A method and system for dynamic threshold early warning of power grid resources based on multi-source data

Abstract

This invention relates to the field of power grid resource management technology, and particularly to a method and system for dynamic threshold early warning of power grid resources based on multi-source data. The method includes the following steps: dynamically weighting and combining standard historical multi-source power grid load data to obtain dynamic baseline load data; applying adaptive load early warning threshold processing to the dynamic baseline load data to obtain dynamic load early warning threshold curve data; constructing a power grid load prediction model; using the power grid load prediction model to perform short-term load prediction to obtain short-term load prediction data; calculating the threshold load difference between the short-term load prediction data and the dynamic load early warning threshold curve data, and performing multi-level early warning triggering processing to obtain a multi-level early warning triggering strategy for the power grid. This invention achieves real-time monitoring and adaptive early warning by dynamically assessing changes in power grid load, effectively reducing false alarms and missed alarms, and improving the level of power grid safety management.

Description

Technical Field

[0001] This invention relates to the field of power grid resource management technology, and in particular to a method and system for dynamic threshold early warning of power grid resources based on multi-source data. Background Technology

[0002] With continuously growing electricity demand, expanding power grid scale, and increasingly complex grid structure, the safe and stable operation of the power grid is crucial for national economic and social development. The power grid operates in a complex and ever-changing environment, and various uncertainties, such as natural disasters, equipment failures, and human error, can threaten its safe and stable operation. Therefore, effectively assessing and providing early warnings about the risks to power grid resources is of significant practical importance for ensuring the safe and stable operation of the power grid and improving power supply reliability. Power grid load, as a crucial parameter for power grid operation, directly affects the safety and economy of the grid. Power grid load itself exhibits significant dynamic characteristics. Load magnitude is influenced by external factors such as seasons, weather, and holidays. The combined effect of these factors leads to complex time-varying characteristics in power grid load. For example, daily load curves show peak-valley variations, weekly load curves show differences between weekdays and non-weekdays, and seasonal load curves show summer and winter peaks. Fixed thresholds cannot capture these dynamic changes and are insufficient to accurately reflect the true state of the load. When the load fluctuates within the normal range, alarms are triggered if it exceeds a set fixed threshold, leading to frequent false alarms. This not only increases the workload of maintenance personnel but may also cause misjudgments of the power grid's operating status, reducing the reliability of the early warning system. However, traditional power grid load monitoring and early warning methods often use fixed thresholds, i.e., a fixed upper or lower load limit is preset. When the actual load exceeds these thresholds, the system issues an alarm. This approach is difficult to adapt to the load characteristics of different regions and time periods, and is prone to over-warning or under-warning. Summary of the Invention

[0003] Based on this, the present invention provides a method and system for early warning of dynamic thresholds for power grid resources based on multi-source data, in order to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a dynamic threshold early warning method for power grid resources based on multi-source data includes the following steps:

[0005] Step S1: Obtain standard historical multi-source power grid load data; classify the standard historical multi-source power grid load data by historical load monitoring date, and decompose the power grid load time-frequency components to generate power grid load time-frequency component data;

[0006] Step S2: Set the rolling window for grid load based on the time-frequency component data of grid load, and perform dynamic baseline value weighted combination to obtain the dynamic baseline value data of grid load;

[0007] Step S3: Filter the standard historical multi-source power grid load data for effective historical load data and calculate the coefficient of variation to generate effective historical power grid load variation coefficients; perform sensitivity parameter analysis of early warning sources based on effective historical power grid load variation coefficients to generate optimal alarm sensitivity parameters; perform adaptive load early warning threshold processing on the dynamic baseline value data of power grid load based on effective historical power grid load variation coefficients and optimal alarm sensitivity parameters to obtain dynamic load early warning threshold curve data.

[0008] Step S4: Obtain real-time power grid resource and equipment status data; construct the power grid monitoring resource topology based on the real-time power grid resource and equipment status data to generate power grid monitoring resource topology data; construct a power grid load forecasting model based on the power grid monitoring resource topology data; use the power grid load forecasting model to perform short-term load forecasting to obtain short-term load forecasting data.

[0009] Step S5: Calculate the threshold load difference of short-term load forecast data using dynamic load early warning threshold curve data, and perform multi-level early warning triggering processing to obtain the power grid multi-level early warning triggering strategy.

[0010] This invention utilizes multi-source historical power grid load data and performs time-frequency component decomposition to extract the periodic patterns and random fluctuations in load changes. This refined data processing approach enables the method to more accurately capture the dynamic characteristics of load changes, avoiding the shortcomings of traditional methods in understanding load change patterns. By classifying historical load data by monitoring date, different types of date load patterns (e.g., weekdays, weekends, holidays) are distinguished, and corresponding early warning models are constructed for each, further improving the accuracy of early warnings and avoiding the problem that a single model cannot adapt to the differences in load patterns on different dates. A rolling window mechanism and a weighted combination algorithm are used to construct a dynamic baseline value, which can track load change trends in real time and adaptively adjust the baseline level according to changes in historical data. This effectively solves the problem that fixed threshold methods cannot adapt to periodic load fluctuations and long-term trend changes. The introduction of dynamic baseline values ​​means that the early warning threshold is no longer a static value, but a curve that can be dynamically adjusted with load changes, thus better reflecting actual load changes and reducing false alarms and missed alarms. Based on the coefficient of variation of historical load data, the sensitivity parameters of the early warning source are analyzed, and the optimal alarm sensitivity parameters are finally determined. Based on the load fluctuation characteristics of different regions and time periods, the system adaptively adjusts the sensitivity of early warnings. For regions or time periods with large load fluctuations, the system automatically increases the early warning sensitivity to promptly detect potential risks; while for regions or time periods with small load fluctuations, the system correspondingly decreases the early warning sensitivity to avoid over-warning and reduce unnecessary maintenance work. By comparing the prediction results with dynamic early warning thresholds, the system achieves prediction and early warning of future load changes. This prediction-early warning linkage mechanism can detect potential load risks in advance, providing sufficient response time for power grid dispatching and operation personnel, effectively preventing accidents, and ensuring the safe and stable operation of the power grid. The application of a multi-level early warning triggering strategy further enhances the effectiveness and practicality of early warnings, enabling the issuance of early warning signals in stages according to the degree to which the load exceeds the threshold, guiding different levels of emergency response measures, and improving the level of refined management of power grid operation. Therefore, this invention's dynamic threshold early warning method for power grid resources based on multi-source data comprehensively analyzes historical load data and real-time power grid resource equipment status, uses time-frequency component decomposition technology to capture dynamic changes in power grid load, and combines rolling windows and weighted combination algorithms to construct dynamic baseline values, achieving refined monitoring and adaptive early warning of power grid resource load changes.

[0011] Preferably, step S1 includes the following steps:

[0012] Step S11: Obtain historical multi-source power grid resource load data;

[0013] Step S12: Unify the format of historical multi-source power grid resource load data from multiple data sources, and perform data preprocessing to generate standard historical multi-source power grid load data;

[0014] Step S13: Classify the historical load monitoring dates according to the standard historical multi-source power grid load data to obtain classified historical multi-source power grid load data;

[0015] Step S14: Perform Fast Fourier Transform on the classified historical multi-source power grid load data to generate historical power grid load spectrum data;

[0016] Step S15: Decompose the time-frequency components of the power grid load based on the historical power grid load spectrum data to generate the time-frequency component data of the power grid load.

[0017] This invention classifies historical load monitoring dates, enabling the differentiation of load patterns across different date types (e.g., weekdays, weekends, holidays) and allowing for separate analysis and modeling for each date type. This effectively solves the problem of traditional methods struggling to distinguish load patterns across different dates, improving the targeting and accuracy of early warnings. For example, weekday load curves typically exhibit distinct peak-valley characteristics, while weekend and holiday load curves are relatively stable. Date classification allows for the construction of more accurate load forecasting models for different date types, thereby improving the accuracy of early warnings. Time-frequency component decomposition further breaks down load data into components of different frequencies, thus more finely characterizing the dynamics of load changes and more accurately capturing the patterns of load variation. By extracting the periodic components of load changes, future load change trends can be predicted more accurately; while by analyzing random fluctuation components, the uncertainty of load changes can be better assessed, thereby improving the reliability of early warnings.

[0018] Preferably, step S15 includes the following steps:

[0019] Step S151: Analyze the frequency range of major loads based on historical power grid load spectrum data to obtain historical load frequency range data;

[0020] Step S152: Perform maximum wavelet decomposition layer processing based on historical load frequency range data to obtain the maximum wavelet decomposition layer;

[0021] Step S153: Divide the frequency range of the decomposition level according to the maximum wavelet decomposition level, and perform decomposition level mapping on the historical load frequency range data to obtain load frequency-decomposition level mapping data.

[0022] Step S154: Determine the coverage range of the decomposition layer for the load frequency-decomposition level mapping data, and optimize the wavelet decomposition level to generate the optimal decomposition layer data;

[0023] Step S155: Perform wavelet transform decomposition on historical power grid load spectrum data based on the optimal decomposition level data to generate wavelet decomposition coefficients;

[0024] Step S156: Extract wavelet approximate components based on wavelet decomposition coefficients to obtain long-term trend data of power grid load;

[0025] Step S157: Extract wavelet detail components based on wavelet decomposition coefficients to obtain intraday power grid load fluctuation data, weekly power grid load fluctuation data, and seasonal power grid load fluctuation data, respectively.

[0026] Step S158: Combine the long-term trend data of power grid load, the intraday fluctuation data of power grid load, the intraweek fluctuation data of power grid load, and the seasonal fluctuation data of power grid load with the time-frequency component of power grid resource monitoring to obtain the time-frequency component data of power grid load.

[0027] This invention combines the division of frequency ranges for decomposition levels with the determination of the coverage of the decomposition level, enabling intelligent determination of the optimal number of wavelet decomposition levels. This ensures that wavelet decomposition can better capture load change characteristics at different time scales. Historical power grid load data is decomposed into fluctuation components at different time scales, such as long-term trends, intraday fluctuations, weekly fluctuations, and seasonal fluctuations. This multi-scale decomposition method can more comprehensively characterize the dynamic characteristics of load changes, avoiding the limitations of traditional methods that only focus on single-scale fluctuation information. By combining the extracted time-frequency component data, a complete power grid load time-frequency component data is constructed. This multi-level, multi-angle load data analysis can more accurately reflect the true situation of load changes, thereby improving the accuracy and reliability of early warning.

[0028] Preferably, step S2 includes the following steps:

[0029] Step S21: Set the rolling window for grid load based on the time-frequency component data of grid load, and obtain the annual trend baseline rolling window, intraday fluctuation baseline rolling window, weekly fluctuation baseline rolling window and seasonal fluctuation baseline rolling window respectively;

[0030] Step S22: Calculate the long-term trend baseline of the power grid load using the annual trend baseline rolling window to generate long-term trend baseline data;

[0031] Step S23: Calculate the intraday fluctuation baseline of the grid load intraday fluctuation data through the intraday fluctuation baseline rolling window to generate intraday fluctuation baseline data;

[0032] Step S24: Calculate the weekly fluctuation baseline of the grid load within the week using the weekly fluctuation baseline rolling window to generate weekly fluctuation baseline data;

[0033] Step S25: Calculate the seasonal fluctuation baseline for the power grid load seasonal fluctuation data using the seasonal fluctuation baseline rolling window to generate seasonal fluctuation baseline data;

[0034] Step S26: Perform a dynamic baseline value weighted combination based on the long-term trend baseline data, intraday fluctuation baseline data, weekly fluctuation baseline data, and seasonal fluctuation baseline data to obtain the dynamic baseline value data of the power grid load.

[0035] This invention achieves accurate capture of load change trends at different time scales by separately setting rolling windows for annual trend baselines, intraday fluctuation baselines, weekly fluctuation baselines, and seasonal fluctuation baselines, and combining load data at different time scales. This divide-and-conquer strategy can more effectively handle the complexity of load changes at different time scales, avoiding the problem that a single baseline cannot simultaneously adapt to fluctuations at multiple time scales. By weighted combining the long-term trend baseline data, intraday fluctuation baseline data, weekly fluctuation baseline data, and seasonal fluctuation baseline data, this method constructs the final dynamic baseline value data for the power grid load. This weighted combination method can comprehensively consider the impact of load changes at different time scales and assign different weights according to the importance of fluctuation characteristics at different time scales, thus making the dynamic baseline value more representative and reliable. For example, during the summer peak period, the weight of the seasonal fluctuation baseline will be higher, while during other time periods, the weight of the intraday fluctuation baseline will be higher.

[0036] Preferably, step S3 includes the following steps:

[0037] Step S31: Obtain the real-time power grid monitoring timestamp;

[0038] Step S32: Filter the standard historical multi-source power grid load data for effective historical load data by using the real-time power grid monitoring timestamp, and calculate the coefficient of variation to generate the effective historical load variation coefficient of the power grid.

[0039] Step S33: Analyze the sensitivity parameters of the early warning source based on the effective historical load variation coefficient of the power grid, and generate the optimal alarm sensitivity parameters;

[0040] Step S34: Calculate the load warning threshold based on the effective historical load variation coefficient of the power grid and the optimal alarm sensitivity parameter, and apply adaptive threshold constraints to obtain dynamic load warning threshold curve data.

[0041] This invention obtains real-time power grid monitoring timestamps and uses them to filter effective historical load data from standard historical multi-source power grid data. This method ensures that the historical data used to calculate the coefficient of variation is correlated with the load characteristics at the current time. This avoids warning bias caused by using outdated or irrelevant data, improving the accuracy and timeliness of warnings. Calculating the coefficient of variation based on the filtered effective historical load data allows for a more accurate quantification of load fluctuation. The coefficient of variation reflects the dispersion of load data and can be used to measure the stability of load changes. A larger coefficient of variation indicates greater load fluctuation, requiring higher warning sensitivity; while a smaller coefficient of variation indicates less load fluctuation, allowing for a more appropriate reduction in warning sensitivity. By analyzing the warning source sensitivity parameters of the effective historical power grid load coefficient of variation and generating optimal alarm sensitivity parameters, this method achieves adaptive adjustment of warning sensitivity. This effectively solves the problem that traditional fixed-threshold warning methods cannot adapt to the differences in load fluctuation characteristics in different regions and time periods, avoiding false alarms and missed alarms.

[0042] Preferably, step S32 includes the following steps:

[0043] Step S321: Select the historical effective load window based on the real-time power grid monitoring timestamp to obtain the effective historical load window data;

[0044] Step S322: Filter the standard historical multi-source power grid load data using the effective historical load window data to obtain the effective historical load data of the power grid;

[0045] Step S323: Calculate the historical load average based on the effective historical load data of the power grid to generate effective historical load average data;

[0046] Step S324: Calculate the historical load standard deviation based on the effective historical load data of the power grid to generate effective historical standard deviation data;

[0047] Step S325: Calculate the coefficient of variation based on the effective historical load mean data and the effective historical standard deviation data to generate the effective historical load variation coefficient of the power grid.

[0048] This invention selects a historical effective load window based on real-time power grid monitoring timestamps. This method intelligently determines the range of historical data used to calculate the coefficient of variation. This ensures that the selected historical data is highly correlated with the load characteristics at the current time point, avoiding interference from outdated or irrelevant data and thus improving the accuracy and representativeness of the coefficient of variation. By using the effective historical load window data to filter standard historical multi-source power grid load data, this ensures that the data used to calculate the coefficient of variation comes from historical periods with load characteristics similar to the current time point, thereby more accurately reflecting the fluctuations in the current load. After filtering the effective historical load data, the method calculates the historical load mean and the historical load standard deviation. The mean reflects the average level of the historical load, while the standard deviation reflects the dispersion of the historical load.

[0049] Preferably, step S33 includes the following steps:

[0050] Step S331: Analyze the distribution characteristics of the coefficient of variation based on the effective historical load variation coefficient of the power grid, and set the initial sensitivity parameters to obtain the range data of the initial sensitivity parameters;

[0051] Step S332: Obtain power grid resource load early warning source configuration data;

[0052] Step S333: Connect the early warning source according to the power grid resource load early warning source configuration data, and read the historical early warning records to obtain the original power grid early warning record data;

[0053] Step S334: Filter the actual alarm records based on the original power grid early warning record data to obtain the actual power grid alarm record data;

[0054] Step S335: Filter false alarm records based on the original power grid early warning record data to obtain power grid false alarm record data;

[0055] Step S336: Use the preset sensitivity traversal step size data to iterate through the initial sensitivity parameter range data to obtain the traversal sensitivity parameters;

[0056] Step S337: Calculate the alarm recall rate by traversing the sensitivity parameters to obtain the power grid alarm recall rate data;

[0057] Step S338: Calculate the alarm accuracy by iterating through the sensitivity parameters to obtain the power grid alarm accuracy data by analyzing the actual alarm record data and the false alarm record data.

[0058] Step S339: Calculate the early warning performance index of the traversal sensitivity parameter based on the power grid alarm precision rate data and the power grid alarm recall rate data, and select the optimal sensitivity parameter to obtain the optimal alarm sensitivity parameter.

[0059] This invention filters raw power grid early warning record data to obtain both actual and false alarm records. It iterates through the initial sensitivity parameter range using a preset sensitivity traversal step size, and calculates alarm recall and precision based on both data. This allows for a comprehensive evaluation of the early warning system performance under different sensitivity parameters. Recall reflects the proportion of true alarms correctly identified by the early warning system, while precision reflects the proportion of true alarms among the alarms issued by the system. These two metrics together constitute key indicators for evaluating early warning system performance. Based on the power grid alarm precision and recall data, early warning performance indicators are calculated for the traversed sensitivity parameters, and the optimal sensitivity parameter is selected. By comprehensively considering both recall and precision, and selecting the sensitivity parameter that achieves the best balance between the two, early warning performance is optimized.

[0060] Preferably, step S4 includes the following steps:

[0061] Step S41: Obtain real-time power grid resource equipment status data;

[0062] Step S42: Construct the power grid monitoring resource topology based on real-time power grid resource equipment status data, and generate power grid monitoring resource topology data;

[0063] Step S43: Collect real-time load data of power grid monitoring nodes based on power grid monitoring resource topology data to obtain real-time load data of power grid resources;

[0064] Step S44: Perform Bayesian network node processing based on power grid monitoring resource topology data, and construct a power grid load prediction model through a preset convolutional neural network model to obtain the power grid load prediction model to be trained.

[0065] Step S45: Process the standard historical multi-source power grid load data into a model training set, and train the power grid load prediction model to be trained to obtain the power grid load prediction model;

[0066] Step S46: Use the power grid load forecasting model to perform short-term load forecasting on the real-time load data of power grid resources to obtain short-term load forecasting data.

[0067] This invention utilizes real-time load acquisition from power grid monitoring nodes based on power grid monitoring resource topology data to obtain the latest load data, providing real-time data input for short-term load forecasting. This ensures the timeliness of the forecast results, enabling a more timely reflection of current power grid load changes. A power grid load forecasting model is constructed using Bayesian network node processing and a pre-defined convolutional neural network model. The model is trained using standard historical multi-source power grid load data, allowing it to learn the changing patterns of historical load data and apply them to future load forecasts. Bayesian networks effectively handle uncertainties and complex relationships, while convolutional neural networks excel at capturing spatial and temporal features in time-series data; the combination of the two improves the accuracy of load forecasting. Using the trained power grid load forecasting model, short-term load forecasting is performed on real-time power grid resource load data, yielding predicted short-term load data.

[0068] Preferably, step S5 includes the following steps:

[0069] Step S51: Extract the forecast time points from the short-term load forecast data to obtain the power grid resource forecast time point data;

[0070] Step S52: Identify the real-time load warning threshold range of the dynamic load warning threshold curve data using the power grid resource prediction time point data, and calculate the threshold load difference of the short-term load prediction data to obtain the predicted load difference data;

[0071] Step S53: Calculate the load over-limit range based on the predicted load difference data to obtain the predicted load over-limit range data of the power grid;

[0072] Step S54: Divide the early warning levels according to the power grid predicted load over-limit data, and perform multi-level early warning triggering processing through the preset early warning level execution strategy to obtain the power grid multi-level early warning triggering strategy.

[0073] This invention extracts the prediction time points from short-term load forecast data and combines this with dynamic load warning threshold curve data to identify real-time load warning threshold ranges, accurately determining the warning threshold corresponding to each prediction time point. The dynamic threshold curve automatically adjusts the threshold level according to the load characteristics of different time periods, thus more effectively identifying abnormal load changes. By calculating the threshold load difference from the short-term load forecast data, predicted load difference data is obtained, reflecting the gap between the predicted load and the warning threshold. Calculating the load over-limit magnitude based on the predicted load difference data allows for a more precise quantification of the degree of load over-limit. The larger the over-limit magnitude, the higher the risk of load over-limit, requiring more urgent countermeasures. Different levels of warnings are triggered based on the degree of load over-limit, and corresponding warning strategies are executed, thereby improving the effectiveness and practicality of the warnings. For example, for minor load over-limit, the system only issues a reminder message; while for severe load over-limit, the system triggers a higher-level warning and takes corresponding control measures to ensure the safe and stable operation of the power grid.

[0074] Preferably, the present invention also provides a power grid resource dynamic threshold early warning system based on multi-source data, which executes the power grid resource dynamic threshold early warning method based on multi-source data as described above. The power grid resource dynamic threshold early warning system based on multi-source data includes:

[0075] The historical load time-frequency decomposition module is used to acquire standard historical multi-source power grid load data; classify the standard historical multi-source power grid load data by historical load monitoring date; and perform time-frequency component decomposition of power grid load to generate power grid load time-frequency component data.

[0076] The load dynamic baseline construction module is used to set the grid load rolling window based on the grid load time-frequency component data and to perform dynamic baseline value weighting combination to obtain grid load dynamic baseline value data.

[0077] The adaptive threshold generation module is used to filter effective historical load data from standard historical multi-source power grid load data and calculate the coefficient of variation to generate effective historical load variation coefficients; it analyzes the sensitivity parameters of the early warning source based on the effective historical load variation coefficients to generate optimal alarm sensitivity parameters; and it performs adaptive load early warning threshold processing on the dynamic baseline value data of power grid load based on the effective historical load variation coefficients and the optimal alarm sensitivity parameters to obtain dynamic load early warning threshold curve data.

[0078] The short-term load forecasting module is used to acquire real-time power grid resource and equipment status data; construct the power grid monitoring resource topology based on the real-time power grid resource and equipment status data to generate power grid monitoring resource topology data; construct a power grid load forecasting model based on the power grid monitoring resource topology data; and use the power grid load forecasting model to perform short-term load forecasting to obtain short-term load forecasting data.

[0079] The multi-level early warning triggering module is used to calculate the threshold load difference of short-term load forecast data through dynamic load early warning threshold curve data, and to perform multi-level early warning triggering processing to obtain the power grid multi-level early warning triggering strategy. Attached Figure Description

[0080] Figure 1 This is a flowchart illustrating the steps of the dynamic threshold early warning method for power grid resources based on multi-source data according to the present invention.

[0081] Figure 2 for Figure 1 A detailed flowchart illustrating the implementation steps of step S4.

[0082] Figure 3 for Figure 1 A detailed flowchart illustrating the implementation steps of step S5.

[0083] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0084] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0085] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0086] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0087] To achieve the above objectives, please refer to Figures 1 to 3 This invention provides a method for dynamic threshold early warning of power grid resources based on multi-source data, comprising the following steps:

[0088] Step S1: Obtain standard historical multi-source power grid load data; classify the standard historical multi-source power grid load data by historical load monitoring date, and decompose the power grid load time-frequency components to generate power grid load time-frequency component data;

[0089] Step S2: Set the grid load rolling window based on the grid load time-frequency component data, and perform dynamic baseline value weighted combination to obtain grid load dynamic baseline value data;

[0090] Step S3: Filter the standard historical multi-source power grid load data for effective historical load data and calculate the coefficient of variation to generate effective historical power grid load variation coefficients; perform sensitivity parameter analysis of early warning sources based on effective historical power grid load variation coefficients to generate optimal alarm sensitivity parameters; perform adaptive load early warning threshold processing on the dynamic baseline value data of power grid load based on effective historical power grid load variation coefficients and optimal alarm sensitivity parameters to obtain dynamic load early warning threshold curve data.

[0091] Step S4: Obtain real-time power grid resource and equipment status data; construct the power grid monitoring resource topology based on the real-time power grid resource and equipment status data to generate power grid monitoring resource topology data; construct a power grid load forecasting model based on the power grid monitoring resource topology data; use the power grid load forecasting model to perform short-term load forecasting to obtain short-term load forecasting data.

[0092] Step S5: Calculate the threshold load difference of short-term load forecast data using dynamic load early warning threshold curve data, and perform multi-level early warning triggering processing to obtain the power grid multi-level early warning triggering strategy.

[0093] In this embodiment of the invention, reference Figure 1 The diagram shown is a flowchart illustrating the steps of the dynamic threshold early warning method for power grid resources based on multi-source data according to the present invention. In this embodiment, the dynamic threshold early warning method for power grid resources based on multi-source data includes the following steps:

[0094] Step S1: Obtain standard historical multi-source power grid load data; classify the standard historical multi-source power grid load data by historical load monitoring date, and decompose the power grid load time-frequency components to generate power grid load time-frequency component data;

[0095] In this embodiment of the invention, historical load data is acquired from multiple sources, including the power grid dispatch control center, smart substations, distribution automation systems, and advanced metering systems (AMI). This data includes electricity load information for different voltage levels and regions, with a time granularity of 15 minutes, 30 minutes, or 1 hour. The data format is uniformly a table of data composed of fields such as timestamps and load values. Historical load data is categorized according to date attributes, for example, by weekdays, weekends, and holidays, and seasonal factors such as summer and winter can also be considered. Wavelet transform or empirical mode decomposition (EMD) methods are used to decompose the categorized historical load data into time-frequency components. The original load data is decomposed into fluctuation components of different frequencies, such as daily cycle components, weekly cycle components, and random fluctuation components. The decomposed time-frequency component data can more clearly reflect the patterns of load changes, such as baseline components, daily cycle components, weekly cycle components, and seasonal fluctuation components. For example, using db4 wavelets for 4-level decomposition, wavelet coefficients at each level are obtained, and approximate and detail components are extracted, representing long-term trends, intraday fluctuations, weekly fluctuations, and seasonal fluctuations, respectively.

[0096] Step S2: Set the rolling window for grid load based on the time-frequency component data of grid load, and perform dynamic baseline value weighted combination to obtain the dynamic baseline value data of grid load;

[0097] In this embodiment of the invention, different rolling window sizes are set for the different time-frequency components (long-term trend, intraday fluctuation, weekly fluctuation, and seasonal fluctuation) decomposed in step S1. For example, a 365-day rolling window is used for the long-term trend, a 24-hour rolling window for intraday fluctuation, a 168-hour rolling window for weekly fluctuation, and a 90-day rolling window for seasonal fluctuation. Then, within the corresponding rolling window, baseline values ​​are calculated for each component. For example, for the long-term trend component, the average load on the corresponding date over the past 365 days is calculated as the baseline value; for the intraday fluctuation component, the average load at the same time over the past 24 hours is calculated as the baseline value. Finally, the baseline values ​​of each component are weighted and combined to obtain the final dynamic baseline value. The weighting coefficients can be set according to the degree of contribution of different components to load changes; for example, the weight of the long-term trend is 0.5, the weight of the intraday fluctuation is 0.2, the weight of the weekly fluctuation is 0.2, and the weight of the seasonal fluctuation is 0.1.

[0098] Step S3: Filter the standard historical multi-source power grid load data for effective historical load data and calculate the coefficient of variation to generate effective historical power grid load variation coefficients; perform sensitivity parameter analysis of early warning sources based on effective historical power grid load variation coefficients to generate optimal alarm sensitivity parameters; perform adaptive load early warning threshold processing on the dynamic baseline value data of power grid load based on effective historical power grid load variation coefficients and optimal alarm sensitivity parameters to obtain dynamic load early warning threshold curve data.

[0099] In this embodiment of the invention, the current real-time power grid monitoring timestamp is obtained. Based on the timestamp, historical load data from the same date and time over the past several years (e.g., 5 years) are selected as valid historical load data. The mean and standard deviation of the valid historical load data are calculated, and then the coefficient of variation (standard deviation / mean) is calculated. Next, the distribution characteristics of the coefficient of variation are analyzed, an initial sensitivity parameter range is set, for example [1.5, 3.0], and a traversal step size is set, for example 0.1. The warning source database is connected, historical warning records are read, and real and false alarm records are filtered out. Using different sensitivity parameters, recall and precision are calculated, and performance indicators such as the F1 value are calculated. The sensitivity parameter with the highest F1 value is selected as the optimal alarm sensitivity parameter. Finally, based on the optimal sensitivity parameter and the coefficient of variation, the upper and lower limits of the dynamic load warning threshold curve are calculated. Upper limit = dynamic baseline value + (sensitivity parameter × coefficient of variation), lower limit = dynamic baseline value - (sensitivity parameter × coefficient of variation).

[0100] Step S4: Obtain real-time power grid resource and equipment status data; construct the power grid monitoring resource topology based on the real-time power grid resource and equipment status data to generate power grid monitoring resource topology data; construct a power grid load forecasting model based on the power grid monitoring resource topology data; use the power grid load forecasting model to perform short-term load forecasting to obtain short-term load forecasting data.

[0101] In this embodiment of the invention, real-time status data of power grid resources and equipment is collected, such as voltage, current, and power data of substations obtained from the power grid dispatch and control center system. A power grid monitoring resource topology is constructed based on the real-time data, for example, using a graph data structure to represent the connection relationships between substations, lines, and loads. Based on the topology, a load forecasting model is constructed, for example, using a graph convolutional neural network (GCN) combined with a Bayesian network to incorporate topology information into the model. Historical load data and weather data are used as training data to train the model. After training, real-time load data is used as input to perform short-term load forecasting, for example, predicting the load value for the next 24 hours.

[0102] Step S5: Calculate the threshold load difference of short-term load forecast data using dynamic load early warning threshold curve data, and perform multi-level early warning triggering processing to obtain the power grid multi-level early warning triggering strategy.

[0103] In this embodiment of the invention, the short-term load forecast data obtained in step S4 is compared with the dynamic load warning threshold curve data obtained in step S3 to calculate the threshold load difference. For example, the hourly load value predicted for the next 24 hours is compared with the upper and lower limits of the corresponding dynamic warning threshold curve to calculate the difference. Next, multi-level warning triggering is performed based on the magnitude of the threshold load difference. For example, three warning levels can be set: Level 1 warning, Level 2 warning, and Level 3 warning. A Level 1 warning is triggered when the predicted load exceeds the upper limit of the warning threshold curve by a certain margin; a Level 2 warning is triggered when the predicted load exceeds the upper limit of the warning threshold curve but does not meet the Level 1 warning triggering condition; and a Level 3 warning is triggered when the predicted load is below the lower limit of the warning threshold curve by a certain margin. For example, a Level 1 warning is triggered when the predicted load exceeds the upper limit by 10%; a Level 2 warning is triggered when the predicted load exceeds the upper limit by 5% but is below 10%; and a Level 3 warning is triggered when the predicted load is below the lower limit by 5%. For example, after a Level 1 warning is triggered, an emergency plan is activated, and power rationing measures are implemented; after a Level 2 warning is triggered, relevant departments are notified to prepare; and after a Level 3 warning is triggered, load changes are closely monitored.

[0104] Preferably, step S1 includes the following steps:

[0105] Step S11: Obtain historical multi-source power grid resource load data;

[0106] Step S12: Unify the format of historical multi-source power grid resource load data from multiple data sources, and perform data preprocessing to generate standard historical multi-source power grid load data;

[0107] Step S13: Classify the historical load monitoring dates according to the standard historical multi-source power grid load data to obtain classified historical multi-source power grid load data;

[0108] Step S14: Perform Fast Fourier Transform on the classified historical multi-source power grid load data to generate historical power grid load spectrum data;

[0109] Step S15: Decompose the time-frequency components of the power grid load based on the historical power grid load spectrum data to generate the time-frequency component data of the power grid load.

[0110] In this embodiment of the invention, historical power grid resource load data is obtained from different data sources. These data sources may include the power grid dispatch and control center database, smart substation data acquisition systems, distribution automation systems, and Advanced Metering Systems (AMI). The acquired data types include power load information for different voltage levels and regions, such as active and reactive power of 110kV substations, 35kV substations, and 10kV feeders. The time granularity of the data can be 15 minutes, 30 minutes, or 1 hour. For example, hourly active and reactive power data covering all 110kV substations, 35kV substations, and some 10kV feeders in a city over the past two years can be obtained. All data is converted into a unified CSV file format, containing fields such as timestamp, substation / feeder ID, voltage level, active power, and reactive power. After format unification, the data is preprocessed, including handling missing values ​​and outliers. For missing values, methods such as linear interpolation, spline interpolation, or mean imputation can be used. Outliers can be identified and removed using the Raida criterion or the 3σ criterion, or replaced with reasonable values. The standard historical multi-source power grid load data is classified according to date attributes. Classification criteria can include weekdays, weekends, holidays, and seasons. For example, dates throughout the year can be divided into weekdays, weekends, and statutory holidays such as Spring Festival and National Day. Data can also be classified into spring, summer, autumn, and winter based on month or temperature. The classified historical multi-source power grid load data obtained in step S13 is processed using Fast Fourier Transform (FFT). FFT converts time-domain signals into frequency-domain signals, thereby analyzing the amplitude and phase information of different frequency components in the load data. For example, performing FFT on the classified weekday load data yields the amplitudes of different frequency components, such as the amplitudes of daily (24-hour) and weekly (168-hour) frequency components. Time-frequency component decomposition can decompose the original load data into fluctuation components of different frequencies, such as baseline components, daily cycle components, weekly cycle components, and random fluctuation components. For example, wavelet transform can be used to decompose weekday load data into baseline components, daily periodic components, and random fluctuation components.

[0111] Preferably, step S15 includes the following steps:

[0112] Step S151: Analyze the frequency range of major loads based on historical power grid load spectrum data to obtain historical load frequency range data;

[0113] Step S152: Perform maximum wavelet decomposition layer processing based on historical load frequency range data to obtain the maximum wavelet decomposition layer;

[0114] Step S153: Divide the frequency range of the decomposition level according to the maximum wavelet decomposition level, and perform decomposition level mapping on the historical load frequency range data to obtain load frequency-decomposition level mapping data.

[0115] Step S154: Determine the coverage range of the decomposition layer for the load frequency-decomposition level mapping data, and optimize the wavelet decomposition level to generate the optimal decomposition layer data;

[0116] Step S155: Perform wavelet transform decomposition on historical power grid load spectrum data based on the optimal decomposition level data to generate wavelet decomposition coefficients;

[0117] Step S156: Extract wavelet approximate components based on wavelet decomposition coefficients to obtain long-term trend data of power grid load;

[0118] Step S157: Extract wavelet detail components based on wavelet decomposition coefficients to obtain intraday power grid load fluctuation data, weekly power grid load fluctuation data, and seasonal power grid load fluctuation data, respectively.

[0119] Step S158: Combine the long-term trend data of power grid load, the intraday fluctuation data of power grid load, the intraweek fluctuation data of power grid load, and the seasonal fluctuation data of power grid load with the time-frequency component of power grid resource monitoring to obtain the time-frequency component data of power grid load.

[0120] In this embodiment of the invention, the spectral data corresponding to each date type is analyzed. For example, for the spectral data of weekdays, its frequency distribution characteristics can be observed by plotting a spectrum graph. In the spectrum graph, the horizontal axis represents frequency, and the vertical axis represents the energy or amplitude of the corresponding frequency. By observing the spectrum graph, frequency bands with concentrated energy can be identified. These frequency bands with concentrated energy usually correspond to the main frequency components of load changes. Suppose the analysis finds that the energy of the weekday load is mainly concentrated in three frequency bands: the first frequency band is the low frequency band, corresponding to a cycle of about 24 hours (i.e., daily load changes); the second frequency band is the mid frequency band, corresponding to a cycle of about 6-8 hours (i.e., peak electricity consumption); and the third frequency band is the high frequency band, corresponding to a cycle of less than 1 hour (i.e., some sudden electricity consumption events). Then, the frequency range of these main frequency bands is recorded. For example, the frequency range of the low frequency band is 0-0.0417Hz (corresponding to a 24-hour cycle), the frequency range of the mid frequency band is 0.125-0.1667Hz (corresponding to a 6-8 hour cycle), and the frequency range of the high frequency band is above 1Hz. Similar analysis is performed on the spectral data for weekends and holidays to obtain their respective main frequency ranges. The lowest frequency value in the historical load frequency range data is analyzed; this lowest frequency typically corresponds to the long-term trend of the load signal, such as the daily load cycle. In step S151, the lowest frequency on weekdays is 0.0417Hz (corresponding to 24 hours), and the lowest frequency on weekends is 0.035Hz. Then, a suitable wavelet basis function is selected; for example, the Daubechies wavelet (dbN) can be used. Each level of the wavelet transform halves the frequency range; therefore, a suitable number of decomposition levels needs to be selected so that the lowest frequency after decomposition covers the lowest frequency in the historical load frequency range data. The maximum number of decomposition levels can be calculated using the following formula: number of levels = ceil(log2(Fs / f_min)). Where Fs represents the sampling frequency, f_min represents the lowest frequency value, and the ceil function represents rounding up. The frequency range is divided according to the maximum number of wavelet decomposition levels. Assuming the maximum number of wavelet decomposition levels is 6, then each level of wavelet decomposition halves the frequency range. Starting from the maximum frequency (half of the sampling frequency, i.e., the Nyquist frequency), the frequency is decomposed layer by layer. For a sampling frequency of once every 15 minutes (0.0011Hz), the corresponding Nyquist frequency is 0.00055Hz. Therefore, the frequency range of the first layer can be set to 0.00055Hz / 2 to 0.00055Hz, the frequency range of the second layer is 0.00055Hz / 4 to 0.00055Hz / 2, and so on, with the frequency range of the sixth layer being 0.00055Hz / 64 to 0.00055Hz / 32. Then, the historical load frequency range data obtained in step S151 needs to be mapped to these decomposition levels.For each historical load frequency range (e.g., low frequency 0-0.0417Hz, mid frequency 0.125-0.1667Hz, and high frequency above 1Hz for weekdays), it is necessary to determine which decomposition level(s) it falls within. Analyze the load frequency-decomposition level mapping data obtained in step S153 to determine the actual number of wavelet decomposition levels required. To avoid computational burden and noise interference from over-decomposition, the number of decomposition levels needs to be optimized. The minimum number of decomposition levels covering the main load frequency range can be selected as the optimal number of decomposition levels. For example, if the main load frequency range maps to levels 5 to 8, then 4 can be selected as the optimal number of decomposition levels, decomposing only to level 4 to cover the target frequency range. Select a suitable wavelet basis function, for example, the Daubechies wavelet (dbN), where N can be chosen according to the actual situation. Then, for each date type of load data, such as weekdays, weekends, and holidays, perform wavelet transform decomposition separately. For example, for weekday load data, firstly, based on the optimal decomposition level data obtained in step S154, determine the number of levels to be decomposed, assuming it to be level 1, level 3, and level 5. Then, use the selected wavelet basis function to perform wavelet decomposition on the load data. Wavelet decomposition produces two parts of coefficients: approximation coefficients (cA) and detail coefficients (cD). The decomposition of level 1 yields cA1 and cD1; the decomposition of level 3 is based on cA1 obtained from level 1, yielding cA3 and cD3; and the decomposition of level 5 is based on cA3 obtained from level 3, yielding cA5 and cD5. According to the decomposition level determination logic in step S154, only cD1, cD3, and cD5 are retained as decomposition coefficients for different frequency bands. For each date type (e.g., weekday, weekend, holiday), approximation components need to be extracted from its wavelet decomposition coefficients. In step S155, assuming the decomposition uses layers 1, 3, and 5, cA1, cA3, and cA5 are obtained, along with corresponding detail coefficients cD1, cD3, and cD5. cA5 is an approximate coefficient obtained after five layers of decomposition, containing lower-frequency information and can be considered as the long-term load trend; therefore, cA5 needs to be extracted as long-term load trend data. For different load types, since the decomposition levels differ, reconstruction using the corresponding decomposition level and approximate coefficients is required to obtain the corresponding wavelet approximate components. Wavelet reconstruction can be achieved by performing an inverse wavelet transform on the approximate coefficients, converting the frequency domain signal back to the time domain signal to obtain the actual load value. Based on different frequency ranges, the corresponding detail coefficients are extracted. In step S155, assuming wavelet decomposition using layers 1, 3, and 5, cD1, cD3, and cD5 are obtained.Among them, cD1 reflects the highest frequency fluctuations, such as the impact of some sudden events, and can therefore be considered as intraday fluctuations; cD3 reflects medium frequency fluctuations, such as the load difference between weekdays and weekends, and can be considered as weekly fluctuations; cD5 reflects low frequency fluctuations, such as load changes in different seasons, and can be considered as seasonal fluctuations. The component data extracted in steps S156 and S157 are combined to form the final time-frequency component data of the power grid load. These component data include long-term trend data, intraday fluctuation data, weekly fluctuation data, and seasonal fluctuation data. These component data can be stored in a data structure, such as a multidimensional array or a data frame containing multiple time series. For example, long-term trend data, intraday fluctuation data, weekly fluctuation data, and seasonal fluctuation data can be stored as different columns, with timestamps as indexes, forming a complete data table.

[0121] Preferably, step S2 includes the following steps:

[0122] Step S21: Set the rolling window for grid load based on the time-frequency component data of grid load, and obtain the annual trend baseline rolling window, intraday fluctuation baseline rolling window, weekly fluctuation baseline rolling window and seasonal fluctuation baseline rolling window respectively;

[0123] Step S22: Calculate the long-term trend baseline of the power grid load using the annual trend baseline rolling window to generate long-term trend baseline data;

[0124] Step S23: Calculate the intraday fluctuation baseline of the grid load intraday fluctuation data through the intraday fluctuation baseline rolling window to generate intraday fluctuation baseline data;

[0125] Step S24: Calculate the weekly fluctuation baseline of the grid load within the week using the weekly fluctuation baseline rolling window to generate weekly fluctuation baseline data;

[0126] Step S25: Calculate the seasonal fluctuation baseline for the power grid load seasonal fluctuation data using the seasonal fluctuation baseline rolling window to generate seasonal fluctuation baseline data;

[0127] Step S26: Perform a dynamic baseline value weighted combination based on the long-term trend baseline data, intraday fluctuation baseline data, weekly fluctuation baseline data, and seasonal fluctuation baseline data to obtain the dynamic baseline value data of the power grid load.

[0128] In this embodiment of the invention, different rolling windows are set according to the characteristics of the time-frequency component data of the power grid load, respectively used to calculate the annual trend baseline, intraday fluctuation baseline, weekly fluctuation baseline, and seasonal fluctuation baseline. The size of the rolling window needs to be adjusted according to the periodic characteristics of different components. For example, a larger rolling window, such as 365 days, can be set for the annual trend baseline; a 24-hour rolling window can be set for the intraday fluctuation baseline; a 168-hour rolling window can be set for the weekly fluctuation baseline; and a 90-day rolling window can be set for the seasonal fluctuation baseline, representing the four seasons of spring, summer, autumn, and winter, respectively. These rolling windows can be adjusted according to actual conditions, for example, they can be fine-tuned according to the changing patterns of historical data and the sensitivity requirements of early warning. Using the annual trend baseline rolling window determined in step S21, the long-term trend data of the power grid load is calculated as a baseline. For example, a 365-day rolling window is used to calculate the long-term trend baseline value for each day. The calculation method for the baseline value can be the moving average method, the moving median method, etc. For example, the average value of the load data of the past 365 days can be used as the long-term trend baseline value at the current moment. Using the intraday fluctuation baseline rolling window determined in step S21, baseline calculations are performed on the intraday fluctuation data of the power grid load. For example, a 24-hour rolling window is used to calculate the intraday fluctuation baseline value for each hour. The calculation method for the baseline value can be selected according to the characteristics of the intraday fluctuation. For example, the average load data at the same time in the past 24 hours can be calculated as the intraday fluctuation baseline value for the current time, thereby capturing the periodic variation pattern of intraday load. Using the weekly fluctuation baseline rolling window determined in step S21, baseline calculations are performed on the weekly fluctuation data of the power grid load. For example, a 168-hour rolling window is used to calculate the weekly fluctuation baseline value for each hour. The average load data at the same time on the same day of the past week can be calculated as the weekly fluctuation baseline value for the current time, thereby capturing the difference between weekday and weekend loads. Using the seasonal fluctuation baseline rolling window determined in step S21, baseline calculations are performed on the seasonal fluctuation data of the power grid load. For example, a 90-day rolling window is used to calculate the seasonal fluctuation baseline value for each day. The average load on the same date in the past 90 days can be calculated as the seasonal fluctuation baseline value for the current date to capture the variation pattern of load in different seasons. The baseline data obtained in steps S22 to S25 are weighted and combined to obtain the final dynamic baseline value of the power grid load. The weighting coefficients can be set according to the contribution of different components to load changes and can be adjusted according to actual conditions. For example, based on historical data analysis, the weights of the long-term trend baseline, intraday fluctuation baseline, weekly fluctuation baseline, and seasonal fluctuation baseline can be determined to be 0.5, 0.2, 0.2, and 0.1, respectively. The final dynamic baseline value is obtained by multiplying the baseline value of each component by its corresponding weight and then summing the results.

[0129] Preferably, step S3 includes the following steps:

[0130] Step S31: Obtain the real-time power grid monitoring timestamp;

[0131] Step S32: Filter the standard historical multi-source power grid load data for effective historical load data by using the real-time power grid monitoring timestamp, and calculate the coefficient of variation to generate the effective historical load variation coefficient of the power grid.

[0132] Step S33: Analyze the sensitivity parameters of the early warning source based on the effective historical load variation coefficient of the power grid, and generate the optimal alarm sensitivity parameters;

[0133] Step S34: Calculate the load warning threshold based on the effective historical load variation coefficient of the power grid and the optimal alarm sensitivity parameter, and apply adaptive threshold constraints to obtain dynamic load warning threshold curve data.

[0134] In this embodiment of the invention, the real-time timestamp of the current power grid monitoring is obtained. This timestamp is used to filter out valid historical load data corresponding to the current time from historical data. Based on the real-time power grid monitoring timestamp, the time range to be filtered from the historical data is determined. To ensure data validity, a time period similar to the current time is usually selected. For example, data from the same time period before and after each day over the past three years are selected as valid historical load data. During the filtering process, abnormal data is also excluded; for example, data points exceeding plus or minus three standard deviations of the historical data average are removed. For the filtered valid historical load data, the coefficient of variation needs to be calculated. The coefficient of variation reflects the dispersion of the data, and the calculation formula is: Coefficient of Variation = Standard Deviation / Average. For each set of valid historical load data, the average and standard deviation of the data set are calculated, and then the coefficient of variation is calculated to form the coefficient of variation of the valid power grid historical load. The range and step size of the sensitivity parameter are defined. For example, the sensitivity parameter can be defined as between 1.0 and 2.0, with a step size of 0.1. Then, different sensitivity parameters need to be set for different coefficient of variation ranges. For example, the coefficient of variation can be divided into three intervals: low coefficient of variation (e.g., less than 0.1), medium coefficient of variation (e.g., between 0.1 and 0.3), and high coefficient of variation (e.g., greater than 0.3). A sensitivity coefficient k can be set and multiplied by the coefficient of variation to obtain the fluctuation range of the warning threshold. By analyzing the warning effect under different sensitivity parameters, such as false alarm rate and false negative rate, the optimal alarm sensitivity parameter can be selected. For example, by simulating different values ​​of k such as 1.5, 2.0, and 2.5, and combining backtesting with known anomalies in historical data, the k value with both low false alarm rate and low false negative rate can be selected as the optimal sensitivity parameter. Based on the effective historical load variation coefficient of the power grid calculated in step S32 and the optimal alarm sensitivity parameter k obtained in step S33, the load warning threshold is calculated for the dynamic baseline value data of the power grid load obtained in step S26. The upper and lower limits of the dynamic load warning threshold curve are obtained by adding and subtracting the product of the sensitivity parameter k and the coefficient of variation from the dynamic baseline value. For example, adding k multiplied by the coefficient of variation to the dynamic baseline value yields the upper limit of the warning threshold curve; subtracting k multiplied by the coefficient of variation from the dynamic baseline value yields the lower limit. To avoid thresholds being too large or too small, an adaptive threshold constraint can be set. For instance, a maximum and minimum threshold value can be set to limit the calculated threshold within a reasonable range. For example, the maximum and minimum threshold values ​​can be set based on historical extreme values ​​or empirical values. The final dynamic load warning threshold curve data includes both upper and lower limit curves.

[0135] Preferably, step S32 includes the following steps:

[0136] Step S321: Select the historical effective load window based on the real-time power grid monitoring timestamp to obtain the effective historical load window data;

[0137] Step S322: Filter the standard historical multi-source power grid load data using the effective historical load window data to obtain the effective historical load data of the power grid;

[0138] Step S323: Calculate the historical load average based on the effective historical load data of the power grid to generate effective historical load average data;

[0139] Step S324: Calculate the historical load standard deviation based on the effective historical load data of the power grid to generate effective historical standard deviation data;

[0140] Step S325: Calculate the coefficient of variation based on the effective historical load mean data and the effective historical standard deviation data to generate the effective historical load variation coefficient of the power grid.

[0141] In this embodiment of the invention, the real-time power grid monitoring timestamps are analyzed to extract date and time information. Then, the width of a historical effective load window is defined; for example, it can be set to a period of time near the same time point of the previous day or several days. To ensure data representativeness, data from the same date and time over the past few years are typically selected. Assume the selected window width is 1 hour before and after, and data from the same day over the past three years are selected. For example, if load fluctuations are rapid, a narrower window can be selected; if load fluctuations are slow, a wider window can be selected. The standard historical multi-source power grid load data is traversed, checking whether the timestamp of each load data point falls within any of the aforementioned time windows. If the timestamp of a load data point falls within a time window, that load data is filtered out and added to the effective historical load dataset. During the filtering process, outliers also need to be excluded, for example, data exceeding the average by plus or minus three standard deviations, to avoid outliers affecting subsequent coefficient of variation calculations. After excluding outliers, all filtered effective historical load data are stored, for example, in a list, where each element represents one effective historical load data point. The average value of all load data within each of the three time windows is calculated separately. Calculate an average value for each window. Then, these three average values ​​can be combined to form an overall average value, which is the final effective historical load average, or these three average values ​​can be retained directly. To improve data stability, the calculated average value can be smoothed, for example, using a moving average method. For example, the average of the average values ​​of each time window and the two time windows before and after it can be calculated as the final average value. Similar to step S323, calculate the standard deviation for the load data within each time window. Calculate the standard deviation for all load data within these three time windows. Calculate a standard deviation for each window. Then, these three standard deviations can be combined to form an overall standard deviation, which is the final effective historical standard deviation, or these three standard deviations can be retained directly. To improve data stability, the calculated standard deviation can be smoothed, for example, using a moving average method. For example, the average of the standard deviations of each time window and the two time windows before and after it can be calculated as the final standard deviation. For each time window, or the overall mean and standard deviation, calculate the coefficient of variation using the formula: Coefficient of Variation = Standard Deviation / Average. For example, for each time window, divide the standard deviation of that time window by its mean to obtain the coefficient of variation for that time window. Alternatively, divide the overall standard deviation by the overall mean to obtain a coefficient of variation.

[0142] Preferably, step S33 includes the following steps:

[0143] Step S331: Analyze the distribution characteristics of the coefficient of variation based on the effective historical load variation coefficient of the power grid, and set the initial sensitivity parameters to obtain the range data of the initial sensitivity parameters;

[0144] Step S332: Obtain power grid resource load early warning source configuration data;

[0145] Step S333: Connect the early warning source according to the power grid resource load early warning source configuration data, and read the historical early warning records to obtain the original power grid early warning record data;

[0146] Step S334: Filter the actual alarm records based on the original power grid early warning record data to obtain the actual power grid alarm record data;

[0147] Step S335: Filter false alarm records based on the original power grid early warning record data to obtain power grid false alarm record data;

[0148] Step S336: Use the preset sensitivity traversal step size data to iterate through the initial sensitivity parameter range data to obtain the traversal sensitivity parameters;

[0149] Step S337: Calculate the alarm recall rate by traversing the sensitivity parameters to obtain the power grid alarm recall rate data;

[0150] Step S338: Calculate the alarm accuracy by iterating through the sensitivity parameters to obtain the power grid alarm accuracy data by analyzing the actual alarm record data and the false alarm record data.

[0151] Step S339: Calculate the early warning performance index of the traversal sensitivity parameter based on the power grid alarm precision rate data and the power grid alarm recall rate data, and select the optimal sensitivity parameter to obtain the optimal alarm sensitivity parameter.

[0152] In this embodiment of the invention, the distribution characteristics of these coefficients of variation are analyzed. For example, the minimum, maximum, average, median, and standard deviation of the coefficients of variation are calculated, and histograms or box plots of the coefficients of variation are plotted to observe their distribution. For example, the analysis may suggest that the minimum coefficient of variation is close to 0, the maximum is close to 0.5, the average is close to 0.2, and most coefficients of variation are distributed between 0.1 and 0.3. Then, based on the analysis results, an initial sensitivity parameter range is set. The sensitivity parameter is typically used to adjust the warning threshold; the larger the coefficient of variation, the smaller the sensitivity parameter should be, and vice versa. For example, the initial sensitivity parameter range can be set between 1.0 and 2.0, or this range can be adjusted according to the actual situation. The sensitivity parameter range can be divided into multiple intervals based on the distribution of the coefficients of variation. The configuration data of the power grid resource load warning source is obtained. This configuration data includes the connection information of the warning source, such as the database address, username, and password, as well as relevant parameters of the warning system, such as the classification criteria for warning levels and the method of sending warning information. The connection method can be either an API-based connection or a database-based connection, depending on the type of early warning system. Assuming the early warning system provides an HTTP-based API, the Python requests library can be used for connection and data requests. Then, historical early warning records are read. This data contains information about historical early warning events, such as warning time, warning level, and warning reason. For example, early warning records from the past year can be read from the database, including warning time, warning level (Level 1, 2, or 3), and warning reason (e.g., overload or underload). Real alarm records are filtered from the raw power grid early warning record data obtained in step S333. Real alarms refer to early warning events where actual power grid anomalies have occurred. This requires judgment based on historical power grid operation data and expert experience. For example, information such as power outages and equipment failures recorded in historical records can be used to determine which early warning records are related to actual power grid anomalies. False alarm records are also filtered from the raw power grid early warning record data obtained in step S333. False alarms refer to early warnings issued by the system when no actual power grid anomalies occurred. This also requires judgment based on historical power grid operation data and expert experience. For example, if a warning event occurs but the power grid operates normally and no equipment malfunctions or abnormalities are detected, it can be judged as a false alarm. Using a preset sensitivity traversal step size, the initial sensitivity parameter range obtained in step S331 is iterated through to generate a series of traversed sensitivity parameters. For example, assuming the initial sensitivity parameter range is [1.5, 3.0] and the preset traversal step size is 0.1, the generated traversed sensitivity parameter sequence is 1.5, 1.6, 1.7, ..., 2.9, 3.0.Using the traversal sensitivity parameters generated in step S336, alarm recall is calculated for both real and false alarm records of the power grid. Alarm recall refers to the proportion of real alarm events successfully detected by the early warning system. For each traversal sensitivity parameter, using historical real alarm events as a baseline, the number of real alarm events successfully detected by the early warning system under that sensitivity parameter is calculated, and then divided by the total number of real alarm events to obtain the alarm recall corresponding to that sensitivity parameter. Using the traversal sensitivity parameters generated in step S336, alarm precision is calculated for both real and false alarm records of the power grid. Alarm precision refers to the proportion of all early warnings issued by the early warning system that are actually real alarms. For each traversal sensitivity parameter, the number of all early warnings issued by the early warning system under that sensitivity parameter, and the number of real alarms among them, are calculated, and then the number of real alarms is divided by the number of all early warnings to obtain the alarm precision corresponding to that sensitivity parameter. Define one or more early warning performance indicators. Commonly used early warning performance metrics include F1 score, F2 score, or other custom metrics. The F1 score is the harmonic mean of recall and precision, calculated as: F1 = 2 × (Recall × Precision) / (Recall + Precision). The F2 score is the harmonic mean weighted by recall, calculated as: F2 = 5 × (Recall × Precision) / (4 × Precision + Recall). Depending on the specific needs, you can choose the F1 score or a custom weighted combination to suit different requirements. For example, if you prioritize coverage of actual alarms, you can choose a higher recall weight; if you prioritize accuracy, you can choose a higher precision weight. Then, iterate through all sensitivity parameters and calculate the corresponding early warning performance metric value for each sensitivity parameter, such as the F1 score or custom metric value for each sensitivity parameter. For example, suppose the calculated F1 value for sensitivity parameter 1.5 is 0.6, 1.6 is 0.7, 1.7 is 0.8, 1.8 is 0.85, 1.9 is 0.9, and 2.0 is 0.8. Compare the warning performance index values ​​corresponding to all sensitivity parameters, and select the sensitivity parameter with the largest warning performance index value as the optimal alarm sensitivity parameter.

[0153] As an example of the present invention, reference is made to... Figure 2 As shown, Figure 1 A detailed flowchart illustrating the implementation steps of step S4 is provided in this example. Step S4 includes:

[0154] Step S41: Obtain real-time power grid resource equipment status data;

[0155] In this embodiment of the invention, real-time status data of power grid resources and equipment is collected from various data sources. These data sources include, but are not limited to, SCADA systems, PMUs (phasor measurement units), smart meters, and other sensor devices. The types of data collected cover various key parameters reflecting the operating status of the power grid, such as transformer load rate, line current, voltage, frequency, active power, and reactive power. The data collection frequency can be adjusted according to actual needs, for example, data can be collected once per second, per minute, or per hour. For example, real-time active power, reactive power, voltage, and current data of all 110kV substations can be collected once per minute from the SCADA system; simultaneously, voltage and current phasor data can be collected once per second from PMUs deployed on critical lines. In addition, real-time electricity consumption data of users can also be collected from smart meters. The collected data needs to be quality verified, for example, checking whether the data is complete and whether there are outliers. For missing or erroneous data, appropriate processing is required, such as using interpolation methods to fill missing values ​​or using filtering algorithms to remove outliers.

[0156] Step S42: Construct the power grid monitoring resource topology based on real-time power grid resource equipment status data, and generate power grid monitoring resource topology data;

[0157] In this embodiment of the invention, a power grid monitoring resource topology is constructed based on the real-time power grid resource equipment status data obtained in step S41. The topology describes the connection relationships between various devices in the power grid, such as the connections between substations, lines, and loads. The topology can be constructed using graphical modeling tools or data structures, such as adjacency matrices or adjacency lists. For example, a topology graph containing all 110kV substations, 35kV substations, 10kV feeders, and user loads can be constructed. In the topology graph, nodes represent substations, feeders, and users, and edges represent line connections. Each node and edge contains its corresponding attribute information; for example, a substation node contains its name, voltage level, and geographical location; a line edge contains its length and impedance.

[0158] Step S43: Collect real-time load data of power grid monitoring nodes based on power grid monitoring resource topology data to obtain real-time load data of power grid resources;

[0159] In this embodiment of the invention, real-time load data is collected from relevant monitoring nodes based on the power grid monitoring resource topology data constructed in step S42. Monitoring nodes can be substations, distribution lines, etc. For example, based on the topology data, it can be determined which substations and distribution lines require load data collection. Then, real-time load data of these nodes, such as active power and reactive power, is read from data sources such as SCADA systems and smart meters. The collected data needs to be associated with the topology data; for example, the collected load data is associated with the corresponding substation or distribution line nodes. This allows the load data to be analyzed in conjunction with the power grid topology; for example, the total load of each substation and the load flow of each line can be calculated.

[0160] Step S44: Perform Bayesian network node processing based on power grid monitoring resource topology data, and construct a power grid load prediction model through a preset convolutional neural network model to obtain the power grid load prediction model to be trained.

[0161] In this embodiment of the invention, Bayesian network node processing is applied to the topology data of power grid monitoring resources. A Bayesian network is a probabilistic graphical model that can be used to represent dependencies between variables. Nodes in the power grid topology (such as substations and transformers) can be considered as nodes in a Bayesian network, and the connections between nodes can be considered as edges. For each node, a conditional probability distribution can be constructed based on its historical load data and other relevant factors (such as temperature and humidity) to reflect the influence of other nodes and environmental factors on the node's load. For example, a Bayesian network can be established where the load of a transformer node is affected by its upstream nodes and factors such as temperature and humidity. The parameters in this Bayesian network can be learned using historical data. The output of the Bayesian network is the load prediction value and corresponding confidence level for each node. Secondly, a convolutional neural network (CNN) model needs to be pre-set for load prediction. CNN models excel at processing data with spatial structure and are very suitable for analyzing power grid topology data. The input to the CNN model can be the historical load information of each node output by the Bayesian network, as well as real-time topology data; the output is the load prediction value of each node over a future period. A CNN model's structure can include convolutional layers, pooling layers, and fully connected layers. For example, a CNN model with multiple convolutional and pooling layers can be designed to extract features from the topology, and then fully connected layers can be used for load prediction. For instance, the following configuration could be used: the first layer is a convolutional layer with 32 kernels of 3x3 size and ReLU activation; the second layer is a pooling layer using max pooling with a 2x2 window; the third layer is a convolutional layer with 64 kernels of 3x3 size and ReLU activation; the fourth layer is a pooling layer using max pooling with a 2x2 window; and the fifth layer is a fully connected layer, with the number of output layers equal to the number of predicted loads. Finally, the output of the Bayesian network is used as input to the CNN model to construct a power grid load prediction model to be trained.

[0162] Step S45: Process the standard historical multi-source power grid load data into a model training set, and train the power grid load prediction model to be trained to obtain the power grid load prediction model;

[0163] In this embodiment of the invention, standard historical multi-source power grid load data is processed into a model training set. This includes steps such as data cleaning, feature engineering, and data normalization. For example, historical load data can be arranged in chronological order, and features related to load forecasting can be extracted, such as date type (weekday, weekend, holiday), weather data (temperature, humidity), and historical load values. Then, these features are normalized, for example, by scaling the feature values ​​to the range of [0,1] to improve the efficiency and stability of model training. The processed training set is then input into the power grid load forecasting model to be trained, constructed in step S44, for training. During training, the model learns the patterns of load changes based on the training data and adjusts the model parameters to minimize prediction errors. Various optimization algorithms can be used to train the model, such as stochastic gradient descent (SGD) and the Adam optimizer. After training, a trained power grid load forecasting model is obtained, which can be used for short-term load forecasting.

[0164] Step S46: Use the power grid load forecasting model to perform short-term load forecasting on the real-time load data of power grid resources to obtain short-term load forecasting data.

[0165] In this embodiment of the invention, the real-time load data of the power grid resources obtained in step S43 is used as input to the power grid load prediction model trained in step S45 for short-term load prediction. For example, the load value for the next 24 hours can be predicted. The prediction results are output in the form of a time series, for example, the predicted load value for each hour within the next 24 hours can be output. This prediction data will be used for subsequent comparison of warning thresholds and warning triggering. The prediction duration can be adjusted according to actual needs, for example, the load value for the next 12 hours, 48 ​​hours or longer can be predicted.

[0166] As an example of the present invention, reference is made to... Figure 3 As shown, Figure 1 A detailed flowchart illustrating the implementation steps of step S5 is provided in this example. Step S5 includes:

[0167] Step S51: Extract the forecast time points from the short-term load forecast data to obtain the power grid resource forecast time point data;

[0168] In this embodiment of the invention, short-term load forecast data is typically stored in time series format, containing forecast time points and corresponding forecast load values. For example, if the load for the next 24 hours is forecasted, then each hour represents a forecast time point. The extracted power grid resource forecast time point data is a timestamp sequence.

[0169] Step S52: Identify the real-time load warning threshold range of the dynamic load warning threshold curve data using the power grid resource prediction time point data, and calculate the threshold load difference of the short-term load prediction data to obtain the predicted load difference data;

[0170] In this embodiment of the invention, the prediction time point data is traversed. For each prediction time point, the corresponding warning threshold needs to be found in the dynamic load warning threshold curve data. If the threshold curve data does not exactly match the prediction time point, the closest warning threshold data can be selected, or the corresponding warning threshold can be calculated using an interpolation method. Then, based on the warning threshold, the difference between the short-term load prediction data for that time point and the corresponding warning threshold is calculated. The predicted load difference for each time point and each monitoring node is stored, for example, as a dictionary, with the key being the time point and the monitoring node, and the value being the corresponding load difference.

[0171] Step S53: Calculate the load over-limit range based on the predicted load difference data to obtain the predicted load over-limit range data of the power grid;

[0172] In this embodiment of the invention, the load over-limit magnitude is calculated based on the predicted load difference data obtained in step S52. The load over-limit magnitude refers to the percentage by which the predicted load value exceeds the warning threshold range. For example, if the predicted load difference at a given time point is 100MW, and the corresponding upper threshold is 1000MW, then the load over-limit magnitude at that time point is (100 / 1000)×100% = 10%. If the predicted load value is lower than the lower threshold, the load over-limit magnitude is negative. For example, if the predicted load value is 700MW, and the lower threshold is 800MW, then the load over-limit magnitude is (-100 / 800)×100% = -12.5%.

[0173] Step S54: Divide the early warning levels according to the power grid predicted load over-limit data, and perform multi-level early warning triggering processing through the preset early warning level execution strategy to obtain the power grid multi-level early warning triggering strategy.

[0174] In this embodiment of the invention, the predicted load over-limit data of the power grid calculated in step S53 is used to classify the early warning levels. Early warning levels are typically divided into multiple levels, such as Level 1, Level 2, and Level 3. The criteria for classifying early warning levels can be set according to the magnitude of the load over-limit. For example, the following early warning rules can be set: when the load over-limit exceeds 10%, a Level 1 early warning is triggered; when the load over-limit is between 5% and 10%, a Level 2 early warning is triggered; and when the load over-limit is less than -5%, a Level 3 early warning is triggered. Multi-level early warning triggering is performed according to the preset early warning level execution strategy. For example, after a Level 1 early warning is triggered, an emergency plan is activated, and power rationing measures are taken; after a Level 2 early warning is triggered, relevant departments are notified to prepare; and after a Level 3 early warning is triggered, load changes are closely monitored. The early warning level execution strategy can be adjusted according to the actual situation; for example, different emergency measures can be set according to the severity of different early warning levels.

[0175] Preferably, the present invention also provides a power grid resource dynamic threshold early warning system based on multi-source data, which executes the power grid resource dynamic threshold early warning method based on multi-source data as described above. The power grid resource dynamic threshold early warning system based on multi-source data includes:

[0176] The historical load time-frequency decomposition module is used to acquire standard historical multi-source power grid load data; classify the standard historical multi-source power grid load data by historical load monitoring date; and perform time-frequency component decomposition of power grid load to generate power grid load time-frequency component data.

[0177] The load dynamic baseline construction module is used to set the grid load rolling window based on the grid load time-frequency component data and to perform dynamic baseline value weighting combination to obtain grid load dynamic baseline value data.

[0178] The adaptive threshold generation module is used to filter effective historical load data from standard historical multi-source power grid load data and calculate the coefficient of variation to generate effective historical load variation coefficients; it analyzes the sensitivity parameters of the early warning source based on the effective historical load variation coefficients to generate optimal alarm sensitivity parameters; and it performs adaptive load early warning threshold processing on the dynamic baseline value data of power grid load based on the effective historical load variation coefficients and the optimal alarm sensitivity parameters to obtain dynamic load early warning threshold curve data.

[0179] The short-term load forecasting module is used to acquire real-time power grid resource and equipment status data; construct the power grid monitoring resource topology based on the real-time power grid resource and equipment status data to generate power grid monitoring resource topology data; construct a power grid load forecasting model based on the power grid monitoring resource topology data; and use the power grid load forecasting model to perform short-term load forecasting to obtain short-term load forecasting data.

[0180] The multi-level early warning triggering module is used to calculate the threshold load difference of short-term load forecast data through dynamic load early warning threshold curve data, and to perform multi-level early warning triggering processing to obtain the power grid multi-level early warning triggering strategy.

[0181] This application addresses the issue of extracting periodic patterns and random fluctuations in load changes by performing time-frequency component decomposition on historical multi-source power grid load data. This allows the early warning method to more accurately capture the dynamic characteristics of load changes. By classifying historical load data by monitoring date, differentiating load patterns for different date types, and constructing corresponding early warning models for each, the accuracy of early warnings is further improved, avoiding the problem of a single model being unable to adapt to differences in load patterns across different dates. A rolling window mechanism and weighted combination algorithm are used to construct a dynamic baseline value, enabling real-time tracking of load change trends and adaptive adjustment of the baseline level based on changes in historical data. This effectively solves the problem that fixed threshold methods cannot adapt to periodic load fluctuations and long-term trend changes, making the early warning threshold no longer a static value but a curve that can dynamically adjust with load changes, thus better reflecting actual load changes and reducing false alarms and missed alarms. Based on the coefficient of variation of historical load data, sensitivity parameters for early warning sources are analyzed, and the optimal alarm sensitivity parameters are ultimately determined. The sensitivity of the early warning is adaptively adjusted according to the load fluctuation characteristics of different regions and time periods. For areas or time periods with large load fluctuations, the system will automatically increase the warning sensitivity to detect potential risks in a timely manner; while for areas or time periods with small load fluctuations, the system will reduce the warning sensitivity accordingly to avoid excessive warnings and reduce unnecessary maintenance work.

[0182] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0183] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for dynamic threshold early warning of power grid resources based on multi-source data, characterized in that, Includes the following steps: Step S1: Obtain standard historical multi-source power grid load data; The standard historical multi-source power grid load data is classified by historical load monitoring date, and the power grid load time-frequency component is decomposed to generate power grid load time-frequency component data; Step S2: Set the grid load rolling window based on the grid load time-frequency component data, and perform dynamic baseline value weighted combination to obtain grid load dynamic baseline value data; Step S3: Filter the standard historical multi-source power grid load data for effective historical load data and calculate the coefficient of variation to generate effective historical power grid load variation coefficients; analyze the sensitivity parameters of the early warning source based on the effective historical power grid load variation coefficients to generate optimal alarm sensitivity parameters; perform adaptive load early warning threshold processing on the dynamic baseline value data of the power grid load based on the effective historical power grid load variation coefficients and the optimal alarm sensitivity parameters to obtain dynamic load early warning threshold curve data; wherein, step S3 includes: Step S31: Obtain the real-time power grid monitoring timestamp; Step S32: Filter the standard historical multi-source power grid load data for validity using real-time power grid monitoring timestamps, and calculate the coefficient of variation to generate a valid historical power grid load coefficient of variation; wherein, step S32 includes: Step S321: Select the historical effective load window based on the real-time power grid monitoring timestamp to obtain the effective historical load window data; Step S322: Filter the standard historical multi-source power grid load data using the effective historical load window data to obtain the effective historical load data of the power grid; Step S323: Calculate the historical load average based on the effective historical load data of the power grid to generate effective historical load average data; Step S324: Calculate the historical load standard deviation based on the effective historical load data of the power grid to generate effective historical standard deviation data; Step S325: Calculate the coefficient of variation based on the effective historical load mean data and the effective historical standard deviation data to generate the effective historical load variation coefficient of the power grid; Step S33: Analyze the sensitivity parameters of the early warning source based on the effective historical load variation coefficient of the power grid, and generate the optimal alarm sensitivity parameters; wherein, step S33 includes: Step S331: Analyze the distribution characteristics of the coefficient of variation based on the effective historical load variation coefficient of the power grid, and set the initial sensitivity parameters to obtain the range data of the initial sensitivity parameters; Step S332: Obtain power grid resource load early warning source configuration data; Step S333: Connect the early warning source according to the power grid resource load early warning source configuration data, and read the historical early warning records to obtain the original power grid early warning record data; Step S334: Filter the actual alarm records based on the original power grid early warning record data to obtain the actual power grid alarm record data; Step S335: Filter false alarm records based on the original power grid early warning record data to obtain power grid false alarm record data; Step S336: Use the preset sensitivity traversal step size data to iterate through the initial sensitivity parameter range data to obtain the traversal sensitivity parameters; Step S337: Calculate the alarm recall rate by traversing the sensitivity parameters to obtain the power grid alarm recall rate data; Step S338: Calculate the alarm accuracy by iterating through the sensitivity parameters to obtain the power grid alarm accuracy data by analyzing the actual alarm record data and the false alarm record data. Step S339: Calculate the early warning performance index of the traversal sensitivity parameter based on the power grid alarm precision rate data and the power grid alarm recall rate data, and select the optimal sensitivity parameter to obtain the optimal alarm sensitivity parameter; Step S34: Calculate the load warning threshold based on the effective historical load variation coefficient of the power grid and the optimal alarm sensitivity parameter, and apply adaptive threshold constraints to obtain dynamic load warning threshold curve data; Step S4: Obtain real-time power grid resource and equipment status data; construct the power grid monitoring resource topology based on the real-time power grid resource and equipment status data to generate power grid monitoring resource topology data; construct a power grid load forecasting model based on the power grid monitoring resource topology data; use the power grid load forecasting model to perform short-term load forecasting to obtain short-term load forecasting data. Step S5: Calculate the threshold load difference of short-term load forecast data using dynamic load early warning threshold curve data, and perform multi-level early warning triggering processing to obtain the power grid multi-level early warning triggering strategy.

2. The method for dynamic threshold early warning of power grid resources based on multi-source data according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Obtain historical multi-source power grid resource load data; Step S12: Unify the format of historical multi-source power grid resource load data from multiple data sources, and perform data preprocessing to generate standard historical multi-source power grid load data; Step S13: Classify the historical load monitoring dates according to the standard historical multi-source power grid load data to obtain classified historical multi-source power grid load data; Step S14: Perform Fast Fourier Transform on the classified historical multi-source power grid load data to generate historical power grid load spectrum data; Step S15: Decompose the time-frequency components of the power grid load based on the historical power grid load spectrum data to generate the time-frequency component data of the power grid load.

3. The method for dynamic threshold early warning of power grid resources based on multi-source data according to claim 2, characterized in that, Step S15 includes the following steps: Step S151: Analyze the frequency range of major loads based on historical power grid load spectrum data to obtain historical load frequency range data; Step S152: Perform maximum wavelet decomposition layer processing based on historical load frequency range data to obtain the maximum wavelet decomposition layer; Step S153: Divide the frequency range of the decomposition level according to the maximum wavelet decomposition level, and perform decomposition level mapping on the historical load frequency range data to obtain load frequency-decomposition level mapping data. Step S154: Determine the coverage range of the decomposition layer for the load frequency-decomposition level mapping data, and optimize the wavelet decomposition level to generate the optimal decomposition layer data; Step S155: Perform wavelet transform decomposition on historical power grid load spectrum data based on the optimal decomposition level data to generate wavelet decomposition coefficients; Step S156: Extract wavelet approximate components based on wavelet decomposition coefficients to obtain long-term trend data of power grid load; Step S157: Extract wavelet detail components based on wavelet decomposition coefficients to obtain intraday power grid load fluctuation data, weekly power grid load fluctuation data, and seasonal power grid load fluctuation data, respectively. Step S158: Combine the long-term trend data of power grid load, the daily fluctuation data of power grid load, the weekly fluctuation data of power grid load, and the seasonal fluctuation data of power grid load with the time-frequency component of power grid resource monitoring to obtain the time-frequency component data of power grid load.

4. The method for dynamic threshold early warning of power grid resources based on multi-source data according to claim 3, characterized in that, Step S2 includes the following steps: Step S21: Set the rolling window for grid load based on the time-frequency component data of grid load, and obtain the annual trend baseline rolling window, intraday fluctuation baseline rolling window, weekly fluctuation baseline rolling window and seasonal fluctuation baseline rolling window respectively; Step S22: Calculate the long-term trend baseline of the power grid load using the annual trend baseline rolling window to generate long-term trend baseline data; Step S23: Calculate the intraday fluctuation baseline of the grid load intraday fluctuation data through the intraday fluctuation baseline rolling window to generate intraday fluctuation baseline data; Step S24: Calculate the weekly fluctuation baseline of the grid load within the week using the weekly fluctuation baseline rolling window to generate weekly fluctuation baseline data; Step S25: Calculate the seasonal fluctuation baseline for the power grid load seasonal fluctuation data using the seasonal fluctuation baseline rolling window to generate seasonal fluctuation baseline data; Step S26: Perform a dynamic baseline value weighted combination based on the long-term trend baseline data, intraday fluctuation baseline data, weekly fluctuation baseline data, and seasonal fluctuation baseline data to obtain the dynamic baseline value data of the power grid load.

5. The method for dynamic threshold early warning of power grid resources based on multi-source data according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Obtain real-time power grid resource equipment status data; Step S42: Construct the power grid monitoring resource topology based on real-time power grid resource equipment status data, and generate power grid monitoring resource topology data; Step S43: Collect real-time load data of power grid monitoring nodes based on power grid monitoring resource topology data to obtain real-time load data of power grid resources; Step S44: Perform Bayesian network node processing based on power grid monitoring resource topology data, and construct a power grid load prediction model through a preset convolutional neural network model to obtain the power grid load prediction model to be trained. Step S45: Process the standard historical multi-source power grid load data into a model training set, and train the power grid load prediction model to be trained to obtain the power grid load prediction model; Step S46: Use the power grid load forecasting model to perform short-term load forecasting on the real-time load data of power grid resources to obtain short-term load forecasting data.

6. The method for dynamic threshold early warning of power grid resources based on multi-source data according to claim 1, characterized in that, Step S5 includes the following steps: Step S51: Extract the forecast time points from the short-term load forecast data to obtain the power grid resource forecast time point data; Step S52: Identify the real-time load warning threshold range of the dynamic load warning threshold curve data using the power grid resource prediction time point data, and calculate the threshold load difference of the short-term load prediction data to obtain the predicted load difference data; Step S53: Calculate the load over-limit range based on the predicted load difference data to obtain the predicted load over-limit range data of the power grid; Step S54: Divide the early warning levels according to the power grid predicted load over-limit data, and perform multi-level early warning triggering processing through the preset early warning level execution strategy to obtain the power grid multi-level early warning triggering strategy.

7. A dynamic threshold early warning system for power grid resources based on multi-source data, characterized in that, For executing the power grid resource dynamic threshold early warning method based on multi-source data as described in claim 1, the power grid resource dynamic threshold early warning system based on multi-source data includes: The historical load time-frequency decomposition module is used to acquire standard historical multi-source power grid load data; classify the standard historical multi-source power grid load data by historical load monitoring date; and perform time-frequency component decomposition of power grid load to generate power grid load time-frequency component data. The load dynamic baseline construction module is used to set the grid load rolling window based on the grid load time-frequency component data and to perform dynamic baseline value weighting combination to obtain grid load dynamic baseline value data. The adaptive threshold generation module is used to filter effective historical load data from standard historical multi-source power grid load data and calculate the coefficient of variation to generate effective historical load variation coefficients; it analyzes the sensitivity parameters of the early warning source based on the effective historical load variation coefficients to generate optimal alarm sensitivity parameters; and it performs adaptive load early warning threshold processing on the dynamic baseline value data of power grid load based on the effective historical load variation coefficients and the optimal alarm sensitivity parameters to obtain dynamic load early warning threshold curve data. The short-term load forecasting module is used to acquire real-time power grid resource and equipment status data; construct the power grid monitoring resource topology based on the real-time power grid resource and equipment status data to generate power grid monitoring resource topology data; construct a power grid load forecasting model based on the power grid monitoring resource topology data; and use the power grid load forecasting model to perform short-term load forecasting to obtain short-term load forecasting data. The multi-level early warning triggering module is used to calculate the threshold load difference of short-term load forecast data through dynamic load early warning threshold curve data, and to perform multi-level early warning triggering processing to obtain the power grid multi-level early warning triggering strategy.

Images