

[0094] Specifically, a complete time series refers to time series data that has been processed using the methods described above, with identified and supplemented anomalous data, thus restoring data integrity and consistency. This complete time series data is included in the training dataset of the large-scale time series model, aiming to provide the model with the latest, most accurate, and cleaned data samples so that the model can continuously learn based on this high-quality data. Anomaly type refers to the specific anomaly category identified after causal decomposition of candidate anomaly data, such as isolated anomalies, mutation anomalies, drift anomalies, and continuous loss anomalies. Combining anomaly types allows the large-scale time series model to not only learn normal data patterns during optimization but also understand the characteristics and causes of different anomaly patterns, thereby improving its ability to identify and handle future anomalies. The real time series data within the current time series window can be understood as the original observation data used for anomaly detection and supplementation operations. It, along with the supplemented complete time series, serves as training input, helping the model maintain sensitivity to the features of the original data while learning the supplementation strategy. Optimizing large time series models can specifically include fine-tuning the model's parameters, performing incremental training, or periodically retraining to enable them to better adapt to new data distributions and patterns.

[0095] The above technical solutions effectively address the potential performance degradation of the model during long-term operation, ensuring that the model can consistently adapt to changing power grid operating conditions and data characteristics. Furthermore, by incorporating anomaly type optimization, the model's ability to understand and handle different anomaly patterns is enhanced, thereby reducing the risk of recurring similar prediction biases and providing more reliable data support for the stable operation of the power grid.

[0096] Secondly, see Figure 2 This application also discloses a time series data augmentation system based on large-scale time series model fusion, used to perform the above-mentioned time series data augmentation method based on large-scale time series model fusion. The system includes: The data acquisition and processing module 210 is used to acquire time series data to be processed, discretize the time series data, and obtain a discretized time feature sequence. The time series prediction module 220 is used to input the time feature sequence into the time series large model to obtain the time series prediction value sequence output by the time series large model corresponding to the current time series window; The deviation calculation module 230 is used to calculate the prediction deviation between the predicted value data in the time series prediction value sequence and the original observation value at the corresponding time point in the time series data. The abnormal data marking module 240 is used to obtain the historical prediction deviation sequence within the current time series window, construct a confidence interval based on the historical prediction deviation sequence, and mark the time series data as candidate abnormal data if the prediction deviation exceeds the confidence interval. The data completion module 250 is used to perform causal decomposition on candidate abnormal data, classify them into different abnormal types, match corresponding completion strategies for candidate abnormal data according to the abnormal type, and complete the candidate abnormal data based on the completion strategy to obtain the complete time series sequence.

[0097] This technical solution provides an integrated system that automates and intelligently enhances the processing of power grid time series data, effectively addressing the shortcomings of insufficient end-to-end automation in existing technologies and improving data processing efficiency and reliability. The above descriptions are merely embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A time series data augmentation method based on large-scale time series model fusion, used to process time series data of power grids, characterized in that, The method includes: Obtain the time series data to be processed, and discretize the time series data to obtain the discretized time feature sequence; The time feature sequence is input into the time series large model to obtain the time series prediction value sequence output by the time series large model corresponding to the current time series window; Calculate the prediction deviation between the predicted values ​​in the time-series predicted value sequence and the original observations at the corresponding time points in the time-series data; Obtain the historical prediction deviation sequence within the current time series window, construct a confidence interval based on the historical prediction deviation sequence, and if the prediction deviation exceeds the confidence interval, mark the time series data as candidate anomalous data. The candidate abnormal data is decomposed into different abnormal types. Based on the abnormal type, a corresponding completion strategy is matched for the candidate abnormal data. The candidate abnormal data is completed based on the completion strategy to obtain the completed time series sequence.

2. The time series data augmentation method based on large-scale time series model fusion according to claim 1, characterized in that, After the step of acquiring the time series data to be processed, and before the step of discretizing the time series data, the process includes: Determine whether there is a time series deviation in the time series data that is related to changes in ambient temperature; If the time series deviation is determined to exist, the time series data is corrected. The corrected time series data is reconstructed using signals across multiple time ranges, and the reconstructed time series data is output as the basis for discretization processing.

3. The time series data augmentation method based on large-scale time series model fusion according to claim 2, characterized in that, The step of determining whether there is a time series deviation in the time series data that is related to changes in ambient temperature includes: Monitor the rate of change of ambient temperature; When the rate of change of the ambient temperature exceeds a first preset threshold, the irradiance deviation is predicted based on a preset correlation function between temperature and time deviation. If the difference between the observed irradiance value and the theoretical irradiance value in the time series data is consistent with the irradiance deviation, then it is determined that there is a time series deviation in the time series data that is related to changes in ambient temperature.

4. The time series data augmentation method based on large-scale time series model fusion according to claim 3, characterized in that, The step of correcting the time series data if the time series deviation is determined to exist includes: Calculate a correction factor corresponding to the rate of change of the ambient temperature based on the real-time ambient temperature. The time series data are corrected using the correction factor to eliminate the time series bias.

5. A time series data augmentation method based on large-scale time series model fusion according to claim 4, characterized in that, The steps of reconstructing signals from the corrected time series data across multiple time ranges and outputting the reconstructed time series data as the basis for discretization include: Acquire real-time meteorological observation data within the current time window, wherein the real-time meteorological observation data includes cloud cover change rate and local severe convective weather warning indicators; Based on the real-time meteorological observation data, the process noise covariance and measurement noise covariance parameters of the Kalman filter are adjusted. Specifically, when the cloud cover change rate is detected to exceed the second preset threshold or the local severe convective weather warning sign indicates the existence of a local severe convective weather warning, the process noise covariance is increased while the measurement noise covariance is decreased. The corrected time series data is subjected to Kalman filtering for scaling. Specifically, within a preset minute-level time range, the corrected time series data is subjected to point-by-point prediction and updating using Kalman filtering. Within a preset hour-level time range, the updated time series data is subjected to secondary smoothing to obtain the reconstructed time series data.

6. The time series data augmentation method based on large-scale time series model fusion according to claim 1, characterized in that, The step of constructing a confidence interval based on the historical prediction deviation sequence includes: Calculate the center value of the historical prediction deviation sequence; Based on the central value, the dynamic standard deviation is increased and decreased respectively to obtain the upper and lower boundaries of the dynamic confidence interval. The dynamic standard deviation is dynamically adjusted according to external scene information, which includes weather change warnings, holiday markers, or sudden changes in new energy output. When at least one of the weather change warnings, holiday markers, or sudden changes in new energy output is detected, the dynamic standard deviation is increased by a preset multiple to widen the confidence interval range.

7. The time series data augmentation method based on large-scale temporal model fusion according to claim 1, characterized in that, The step of performing causal decomposition on the candidate abnormal data and classifying it into different abnormal types includes: Extract multidimensional features from the candidate anomaly data. The multidimensional features include: the absolute and relative values ​​of the prediction deviation, the mean and variance of the data before and after the anomaly point, the rate of change of adjacent data points, and the duration of the deviation. The extracted multidimensional features are input into a pre-trained anomaly type classifier; The anomaly type classifier outputs the anomaly type determination result, generating a labeled time series sequence with anomaly type labels, wherein the anomaly type labels include isolated anomalies, mutation anomalies, drift anomalies, and continuous defect anomalies.

8. A time series data augmentation method based on large-scale time series model fusion according to claim 7, characterized in that, The step of matching a corresponding completion strategy to the candidate anomaly data according to the anomaly type, and completing the candidate anomaly data based on the completion strategy to obtain the completed time series sequence includes: Based on the anomaly type marker in the marked time sequence, a corresponding completion strategy is matched from a preset completion strategy library. The completion strategy includes the historical context window length and constraint strength used for completion. A preliminary completion value is generated based on the aforementioned historical context window length; Based on the power grid physical constraint rules corresponding to the constraint strength, the preliminary completion value is corrected to obtain the corrected final completion value. The power grid physical constraint rules include the upper and lower limits of load peak and valley periods, the new energy output being not less than zero and not greater than the installed capacity, and the synchronicity constraint between photovoltaic output and solar irradiance. The candidate abnormal data is completed based on the final completion value to obtain the complete time series sequence.

9. A time series data augmentation method based on large-scale time series model fusion according to claim 1, characterized in that, After obtaining the complete time series sequence, it includes: The complete time series sequence is incorporated into the model training dataset of the large time series model. The large time series model is then optimized by combining the anomaly type and the real time series data within the current time series window.

10. A time series data augmentation system based on large-scale temporal model fusion, used to execute the time series data augmentation method based on large-scale temporal model fusion as described in any one of claims 1 to 9, characterized in that, The system includes: The data acquisition and processing module is used to acquire time series data to be processed, and to discretize the time series data to obtain a discretized time feature sequence. The time series prediction module is used to input the time feature sequence into the time series large model to obtain the time series prediction value sequence output by the time series large model corresponding to the current time series window; The deviation calculation module is used to calculate the prediction deviation between the predicted value data in the time series prediction value sequence and the original observation value at the corresponding time point in the time series data. An anomaly data marking module is used to obtain the historical prediction deviation sequence within the current time series window, construct a confidence interval based on the historical prediction deviation sequence, and mark the time series data as candidate anomaly data if the prediction deviation exceeds the confidence interval. A data completion module is used to perform causal decomposition on the candidate anomaly data, divide it into different anomaly types, match a corresponding completion strategy for the candidate anomaly data according to the anomaly type, and complete the candidate anomaly data based on the completion strategy to obtain the completed time series sequence.

Images