Power grid disaster early warning method and device

Abstract

The invention discloses a power grid disaster early warning method and device, relates to the technical field of electric power safety, and mainly aims to improve the accuracy of an early warning result. According to the main technical scheme, the method comprises the steps of obtaining power grid disaster and meteorological data; for each data region, screening single feature sets meeting conditions according to a preset support degree obtained by a genetic algorithm; determining a plurality of occurrence time periods according to the occurrence time of the meteorological data; constructing a response window deviating backwards by taking each time period as a reference, wherein the window comprises a reference section and non-overlapping time periods deviating backwards for a preset time length; associating the reference section weather with the disaster data and associating the reference section weather with the disaster data in each response window to form a plurality of initial association sets; screening a target item combination feature set from the plurality of initial association sets; according to the preset confidence obtained by the genetic algorithm, screening a combined feature of a confidence load requirement as a target association rule; and inputting the target association rule, the power grid disaster and the meteorological data into a pre-training early warning model to obtain a power grid disaster risk early warning result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power safety, in particular to a power grid disaster early warning method and device. BACKGROUND

[0002] The goal of power grid disaster early warning is to identify the disaster risks that the power grid may face in advance, and timely issue early warning information, so that the relevant departments can take targeted prevention and response measures according to the early warning results, thereby effectively preventing or mitigating the occurrence of disaster accidents. At present, the power grid disaster early warning mainly adopts a data-driven method, which collects historical meteorological data and historical power grid disaster data, deeply mines the correlation between the two, and constructs a rule model that can be used to judge the risk. On this basis, when new real-time or forecast meteorological data is obtained, the system matches these rules to determine whether the early warning condition is met, and if so, the corresponding level of early warning is triggered.

[0003] However, the existing method has significant defects. When constructing the correlation rules between meteorology and power grid disasters, the existing method usually only focuses on the statistical correlation, and tends to extract the frequently co-occurring meteorological and fault data in the same period. This processing method ignores the time lag in the disaster evolution process. For example, after strong wind or freezing rain occurs in a certain area, power grid failure often does not occur immediately, but after several hours, such as the surrounding trees falling due to snow cover, causing line short circuit, and finally leading to power equipment failure. Therefore, the rules generated by the current method will be difficult to reflect the true causal relationship, resulting in inaccurate early warning: it may issue an alarm in a less affected area, but miss the report in a truly high-risk area, seriously affecting the accuracy of the early warning, and further causing the misallocation of operation and maintenance resources, threatening the safe operation of the power grid. SUMMARY

[0004] In view of the above problems, the present application provides a power grid disaster early warning method and device, the main purpose of which is to improve the accuracy of the early warning results and ensure the safe operation of the power grid.

[0005] To solve the above technical problems, the present application proposes the following solutions: In a first aspect, the present application provides a power grid disaster early warning method, which comprises: obtaining pre-processed power grid disaster data and meteorological data in a target time period, wherein each piece of power grid disaster data and meteorological data has a corresponding occurrence time and occurrence area; For each occurrence area, a single feature set that meets the preset support degree is selected from the power grid disaster data and meteorological data according to the preset support degree obtained by the genetic algorithm; For each occurrence area, a plurality of occurrence time periods are determined according to the occurrence time of each meteorological data in the single feature set; a plurality of response windows offset backward are constructed with each occurrence time period as a reference time period, the response window includes the reference time period and a plurality of time periods offset backward by a preset time length relative to the reference time period and not overlapping with each other; the meteorological data corresponding to the reference time period and the power grid disaster data are associated, and the meteorological data corresponding to the reference time period and the power grid disaster data occurring in each response window are associated to form a plurality of initial association sets of meteorological data and disasters; a target item combination feature set is screened out from the plurality of initial association sets in each occurrence area; a combination feature with a confidence higher than the preset confidence is screened out from the target item combination feature set as a target association rule according to the preset confidence obtained by the genetic algorithm; the target association rule, the power grid disaster data and the meteorological data are jointly input into a pre-trained early warning model to obtain a power grid disaster risk early warning result.

[0006] In a second aspect, the present application provides a power grid disaster early warning device, the device comprises: a data acquisition unit for acquiring preprocessed power grid disaster data and meteorological data in a target time period, wherein each piece of power grid disaster data and meteorological data has a corresponding occurrence time and occurrence area; a single set screening unit for screening a single feature set satisfying a preset support degree from the power grid disaster data and meteorological data acquired by the data acquisition unit according to the preset support degree obtained by the genetic algorithm for each occurrence area; a time period determination unit for determining a plurality of occurrence time periods according to the occurrence time of each meteorological data in the single feature set obtained by the single set screening unit for each occurrence area; a window determination unit for constructing a plurality of response windows offset backward with each occurrence time period determined by the time period determination unit as a reference time period, the response window includes the reference time period and a plurality of time periods offset backward by a preset time length relative to the reference time period and not overlapping with each other; an association set determination unit for associating the meteorological data corresponding to the reference time period and the power grid disaster data, and associating the meteorological data corresponding to the reference time period and the power grid disaster data occurring in each response window to form a plurality of initial association sets of meteorological data and disasters; a feature set screening unit for screening a target item combination feature set from the plurality of initial association sets in each occurrence area determined by the association set determination unit; a rule determination unit for screening a combination feature with a confidence higher than the preset confidence from the target item combination feature set obtained by the feature set screening unit as a target association rule according to the preset confidence obtained by the genetic algorithm; The risk early warning unit is configured to input the target association rule determined by the rule determination unit, the power grid disaster data and the meteorological data into a pre-trained early warning model to obtain a power grid disaster risk early warning result.

[0007] In order to achieve the above-mentioned purpose, according to the third aspect of the present application, a storage medium is provided, which comprises a stored program, wherein the storage medium controls the device where the storage medium is located to execute the power grid disaster early warning method of the first aspect when the program is running.

[0008] In order to achieve the above-mentioned purpose, according to the fourth aspect of the present application, a processor is provided, which is used to run a program, wherein the power grid disaster early warning method of the first aspect is executed when the program is running.

[0009] By the technical scheme, the power grid disaster early warning method and device can obtain preprocessed power grid disaster data and meteorological data in a target time period, each piece of data having a corresponding occurrence time and occurrence region. Then, for each occurrence region, a single feature set meeting a preset support degree obtained based on a genetic algorithm is screened from the power grid disaster data and the meteorological data. After obtaining the single feature set, for each occurrence region, a plurality of occurrence time periods are determined according to the occurrence time of the meteorological data, and a plurality of backward offset response windows are constructed by taking the occurrence time periods as reference time periods. The response window includes the reference time period and a plurality of non-overlapping time periods offset backward from the reference time period by a preset time length. By correlating the meteorological data and the power grid disaster data corresponding to the reference time period and the meteorological data and the power grid disaster data occurring in the response window, an initial correlation set of meteorological data and disasters is formed. The design can effectively identify the lagging influence of meteorological anomalies on power grid failures, clearly determine the time causal relationship of'meteorological change first and then failure', avoid false correlations caused by co-occurrence or reverse causality, significantly improve the explainability and early warning foresight of the correlation rules, provide a reliable data basis for subsequent mining of high-confidence correlation rules with time evolution rules, and thus improve the accuracy of the early warning results. Further, a target item combination feature set is screened from a plurality of initial correlation sets in each occurrence region, and a combination feature with a confidence higher than a preset confidence obtained based on a genetic algorithm optimization is screened as a final target correlation rule. The target correlation rule, the power grid disaster data and the meteorological data are jointly input into a pre-trained early warning model to generate a power grid disaster risk early warning result. In addition, compared with the prior art, the present application uses a step-by-step correlation strategy of 'first screening high-frequency single features and then constructing time sequence response windows', effectively eliminates occasional, isolated or low-impact noise data, ensures that subsequent correlation analysis is only carried out on representative high-frequency events, avoids interference of irrelevant data on the model. At the same time, the strategy significantly reduces the data size to be processed, reduces the calculation redundancy, enables the system to quickly focus on key areas and key events, improves the efficiency and stability of the correlation mining, and is more suitable for real-time disaster early warning applications in large-scale power grid scenarios. In summary, the present application not only improves the accuracy of the early warning results by mining high-confidence correlation rules with time evolution rules, but also improves the calculation efficiency of the early warning and ensures the safety of the power grid operation.

[0010] The above description is only a summary of the technical scheme of the present application. In order to more clearly understand the technical means of the present application, the present application can be implemented according to the content of the specification, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application are described below. BRIEF DESCRIPTION OF DRAWINGS

[0011] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a better understanding of the preferred embodiments, and are not intended to constrain the application. Moreover, like reference numerals denote same or similar components throughout the attached drawings. In the drawings: Figure 1 A flow chart of a power grid disaster early warning method according to an embodiment of the present application is shown in FIG. 1; Figure 2 A flow chart of another power grid disaster early warning method according to an embodiment of the present application is shown in FIG. 2; Figure 3 A block diagram of a power grid disaster early warning device according to an embodiment of the present application is shown in FIG. 3; Figure 4 A block diagram of another power grid disaster early warning device according to an embodiment of the present application is shown in FIG. 4. DETAILED DESCRIPTION

[0012] Exemplary embodiments of the present disclosure will be described more fully hereinafter with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be embodied in various forms without being limited by the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art.

[0013] In the existing power grid disaster early warning method, there are two prominent problems: on the one hand, when constructing the association rules of meteorology and power grid disasters, only the co-occurrence on the statistical data is usually concerned, and the frequent simultaneous occurrence of meteorological events and fault records in the same time period is inclined to be extracted. This modeling method based on synchronous correlation is difficult to distinguish the cause and effect relationship from the accidental association, and is easy to misjudge "co-occurrence" as "cause and effect", leading to the lack of physical interpretability of the generated rules, and further causing the inaccuracy of early warning, threatening the safe and stable operation of the power grid. On the other hand, in the association rule mining process, the support and confidence thresholds are mostly set by artificial, introducing strong subjective judgment, leading to the weak generalization ability of the rule mining results affected by experience, further intensifying the unreliability of the early warning results.

[0014] Therefore, there is an urgent need for a power grid disaster early warning method completely different from the existing method to improve the accuracy and practicality of early warning. The fundamental difference between the present application and the traditional association mining method is that the classic paradigm of "dividing fixed time windows first, then counting item set frequency" is abandoned, and a new idea of "directed association modeling based on causal time sequence guidance" is adopted. The traditional method usually divides the meteorological and power grid failure data into transactions according to fixed time intervals (such as every hour), considers that the events occurring in the same interval have relevance, and then mines rules by searching for frequent item sets layer by layer. Its essence is a static association analysis based on co-occurrence frequency, which is difficult to identify the temporal causal relationship of "meteorological changes first, then failure occurs". Moreover, the support and confidence thresholds are mostly set by humans, which is highly subjective and easy to introduce noise and false rules. The present application first filters out the single feature set that meets the preset support threshold obtained by genetic algorithm optimization from the power grid disaster data and meteorological data, preferentially retains high-frequency and high-representative meteorological and failure events, and effectively filters out incidental and low-impact data. Then, the occurrence time of each meteorological data in the single feature set is taken as the starting point to determine multiple occurrence time periods, and these time periods are used as reference time periods to construct multiple response windows that are offset backward and do not overlap with each other, each window containing the reference time period and subsequent time periods. Then, the meteorological data and power grid disaster data corresponding to the reference time period and the power grid disaster data occurring in the response window are associated to form an initial association set with clear time sequence logic, thereby capturing the lagging influence of meteorological anomalies on power grid failures and not ignoring the influence of instant meteorological data on power grid data. On this basis, the target item combination feature set is selected from the multiple initial association sets of each occurrence area, and the combination features with confidence higher than the threshold are selected as the target association rules by combining the preset confidence obtained by genetic algorithm optimization. This method realizes the transformation from "co-occurrence correlation" to "time sequence causality" through the step-by-step strategy of "first filtering high-frequency features, then constructing causal response windows", which not only avoids the misassociation caused by improper time alignment in traditional methods, but also reduces human intervention through data-driven parameter optimization, significantly improving the accuracy, interpretability and early warning foresight of the association rules, and fundamentally solving the false alarm and missed alarm problems caused by ignoring the lagging effect and relying on subjective setting in the existing technology, providing more reliable technical support for accurate early warning of power grid disasters.

[0015] Next, according to Figure 1 A power grid disaster early warning method provided by the present application is described, and the specific execution steps are as shown in Figure 1 The method comprises the following steps: 101. Obtain the preprocessed power grid disaster data and meteorological data in the target time period.

[0016] In this invention, power grid disaster data and meteorological data for the target time period are first acquired. The power grid disaster data includes records of events such as transmission line tripping, tower tilting, and insulator flashover. Each record includes the event type, occurrence time, and occurrence area (which can be represented as geographical coordinates or the power supply area). The meteorological data comes from weather stations, radar, and satellite remote sensing, and includes key parameters such as wind speed, rainfall, relative humidity, and lightning activity. Each data point is also labeled with its corresponding acquisition time and spatial location, i.e., the time and area of ​​occurrence.

[0017] To ensure data quality and consistency in analysis, a series of preprocessing operations are required after acquiring the raw data, including time alignment, outlier cleaning, unit standardization, and discretization. After preprocessing, all data is transformed into a unified format and includes consistent occurrence times and regions, providing a high-quality data foundation for subsequent correlation analysis.

[0018] The deadline for the target time period can be set to the current time, while the start time can be flexibly set according to the analysis needs without specific restrictions.

[0019] 102. For each affected area, based on the preset support obtained by the genetic algorithm, select a set of individual features that meet the preset support from the power grid disaster data and meteorological data.

[0020] In this step, frequency statistics can be performed on the preprocessed data in each affected area to identify frequently occurring meteorological and power grid disaster data.

[0021] Meanwhile, to avoid subjective bias caused by manually setting support and confidence levels, this invention employs a genetic algorithm to calculate support and confidence levels. Subsequently, a set of individual features that meet the preset support levels obtained from the genetic algorithm can be selected from power grid disaster data and meteorological data. This process effectively filters out low-frequency, sporadic events while retaining representative high-risk features.

[0022] Example: In City A, statistics show that "rainfall" occurred 3 times and "line tripping" occurred 5 times, both of which are higher than the preset support obtained by the genetic algorithm optimization. Therefore, "rainfall" and "line tripping" are included in the single feature set; while "instantaneous lightning strike" only occurred once, which is below the threshold and is excluded.

[0023] Support measures the prevalence of an event in a dataset, and its formula is: Support = Number of occurrences of the event / Total number of samples. Setting support aims to filter out "noise" events that occur too infrequently or may be due to random factors, ensuring that the features included in the analysis are sufficiently representative and avoiding the generation of rules that are meaningless.

[0024] Confidence score is used to assess the conditional reliability of a rule. It measures the probability that event B (such as a line tripping) will occur given that event A (such as "rainfall") has occurred. Confidence score = (Number of times A and B occur simultaneously) / (Number of times A occurs alone). This indicator directly reflects the warning effectiveness of the rule ("rainfall → line tripping"), determining whether it has practical guiding significance; that is, what is the probability that event B will occur when event A is observed.

[0025] 103. For each occurrence area, multiple occurrence time periods are determined based on the occurrence time corresponding to each meteorological data in the single feature set.

[0026] 104. Construct multiple backward-shifted response windows using each occurrence time period as the base time period. Each response window includes the base time period and multiple non-overlapping time periods that are shifted backward by a preset time period relative to the base time period.

[0027] 105. Associate the meteorological data and power grid disaster data corresponding to the baseline time period, and associate the meteorological data and power grid disaster data occurring within each response window to form multiple initial association sets of meteorology and disasters.

[0028] Subsequently, in step 103, based on the single feature set constructed in step 102, the occurrence time points corresponding to all meteorological data are extracted, and these time points are summarized and integrated to divide into multiple consecutive occurrence time periods.

[0029] The division process can adopt different strategies according to actual needs: it can be a flexible and free division, or it can be combined with the characteristics of power grid operation, such as dividing according to the peak, flat and low periods of electricity load, so as to enhance the correlation between time periods and the actual operating status of the power grid.

[0030] After being divided, each time period may contain multiple meteorological data points. These data may involve multiple different meteorological parameters (such as wind speed, rainfall, temperature, etc.), or they may correspond to the observation values ​​of the same meteorological parameter at different times. For example, a certain time period may simultaneously contain temperature records at 10:00 and 11:00.

[0031] Example illustration: Suppose a single feature set contains 5 meteorological data points that meet the support threshold, occurring at 10:00, 11:30, 12:00, 13:00, and 13:30 on July 15, 2025. Based on temporal continuity and interval distribution, these time points can be categorized into two time periods: 10:00–11:30 and 12:00–13:30. This division helps organize discrete meteorological events into temporally continuous analytical units, providing structured input for subsequent correlation analysis.

[0032] After determining the plurality of time periods, a plurality of backward-shifted response windows can be constructed in step 104, with each of the time periods as a reference time period. The response window includes the reference time period and a plurality of time periods that are sequentially shifted backward with respect to the reference time period by a preset time length and do not overlap with each other.

[0033] The preset time length can be determined according to the hysteresis characteristics of the response of the power grid equipment under actual meteorological conditions, for example, the longest chain reaction time required for an insulator flashover, line dancing or tower tilting and other faults caused by extreme weather such as strong wind, heavy rain or lightning, so as to ensure covering the potential delay effect of meteorological influence.

[0034] Subsequently, in step 105, for each response window, the meteorological data corresponding to the reference time period and the power grid disaster data are associated, and the meteorological data corresponding to the reference time period and the power grid disaster data occurring in each response window are associated to form a plurality of initial association sets of meteorological and disaster data. Specifically: For the response window constituted by the reference time period itself, the meteorological data inside it is associated with the power grid disaster occurring at the same time; for the backward-shifted response window, the meteorological data of the reference time period is associated with the power grid disaster occurring in the shifted time period, so as to capture the faults possibly caused in the subsequent period after the occurrence of the meteorological event.

[0035] The reason for associating the meteorological data of the reference time period with the power grid disaster event occurring at the same time in the response window is that some meteorological conditions (such as strong wind, lightning, short-time heavy rain, etc.) can cause power grid faults in a short time after the occurrence, and there is a significant immediate correlation between them. Such faults respond quickly, and the cause-and-effect relationship is clear, which is a key mode that cannot be ignored in power grid disaster warning.

[0036] For example, in step 104, the reference time period is “2025-07-15 10:00-11:00”, and the following three response windows are constructed: Reference time period: 10:00-11:00 First response window: 11:00-12:00 Second response window: 12:00-13:00 In step 105, it is assumed that the following meteorological data is recorded in the reference time period (10:00-11:00): rainfall A value, wind speed B value; at the same time, the power grid disaster data occurring in the time period is: the line tripping of a certain line in region A. The initial association set formed is: {rainfall A value, wind speed B value, line tripping of a certain line in region A} Furthermore, if another line in Area A trips within the first response window (11:00–12:00), then based on the same baseline meteorological conditions (i.e., rainfall value A and wind speed value B from 10:00–11:00) and the disaster event during that period, a corresponding initial association set is generated as: {rainfall value A, wind speed value B, another line in Area A trips}. It should be noted that the above process uses the occurrence area as the basic analysis unit. Within each area, corresponding response windows are constructed for multiple occurrence time periods, and multiple initial association sets are generated under each window. This approach allows for the systematic exploration of the potential impact of the same meteorological event on power grid equipment within different time delay intervals. Furthermore, independent modeling at the regional level reflects the differences in meteorological-disaster response characteristics under different geographical environments, improving the model's refinement and practicality.

[0037] It should be noted that in the above method, response windows constructed from different baseline time periods may cover power grid disaster events within the same time period, resulting in the disaster event appearing in multiple initial association sets. This method can preserve the potential correlation between power grid disaster events and different historical meteorological periods, which helps to identify complex meteorological disaster-causing patterns such as "continuous rainfall + instantaneous strong winds" and improves the ability to reconstruct the causes of complex faults. In addition, the occurrence of some power grid faults has a lag and uncertainty, making it difficult to accurately determine their unique meteorological triggering period. By preserving the association records of the event in multiple time windows, it can be ensured that meteorological impacts of different time periods and intensities are fully recorded and assessed.

[0038] 106. Select the target item combination feature set based on multiple initial association sets in each occurrence region.

[0039] 107. Based on the preset confidence level obtained by the genetic algorithm, select combination features with confidence levels higher than the preset confidence level from the target item combination feature set as target association rules.

[0040] 108. Input the target association rules, power grid disaster data and meteorological data into the pre-trained early warning model to obtain the power grid disaster risk early warning result.

[0041] After generating multiple initial association sets in step 105, each initial association set can be regarded as an independent data "transaction", and these transactions constitute the basic dataset for subsequent association rule mining.

[0042] On this basis, step 106 can be entered, and a layer-by-layer progressive frequent item set mining strategy can be used to screen the target item combination feature set meeting the support degree from all transactions. The mining process starts from a binomial combination, that is, a combination of two feature items appearing at the same time and meeting the support degree (for example, a certain meteorological data and a certain power grid disaster data), and then generates and evaluates a trinomial, a quadrinomial and the like based on the high-frequency binomial combination, and iterates continuously until no new combination meeting the support degree requirement can be generated.

[0043] To improve the calculation efficiency, the application can also introduce a phased filtering mechanism in step 106: after screening the binomial combination feature set, all feature items not included in any high-frequency binomial combination are marked and migrated to an isolated storage area. These isolated feature items do not meet the support degree requirement by themselves or their combinations, indicating that their frequency is low and their relevance is weak, and they are most likely to be low-frequency items in subsequent high-order combination mining. Excluding these feature items in advance can significantly reduce the subsequent calculation amount, avoid invalid traversal, and improve the overall mining efficiency.

[0044] It should be noted that the feature items that do not meet the high-frequency binomial combination condition are marked and migrated to the isolated storage area, which aims to preserve the original information of these low-frequency data and avoid permanent deletion. Although such data does not enter the mining path of the target association rule, it can still be input into the early warning model as auxiliary information in subsequent step 108 for context comparison, anomaly detection or model confidence correction, thereby improving the model's ability to identify rare events or new risk patterns.

[0045] Subsequently, in step 107, based on the preset confidence optimized by the genetic algorithm, all candidate rules in the target item combination feature set are screened, and only combinations with a confidence higher than the preset confidence are retained, finally forming a target association rule with high reliability.

[0046] Finally, in step 108, the obtained target association rule is input into the pre-trained power grid disaster risk early warning model together with the power grid disaster data and meteorological data in the target time period. Since the data in the target time period covers the current time, the model can comprehensively judge the fault risk level of each region under the current weather condition, output the corresponding risk warning result, and realize the closed-loop support from data to decision.

[0047] Based on the above Figure 1As can be seen from the implementation method, the power grid disaster early warning method provided by this invention first acquires preprocessed power grid disaster data and meteorological data within a target time period, where each data point has a corresponding occurrence time and location. Then, for each occurrence location, based on a preset support level obtained through genetic algorithm optimization, a set of individual features satisfying the support threshold is selected from the power grid disaster data and meteorological data. After obtaining the individual feature sets, for each occurrence location, multiple occurrence time periods are determined based on the occurrence time of the meteorological data, and these occurrence time periods are used as reference time periods to construct multiple backward-shifted response windows; each response window includes the reference time period and several non-overlapping time periods shifted backward by a preset time relative to the reference time period. By associating the meteorological data corresponding to the reference time period with the power grid disaster data, and by associating the meteorological data corresponding to the reference time period with the power grid disaster data occurring within each response window, multiple initial association sets of meteorology and disasters are formed. This design effectively identifies the lagged impact of meteorological anomalies on power grid faults, clarifies the temporal causal relationship of "meteorological change precedes fault occurrence," avoids false associations caused by co-occurrence or reverse causality, significantly improves the interpretability and early warning foresight of association rules, and provides a reliable data foundation for subsequent mining of high-confidence association rules with temporal evolution patterns, thereby improving the accuracy of early warning results. Furthermore, a target item combination feature set is selected based on multiple initial association sets in each occurrence area; and based on a pre-set confidence level obtained through genetic algorithm optimization, combination features with confidence levels higher than a certain threshold are selected as the final target association rules. The target association rules, power grid disaster data, and meteorological data are jointly input into a pre-trained early warning model to generate power grid disaster risk early warning results. In addition, compared with existing technologies, this invention, through a step-by-step association strategy of "first screening high-frequency single features, then constructing a time-series response window," effectively eliminates occasional, isolated, or low-impact noise data, ensuring that subsequent association analysis is only conducted on representative high-frequency events, avoiding interference from irrelevant data on the model. Meanwhile, this strategy significantly reduces the scale of data to be processed, lowers computational redundancy, and enables the system to quickly focus on key areas and events, improving the efficiency and stability of correlation mining. It is more suitable for real-time disaster early warning applications in large-scale power grid scenarios. In summary, this invention not only improves the accuracy of early warning results and ensures the safety of power grid operation by mining high-confidence correlation rules with temporal evolution patterns, but also significantly improves the computational efficiency of early warning.

[0048] Furthermore, as a response to Figure 1 Further refinement and extension of the illustrated embodiment, this invention also provides another power grid disaster early warning method, such as... Figure 2 As shown, the specific steps are as follows: 201. Obtain preprocessed power grid disaster data and meteorological data within the target time period.

[0049] The implementation of step 201 is the same as that of step 101, and the same technical effects can be achieved, and the same technical problems can be solved, which will not be repeated here.

[0050] 202. Calculate the preset support and preset confidence according to the genetic algorithm.

[0051] In the present application, the user can pre-set the initial support, the initial confidence and the average value between the two according to historical experience or business requirements. At the same time, the weight coefficients of the support, the confidence and the average value in the fitness evaluation are set. Based on these initial parameters, the support and the confidence are adaptively optimized by using the genetic algorithm. The specific process is as follows: First, generate multiple candidate support and candidate confidence in a reasonable preset value range based on the initial support and the initial confidence to form an initial population. Each individual in the population represents a set of support and confidence to be evaluated.

[0052] Subsequently, the individuals in the initial population are subjected to cross and mutation operations to generate a new generation of population. For each individual in the population, a preset fitness function is used to calculate its fitness value.

[0053] The fitness function is as follows: (Formula 1) Where F is the fitness function, , , is the weight, sup is the support, con is the confidence, and avg is the average value of the support and the confidence.

[0054] The higher the fitness value, the more balanced the support and the confidence, and the more conducive to mining high-quality and practical association rules.

[0055] In each generation of evolution, the individual with the highest fitness value is selected to enter the next generation, and the cross, mutation and fitness evaluation are repeated until the preset termination condition (such as reaching the maximum number of iterations or the fitness converges) is met. Finally, the support and the confidence corresponding to the individual with the highest fitness value are output as the preset support and the preset confidence, which are used for subsequent feature screening and rule extraction.

[0056] 203. For each occurrence area, the preset support obtained by the genetic algorithm is used to screen a single feature set from the power grid disaster data and the meteorological data that meets the preset support.

[0057] 204. For each occurrence area, a plurality of occurrence time periods are determined according to the occurrence time corresponding to each meteorological data in the single feature set.

[0058] 205. Construct a plurality of backward offset response windows with each occurrence time period as a reference time period, the response window comprising the reference time period and a plurality of time periods that are not overlapped with each other and offset backward from the reference time period by a preset time length.

[0059] 206. Correlate the weather data corresponding to the reference time period with the power grid disaster data, and correlate the weather data corresponding to the reference time period with the power grid disaster data occurring in each response window to form a plurality of initial correlation sets of weather and disasters.

[0060] The embodiments of steps 203-206 are the same as those of steps 102-105, and can achieve the same technical effects, solve the same technical problems, and will not be repeated here.

[0061] 207. Screen a target item combination feature set from the plurality of initial correlation sets in each occurrence area.

[0062] In this embodiment, the preset item set number is first determined. Generally, association rule mining starts from a combination containing one feature (i.e., a single item set) and iterates. However, in the present application, since the screening of single item feature sets has been completed in the previous step (i.e., high-frequency weather and disaster features have been filtered out by the support threshold), this step can directly use two items as the initial preset item set number, skipping the single item mining process.

[0063] This design is significantly different from the layer-by-layer construction method of traditional mining algorithms, effectively avoiding repeated calculation of low-frequency single items and improving overall mining efficiency.

[0064] Subsequently, in each power supply area, for the plurality of initial correlation sets corresponding to the area (each is regarded as a transaction), all combinations with the number of feature items equal to the current preset item set number are extracted to form the initial feature set between the weather data and the power grid disaster data. If the initial feature set is not empty, the frequency of each combination appearing in all initial correlation sets is further counted, and the combinations with an appearance frequency not lower than the preset support threshold are determined as candidate initial feature sets.

[0065] Then, the specified item number (usually 1) is added to the preset item set number to obtain the updated item set number (i.e., entering the three-item combination stage), and based on the co-occurrence relationship between the feature items in the candidate initial feature set, all updated feature sets with a length of the updated item set number are generated by combination.

[0066] If the updated feature set is not empty, the appearance frequency of each combination in the initial correlation set is also counted, and the combinations that meet the preset support are screened to form the candidate updated feature set.

[0067] Thereafter, the iterative operation of "incrementing the item set length -> generating a new feature set -> support degree screening" is repeatedly performed to mine more complex multi-factor coupling patterns layer by layer until the iteration is terminated when no new feature combination satisfying the preset support degree can be generated. Finally, the candidate update feature set successfully generated in the last time is taken as the target item combination feature set for subsequent confidence evaluation and rule extraction.

[0068] For example, it is assumed that the initial association set in a certain area is as follows: {high wind speed, heavy rainfall, A line trip} {high wind speed, heavy rainfall, B line trip} {high wind speed, low temperature, C line icing} This step directly starts mining from the binomial combination: Generate all binomial combinations: {high wind speed, heavy rainfall} (2 times) {high wind speed, A line trip} (1 time) {heavy rainfall, A line trip} (1 time) ... If the preset support degree is set to 2, only {high wind speed, heavy rainfall} satisfies the condition, and enters the candidate initial feature set.

[0069] Increment to the trinomial combination: based on {high wind speed, heavy rainfall} and other high-frequency item combinations, for example, generate {high wind speed, heavy rainfall, A line trip}, {high wind speed, heavy rainfall, B line trip}.

[0070] If statistics show that the two trinomial combinations each appear once, but the preset support degree is 2, there is no trinomial combination satisfying the condition, and the iteration is terminated, but if the number of occurrences of the two trinomial combinations is greater than the preset support degree, the iteration can continue.

[0071] To further improve the calculation efficiency, after completing the screening of each layer of feature set, the system will mark the feature items that are not included in any high-frequency combination and migrate them to the isolated storage area. For example, "low temperature" and "icing" appear in the original transaction, but do not enter any high-frequency binomial or trinomial combination, indicating that their association strength is weak or is only an occasional event.

[0072] The purpose of this design is not to delete data, but to exclude it from the main path mining, reducing the search space of subsequent high-order combinations, thereby significantly reducing the calculation complexity. At the same time, these isolated data are still retained in the system and can be used as auxiliary input for anomaly detection or model boundary analysis in subsequent step 209, improving the robustness of the early warning system.

[0073] In addition, in the present scheme, mining and analysis need to be performed for each occurrence area in the target item combination feature set generation stage. This is because different areas have significant differences in geographical environment, climate characteristics, and power grid structure, etc. If the data of all areas are mixed and processed, the area-specific meteorological-disaster correlation mode may interfere with each other or even cancel out, making it difficult to mine effective and practically meaningful correlation rules. Through zoned modeling, the local rules of each area can be fully captured, improving the support and confidence of the rules.

[0074] However, when the target correlation rules are subsequently input into the early warning model, there is no need to explicitly distinguish between areas in the model structure. This is because the regional characteristics have been fully integrated in the rule screening stage, and the resulting rules themselves contain regional prior knowledge. At this time, the high-confidence rules of each area can be uniformly input into the global early warning model as features. The "explainability" value of the rules has been embodied in the early stage, while the model focuses on dynamic assessment and prediction of risk probability by integrating multiple rules, realizing the organic combination of knowledge-driven and data-driven.

[0075] 208. From the target item combination feature set, filter combination features with a confidence higher than the preset confidence obtained from the genetic algorithm, as target correlation rules.

[0076] The implementation of step 208 is the same as that of step 107, and can achieve the same technical effects, solve the same technical problems, and will not be repeated here.

[0077] 209. Jointly input the target correlation rules, power grid disaster data, and meteorological data into the pre-trained early warning model to obtain power grid disaster risk early warning results.

[0078] In this step, first, according to the occurrence time and occurrence area of the power grid disaster data and meteorological data, fine temporal and spatial matching is performed to construct structured spatio-temporal sequence data. The specific processing process is as follows: Each data is mapped to a spatio-temporal unit composed of a preset spatial grid and a time interval according to its occurrence time and geographical location; for multiple data of the same parameter belonging to the same spatio-temporal unit, the average value is used for aggregation to generate a unified data value under the spatio-temporal unit; for the target spatio-temporal unit without direct observation data, the inverse distance weighted interpolation method is used based on the observation values of the adjacent spatial grid to generate the data value of the target spatio-temporal unit, to ensure the integrity and continuity of the spatio-temporal sequence.

[0079] After the above processing, all power grid disaster data and meteorological data are integrated into spatio-temporal sequence data of a unified format as input for the subsequent deep learning model.

[0080] Subsequently, a convolutional neural network (CNN) is used to extract features from the spatio-temporal sequence data. CNN effectively captures the local correlation and pattern features of weather and power grid state in space through local receptive fields and weight sharing mechanisms, and outputs corresponding high-order feature sequences.

[0081] Further, the high-order feature sequence and the target association rule generated in the previous step are jointly input into a convolutional attention module (CBAM) embedded in the early warning model. The module is located between the CNN convolutional neural network and the long short-term memory network LSTM, and its function is to give higher attention weights to the corresponding feature channels and spatial regions based on the meteorological data and power grid disaster data indicated by the target association rule, that is, to guide the attention mechanism to focus on key feature channels and sensitive spatial regions.

[0082] Specifically, the CBAM module includes two sub-modules: Channel attention module: according to the meteorological parameters and disaster types involved in the target association rule, the weight of the corresponding feature channel is enhanced; Spatial attention module: higher spatial attention weights are given to the spatial grid regions associated with the above-mentioned key parameters.

[0083] Through this rule-guided attention mechanism, the model can dynamically enhance the feature expression related to known high-risk patterns, output the attention-enhanced feature sequence, and improve the feature discrimination ability.

[0084] Subsequently, the enhanced feature sequence is input into a long short-term memory network (LSTM). LSTM models the state evolution in the time dimension through its internal gating mechanism, captures the temporal dependence and dynamic change trend between weather conditions and power grid responses, and finally generates power grid disaster risk early warning results, including risk level, possible time interval and impact area.

[0085] In addition, after generating the early warning results, when the system receives new power grid disaster data or meteorological data, an online learning mechanism in the early warning model can be triggered to dynamically update the gating parameters of the LSTM network. This mechanism enables the model to continuously adapt to changes in the power grid operating environment (such as equipment aging, climate trend deviation, etc.), improving its long-term prediction performance and self-adaptive ability.

[0086] Further, as an implementation of the method shown in the above Figure 1 As an implementation of the method shown in the above Figure 1The method shown is implemented. The device embodiment corresponds to the foregoing method embodiment, and for the sake of readability, the details of the foregoing method embodiment will not be described one by one, but it should be clear that the device in the embodiment can correspondingly implement all the contents in the foregoing method embodiment. As shown in the method Figure 3 The device comprises: The data acquisition unit 301 is configured to acquire preprocessed power grid disaster data and meteorological data in a target time period, wherein each piece of power grid disaster data and meteorological data has a corresponding occurrence time and occurrence region. The single set screening unit 302 is configured to, for each occurrence region, screen a single feature set satisfying a preset support degree from the power grid disaster data and meteorological data acquired by the data acquisition unit 301 according to the preset support degree obtained by the genetic algorithm. The time period determination unit 303 is configured to, for each occurrence region, determine a plurality of occurrence time periods according to the occurrence time corresponding to each piece of meteorological data in the single feature set obtained by the single set screening unit 302. The window determination unit 304 is configured to construct a plurality of response windows offset backward with each occurrence time period determined by the time period determination unit 303 as a reference time period, wherein the response window comprises the reference time period and a plurality of time periods offset backward by a preset time length with respect to the reference time period and do not overlap with each other. The association set determination unit 305 is configured to associate the meteorological data and the power grid disaster data corresponding to the reference time period determined by the window determination unit 304, and associate the meteorological data corresponding to the reference time period and the power grid disaster data occurring in each response window, to form a plurality of initial association sets of meteorological data and disasters. The feature set screening unit 306 is configured to screen a target item combination feature set from the plurality of initial association sets in each occurrence region determined by the association set determination unit 305. The rule determination unit 307 is configured to screen a combination feature with a confidence higher than a preset confidence from the target item combination feature set obtained by the feature set screening unit 306 as a target association rule according to the preset confidence obtained by the genetic algorithm. The risk early warning unit 308 is configured to input the target association rule determined by the rule determination unit 307, the power grid disaster data and the meteorological data into a pre-trained early warning model to obtain a power grid disaster risk early warning result.

[0087] Further, as an implementation of the foregoing Figure 2 The method shown is implemented. The device embodiment corresponds to the foregoing method embodiment, and for the sake of readability, the details of the foregoing method embodiment will not be described one by one, but it should be clear that the device in the embodiment can correspondingly implement all the contents in the foregoing method embodiment. As shown in the method Figure 2The method is implemented. The device embodiment corresponds to the foregoing method embodiment, and details in the foregoing method embodiment will not be described one by one for the convenience of reading, but it should be clear that the device in the embodiment can correspondingly implement all contents in the foregoing method embodiment. As shown in the embodiment, Figure 4 The device comprises: The data acquisition unit 301 is configured to acquire preprocessed power grid disaster data and meteorological data in a target time period, wherein each piece of power grid disaster data and meteorological data has a corresponding occurrence time and occurrence region; The single set screening unit 302 is configured to, for each occurrence region, screen a single feature set satisfying a preset support degree from the power grid disaster data and meteorological data acquired by the data acquisition unit 301 according to the preset support degree obtained by the genetic algorithm; The time period determination unit 303 is configured to, for each occurrence region, determine a plurality of occurrence time periods according to the occurrence time corresponding to each piece of meteorological data in the single feature set obtained by the single set screening unit 302; The window determination unit 304 is configured to construct a plurality of response windows offset backward with each occurrence time period determined by the time period determination unit 303 as a reference time period, wherein the response window comprises the reference time period and a plurality of time periods offset backward by a preset time length from the reference time period and do not overlap with each other; The association set determination unit 305 is configured to associate the meteorological data and the power grid disaster data corresponding to the reference time period determined by the window determination unit 304, and associate the meteorological data corresponding to the reference time period and the power grid disaster data occurring in each response window, to form a plurality of initial association sets of meteorological data and disasters; The feature set screening unit 306 is configured to screen a target item combination feature set from the plurality of initial association sets in each occurrence region determined by the association set determination unit 305; The rule determination unit 307 is configured to screen a combination feature with a confidence higher than a preset confidence from the target item combination feature set obtained by the feature set screening unit 306 as a target association rule according to the preset confidence obtained by the genetic algorithm; The risk early warning unit 308 is configured to input the target association rule determined by the rule determination unit 307, the power grid disaster data and the meteorological data into a pre-trained early warning model to obtain a power grid disaster risk early warning result.

[0088] In an optional embodiment, before the single set screening unit 302 screens a single feature set satisfying a preset support degree from the power grid disaster data and meteorological data according to the preset support degree obtained by the genetic algorithm for each occurrence region, the device further comprises a calculation unit 309, and the calculation unit 309 is specifically configured to: obtaining an initial support degree, an initial confidence degree, an average value between the initial support degree and the initial confidence degree, and weights corresponding to the initial support degree, the initial confidence degree, and the average value respectively; According to the initial support degree, the initial confidence degree, the average value, and the weights corresponding to the initial support degree, the initial confidence degree, and the average value respectively, a preset fitness function is used to perform iterative optimization through a genetic algorithm, and preset support degree and preset confidence degree are output.

[0089] In an optional implementation, when the computing unit 309 outputs preset support degree and preset confidence degree according to the initial support degree, the initial confidence degree, the average value, and the weights corresponding to the initial support degree, the initial confidence degree, and the average value respectively by using a preset fitness function through a genetic algorithm, the computing unit 309 is specifically configured to: generate a plurality of candidate support degrees and a plurality of candidate confidence degrees in respective preset value ranges based on the initial support degree and the initial confidence degree; construct an initial population according to the initial support degree, the initial confidence degree, the plurality of candidate support degrees, and the plurality of candidate confidence degrees, and each individual in the initial population represents a group of support degree and confidence degree; generate new individuals by performing crossover and mutation on the individuals in the initial population, and calculate fitness values of the individuals by using the preset fitness function in combination with the support degree, the confidence degree, the average value between the support degree and the confidence degree, the support degree and the confidence degree, and the weights corresponding to the support degree, the confidence degree, and the average value in each individual; select individuals with the highest fitness values to enter the next generation, and repeat the process of generating new individuals and calculating the fitness values of the individuals until a preset termination condition is met, and output the support degree and the confidence degree corresponding to the individual with the highest fitness value as the preset support degree and the preset confidence degree.

[0090] In an optional implementation, the feature set screening unit 306 is specifically configured to: obtain a preset item set number; extract, in each occurrence area, all combinations of item sets with a length equal to the preset item set number from the plurality of initial associations as initial feature sets between the meteorological data and the power grid disaster data; if the initial feature sets are not empty, count frequencies of the initial feature sets in the plurality of initial associations, and determine initial feature sets with frequencies satisfying a preset support degree as candidate initial feature sets; increase a specified item number on the basis of the preset item set number to obtain an updated item set number, and generate, based on the meteorological data and the power grid disaster data in the candidate initial feature sets, updated feature sets with an item set length equal to the updated item set number by combination; If the update feature set is not empty, the occurrence frequency of each update feature set in a plurality of initial association sets is counted, and an update feature set with an occurrence frequency satisfying the preset support degree is determined as a candidate update feature set; The operations of incrementing the item set length, generating the feature set, and filtering the support degree are repeatedly performed until a new feature set satisfying the preset support degree cannot be generated, and the candidate update feature set generated last time is taken as the target item combination feature set.

[0091] In an optional implementation, the risk early warning unit 308 is specifically configured to: spatiotemporal matching is performed on the power grid disaster data and the meteorological data according to occurrence time and occurrence region, to obtain spatiotemporal sequence data; The spatiotemporal sequence data is processed by a convolutional neural network to obtain high-order feature sequences; The high-order feature sequences and the target association rule are jointly input into a convolutional attention module embedded between a convolutional neural network and a long short-term memory network in the early warning model, to output feature sequences enhanced by attention, wherein the convolutional attention module is configured to give higher attention weights to corresponding feature channels and spatial regions according to meteorological data and power grid disaster data indicated by the target association rule; The feature sequences enhanced by attention are input into the long short-term memory network, so that the long short-term memory network models time evolution rules through a gating mechanism, to generate a power grid disaster risk early warning result.

[0092] In an optional implementation, when the risk early warning unit 308 performs spatiotemporal matching on the power grid disaster data and the meteorological data according to occurrence time and occurrence region to obtain spatiotemporal sequence data, the risk early warning unit 308 is specifically configured to: According to the occurrence time and the occurrence region of each piece of power grid disaster data and meteorological data, the power grid disaster data and the meteorological data are attributed to a spatiotemporal unit formed by a corresponding spatial grid and a time interval; For a plurality of data of the same parameter attributed to the same spatiotemporal unit, an average value is used for aggregation to generate a unified data value under the spatiotemporal unit; For a target spatiotemporal unit without direct observation data, an inverse distance weighted interpolation method is used to generate a data value of the target spatiotemporal unit based on observation values of adjacent spatial grids; The processed power grid disaster data and meteorological data are determined as spatiotemporal sequence data.

[0093] In an optional implementation, after the risk early warning unit 308 generates a power grid disaster risk early warning result, when new power grid disaster data and meteorological data are received, a gating parameter of the long short-term memory network is updated by using an online learning mechanism in the early warning model.

[0094] Further, the embodiment of the present application also provides a storage medium for storing a computer program, wherein the computer program controls a device where the storage medium is arranged to execute the power grid disaster early warning method described in the above Figures 1-2

[0095] Further, the embodiment of the present application also provides a processor for running a program, wherein the program executes the power grid disaster early warning method described in the above Figures 1-2

[0096] In the above embodiment, the description of each embodiment has its own focus, and the part not described in detail in an embodiment can be referred to the related description of other embodiments.

[0097] It can be understood that the related features in the above method and device can be mutually referred. In addition, the "first", "second" and the like in the above embodiment are used to distinguish the embodiments, and do not represent the advantages and disadvantages of the embodiments.

[0098] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.

[0099] The algorithms and displays provided herein are not inherently related to any particular computer, virtual system, or other apparatus. Various general purpose systems can be used with these teachings, with the structure for a variety of such systems will be apparent from the description above. In addition, the present application is not intended to be limited to any particular programming language. It will be appreciated that a variety of programming languages can be used to implement the present application as described herein, and any references below to specific languages are provided for disclosure of enablement only.

[0100] In addition, the memory can include non-persistent memory in computer readable medium, random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM), and the memory includes at least one memory chip.

[0101] Those skilled in the art will appreciate that embodiments of the present application can be a method, a system or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code thereon.

[0102] ​​The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps in one or more flow or blocks

[0103] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps in one or more flow or blocks

[0104] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps in one or more flow or blocks

[0105] In one typical configuration, the computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0106] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) and / or cache memory, for storing instructions and data used and / or generated by the computing device. The memory can also include non-volatile memory, such as read-only memory (ROM), electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory, or nonvolatile random access memory (RAM) (e.g., fault tolerant RAM), for storing instructions and data used and / or generated by the computing device. The memory is an example of computer readable media.

[0107] Computer-readable media includes permanent and non-permanent, movable and non-movable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer-readable media does not include transitory media such as modulated data signals and carriers.

[0108] It should also be noted that the terms "comprising", "containing", or any other variant thereof are intended to cover non-exclusive inclusions, such that a process, method, article or apparatus that comprises a list of elements does not only include those elements, but can also include other elements not expressly listed or inherent to such process, method, article or apparatus. Without more limitations, the element defined by the phrase "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus that includes the element.

[0109] Those skilled in the art will appreciate that embodiments of the present application can be provided as a method, system or computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer usable program code.

[0110] The above merely provides embodiments of the present application and is not intended to limit the present application. Various modifications and changes can be made to the present application by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the scope of the claims of the present application.

Claims

1. A power grid disaster early warning method, characterized by, The method comprises: obtaining pre-processed power grid disaster data and meteorological data in a target time period, wherein each piece of power grid disaster data and meteorological data has a corresponding occurrence time and occurrence region; for each occurrence region, screening a single feature set that meets a preset support degree from the power grid disaster data and meteorological data according to the preset support degree obtained by a genetic algorithm; for each occurrence region, determining a plurality of occurrence time periods according to the occurrence time corresponding to each piece of meteorological data in the single feature set; constructing a plurality of backward offset response windows with each occurrence time period as a reference time period, wherein the response window comprises the reference time period and a plurality of time periods that are offset backward from the reference time period by a preset time length and do not overlap with each other; associating the meteorological data and power grid disaster data corresponding to the reference time period, and associating the meteorological data corresponding to the reference time period and the power grid disaster data occurring in each response window, to form a plurality of initial association sets of meteorological data and disasters; screening a target combined feature set from the plurality of initial association sets in each occurrence region; screening a combined feature with a confidence degree higher than a preset confidence degree from the target combined feature set as a target association rule according to the preset confidence degree obtained by the genetic algorithm; inputting the target association rule, power grid disaster data and meteorological data into a pre-trained early warning model to obtain a power grid disaster risk early warning result.

2. The method of claim 1, wherein, Before screening a single feature set that meets a preset support degree from the power grid disaster data and meteorological data according to the preset support degree obtained by a genetic algorithm for each occurrence region, the method further comprises: obtaining an initial support degree, an initial confidence degree, an average value between the initial support degree and the initial confidence degree, and weights corresponding to the initial support degree, the initial confidence degree and the average value respectively; iteratively optimizing the initial support degree, the initial confidence degree, the average value and the weights corresponding thereto by a genetic algorithm using a preset fitness function to output the preset support degree and the preset confidence degree.

3. The method of claim 2, wherein, iteratively optimizing the initial support degree, the initial confidence degree, the average value and the weights corresponding thereto by a genetic algorithm using a preset fitness function to output the preset support degree and the preset confidence degree, comprising: generating a plurality of candidate support degrees and a plurality of candidate confidence degrees within a preset value range of the initial support degree and the initial confidence degree; constructing an initial population from the initial support degree, the initial confidence degree, the plurality of candidate support degrees and the plurality of candidate confidence degrees, wherein each individual in the initial population represents a group of support degrees and confidence degrees; generating new individuals by crossing and mutating the individuals in the initial population, and calculating the fitness value of each individual by using the preset fitness function in combination with the support degree, the confidence degree, the average value between the support degree and the confidence degree, the support degree and the confidence degree and the average value corresponding to the weights thereof in each individual; The individual with the highest fitness is selected to enter the next generation, and the process of generating new individuals and calculating individual fitness values is repeated until a preset termination condition is met, and the support and fitness of the individual with the highest fitness value are output as the preset support and preset fitness.

4. The method of claim 1, wherein, The target item combination feature set is filtered from the plurality of initial association sets in each occurrence area, including: A preset item set number is obtained; In each occurrence area, all combinations of item set lengths equal to the preset item set number are extracted from the plurality of initial association sets as initial feature sets between meteorological data and power grid disaster data; If the initial feature set is not empty, the frequency of each initial feature set appearing in the plurality of initial association sets is counted, and the initial feature set with an appearance frequency satisfying a preset support is determined as a candidate initial feature set; On the basis of the preset item set number, a specified item number is added to obtain an updated item set number, and based on the meteorological data and power grid disaster data in the candidate initial feature set, an updated feature set with an item set length of the updated item set number is generated by combination; If the updated feature set is not empty, the appearance frequency of each updated feature set in the plurality of initial association sets is counted, and the updated feature set with an appearance frequency satisfying the preset support is determined as a candidate updated feature set; The operations of increasing the item set length, generating the feature set, and filtering the support are repeatedly performed until no new feature set satisfying the preset support can be generated, and the candidate updated feature set generated last time is taken as the target item combination feature set.

5. The method of claim 1, wherein, The target association rule, the power grid disaster data, and the meteorological data are jointly input into a pre-trained early warning model to obtain a power grid disaster risk early warning result, including: The power grid disaster data and the meteorological data are spatio-temporally matched according to occurrence time and occurrence area to obtain spatio-temporal sequence data; The spatio-temporal sequence data is processed by a convolutional neural network to obtain a high-order feature sequence; The high-order feature sequence and the target association rule are jointly input into a convolutional attention module embedded between a convolutional neural network and a long short-term memory network in the early warning model, and an attention-enhanced feature sequence is output, wherein the convolutional attention module is used to give higher attention weights to corresponding feature channels and spatial regions according to the meteorological data and the power grid disaster data indicated by the target association rule; The attention-enhanced feature sequence is input into the long short-term memory network, so that the long short-term memory network models the time evolution law through the gating mechanism to generate the power grid disaster risk early warning result.

6. The method of claim 5, wherein, The power grid disaster data and the meteorological data are spatio-temporally matched according to occurrence time and occurrence area to obtain spatio-temporal sequence data, including: According to the occurrence time and the occurrence area of each piece of power grid disaster data and meteorological data, it is attributed to the spatio-temporal unit composed of the corresponding spatial grid and time interval; For multiple data of the same parameter belonging to the same spatio-temporal unit, the average value is used for aggregation to generate a unified data value under the spatio-temporal unit; For the target spatio-temporal unit without direct observation data, the inverse distance weighted interpolation method is used to generate the data value of the target spatio-temporal unit based on the observation values of the adjacent spatial grids; The processed power grid disaster data and the meteorological data are determined as spatio-temporal sequence data.

7. The method of claim 5, wherein, After generating the power grid disaster risk early warning result, when receiving new power grid disaster data and meteorological data, the gating parameters of the long short-term memory network are updated by using an online learning mechanism in the early warning model.

8. A power grid disaster early warning device, characterized by, The device comprises: a data acquisition unit configured to acquire preprocessed power grid disaster data and meteorological data in a target time period, wherein each piece of power grid disaster data and meteorological data has a corresponding occurrence time and occurrence region; a single set screening unit configured to, for each occurrence region, screen a single feature set satisfying a preset support degree from the power grid disaster data and meteorological data acquired by the data acquisition unit according to the preset support degree obtained by the genetic algorithm; a time period determination unit configured to, for each occurrence region, determine a plurality of occurrence time periods according to the occurrence times of the meteorological data in the single feature set obtained by the single set screening unit; a window determination unit configured to construct a plurality of response windows offset backward with respect to each occurrence time period determined by the time period determination unit as a reference time period, wherein the response window comprises the reference time period and a plurality of time periods offset backward by a preset time length with respect to the reference time period; an association set determination unit configured to associate the meteorological data and the power grid disaster data corresponding to the reference time period determined by the window determination unit, and associate the meteorological data corresponding to the reference time period and the power grid disaster data occurring in each response window, to form a plurality of initial association sets of meteorological data and power grid disasters; a feature set screening unit configured to screen a target item combination feature set from the plurality of initial association sets in each occurrence region determined by the association set determination unit; a rule determination unit configured to screen a combination feature with a confidence degree higher than a preset confidence degree from the target item combination feature set obtained by the feature set screening unit as a target association rule, according to the preset confidence degree obtained by the genetic algorithm; a risk early warning unit configured to input the target association rule determined by the rule determination unit, the power grid disaster data and the meteorological data into a pre-trained early warning model to obtain a power grid disaster risk early warning result.

9. A storage medium, characterized by The storage medium comprises a stored program, wherein the program controls the device where the storage medium is located to execute the power grid disaster early warning method in any one of claims 1 to 7 when the program is running.

10. A processor, comprising: The processor is configured to run a program, wherein the program executes the power grid disaster early warning method in any one of claims 1 to 7 when the program is running.

Images