Method and system for predicting key factors of drought-flood abrupt change based on causal inference

Abstract

The application provides a drought-flood sudden change key factor prediction method and system based on causal inference, relates to the drought-flood sudden change prediction technical field, and comprises the following steps: acquiring critical state data of a drought-flood system when the drought-flood system transits from a metastable state to a sudden change state; determining a critical starting point of drought-flood sudden change; screening a precursor factor of an order parameter capable of driving the system to transit from a metastable state to a critical state; excluding a pseudo factor that only accompanies a critical state but has no driving effect; verifying the universality of each factor in a candidate factor set in triggering a critical phase change of the system; screening a factor with cross-system domain universal driving performance and determining the factor as a drought-flood sudden change key factor; constructing a critical phase change deduction model; predicting whether the system will have a critical transition in the future 1-15 days and generating a prediction report. By means of a critical state fingerprint matching method, multi-dimensional core fingerprint characteristics are extracted, the critical starting point of drought-flood sudden change can be accurately determined, a complete critical state event list associated with geographical environment information is generated, and the accuracy of critical event identification is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of drought-flood sudden change prediction, and particularly relates to a drought-flood sudden change key factor prediction method and system based on causal inference. BACKGROUND

[0002] Drought-flood sudden change is an extreme hydrological event in which a regional hydrological system quickly changes from a drought state to a flood state under specific meteorological and hydrological conditions. Its occurrence is closely related to multiple factors such as atmospheric circulation anomalies, changes in monsoon activity, and changes in underlying surface conditions. Under the background of global climate change, the frequency of extreme precipitation events increases, and in some areas, the previous prolonged drought leads to soil layer shrinkage and a decrease in vegetation coverage. Subsequent concentrated heavy rainfall is difficult to effectively infiltrate the soil, easily forming surface runoff, and thus triggering drought-flood sudden change. Such events are particularly common in monsoon climate zones and semi-arid and semi-humid transitional zones.

[0003] From the perspective of technical research, the monitoring of drought-flood sudden change relies on multi-source data collection systems such as meteorological observation, hydrological monitoring, and remote sensing perception. Among them, the rainfall monitoring of meteorological stations, the runoff and water level observation of hydrological stations, and the soil moisture and vegetation coverage inversion of satellite remote sensing are the core means to obtain drought state evolution data. In the data processing layer, early methods mainly use moving average method and trend analysis method to identify the mutation characteristics of hydrological elements to determine the drought-flood state transition node. With the introduction of complex system theory, research has begun to focus on the evolution law of the transition of the hydrological system from a metastable state to a critical state, regarding drought-flood sudden change as a manifestation of system critical phase transition, and trying to capture the state transition trend by extracting system critical precursor signals.

[0004] Drought-flood sudden change key factors refer to meteorological, hydrological, and underlying surface elements that have a significant driving or indicating effect on the rapid transition of the hydrological system from a drought state to a flood state. The identification and analysis of these key factors are the core link to improve the accuracy of drought-flood sudden change warning, and are directly related to the depth of understanding of the event mechanism and the pertinence of response measures. In the dimension of meteorological elements, the spatiotemporal distribution characteristics of rainfall are the basic factors that trigger drought-flood sudden change. The lack of precipitation during the previous drought period determines the soil drought base, while the intensity, duration, and spatial coverage of concentrated rainfall in the later period directly affect the formation speed and scale of surface runoff. In addition, temperature changes indirectly regulate the critical conditions of drought-flood transition by affecting evaporation and changing soil moisture conditions, becoming an important auxiliary driving factor.

[0005] General drought-flood sudden change key factor prediction methods lack comprehensive perception of drought-flood characteristics, often only using time division or a single indicator for judgment, resulting in insufficient dimensionality and effectiveness of the collected critical precursor data, making it difficult to fully reflect the true state of the system transition process. In terms of factor screening, the evaluation of factor driving effectiveness lacks data support, and the selected precursor factors lack pertinence and reliability.

[0006] In order to solve the defects in the prior art, the technical scheme provides a drought-flood sudden change key factor prediction method and system based on causal inference. SUMMARY

[0007] The present application provides a drought-flood sudden change key factor prediction method and system based on causal inference to solve the defects in the prior art.

[0008] In one aspect, the present application provides a drought-flood sudden change key factor prediction method based on causal inference, comprising:

[0009] S1: Capture the critical phase change precursor of the complex system, use non-invasive multi-field coupling perception to obtain the critical state data of the drought-flood system when it transits from metastable state to sudden change state, and form a critical precursor data set;

[0010] S2: Use critical state fingerprint matching method to determine the critical starting point of drought-flood sudden change, mark the event type, and form a critical state event list;

[0011] S3: Based on the critical precursor data set and the critical state event list, screen the precursor factors of the order parameter that can drive the system to transit from metastable state to critical state; exclude pseudo factors that only accompany critical state but have no driving effect, and obtain a candidate factor set driven by self-organized criticality;

[0012] S4: Verify the universality of each factor in the candidate factor set in triggering the critical phase change of the system in different hydrological climate systems; screen the factors with cross-system domain universal driving property, determine them as drought-flood sudden change key factors, and form a universal key factor list;

[0013] S5: Based on the causal relationship between the drought-flood sudden change key factor and the self-organized critical phase change of the system, a critical phase change deduction model is constructed; based on the critical phase change deduction model, it is predicted whether the system will occur critical transition in the future 1-15 days, the order parameter contribution degree of each key factor in promoting the critical phase change is output, and a prediction report is generated.

[0014] According to the drought-flood sudden change key factor prediction method based on causal inference provided by the present application, in step S1, the step of obtaining the critical state data of the drought-flood system when it transits from metastable state to sudden change state comprises:

[0015] S11: For historical drought-flood sudden change events, analyze the runoff mutation point by sliding window method, and define the period from 30 days before the mutation point to the day of the mutation point as the transition stage;

[0016] S12: Based on the transition stage, verify the rationality of the soil humidity drop range, output the transition stage data segment labeled with the metastable state starting time to the sudden change state triggering time;

[0017] S13: Based on the transition stage data segment, screen the key indicators of each perception domain;

[0018] S14: Structure integration in time sequence and spatial distribution, supplement of metadata such as data source and collection equipment, and formation of critical precursor data set.

[0019] According to the dry-wet sudden change key factor prediction method based on causal inference provided by the application, the step of determining the critical starting point of dry-wet sudden change in step S2 comprises:

[0020] S21: With the system equilibrium state reference value as a reference, the characteristics deviating from the reference in the critical precursor data are analyzed, the core fingerprint characteristics including daily rainfall increase, 10cm soil moisture daily decrease, runoff-rainfall response lag time and NDVI spatial heterogeneity are extracted, and a fingerprint characteristic type library is formed;

[0021] S22: Based on the critical precursor data set, the value of the core fingerprint characteristic is calculated in combination with the fingerprint characteristic type library, and a quantized fingerprint characteristic matrix is generated;

[0022] S23: According to historical dry-wet sudden change events, the characteristic values of the quantized fingerprint characteristic matrix within 7 days before and after the starting time point are extracted as a critical sample set, the threshold value of each characteristic is determined by using the percentile method, the threshold value effectiveness is verified by using the ROC curve, and a critical threshold table is output;

[0023] S24: The quantized fingerprint characteristic matrix is checked in time and space units, when a unit simultaneously satisfies the supercritical threshold value of the fingerprint characteristic, it is marked as a potential critical starting point, and then compared with historical dry-wet sudden change events, a critical starting point list containing the starting time and spatial coordinates is output, and the event type is marked;

[0024] S25: A unique event ID is assigned to each potential critical starting point, the corresponding precursor data segment is associated, the geographical environment information at the time of event occurrence is supplemented, and a critical state event list is generated.

[0025] According to the dry-wet sudden change key factor prediction method based on causal inference provided by the application, the step of screening the order parameter precursor factor capable of driving the system from a metastable state to a critical state in step S3 comprises:

[0026] S31: According to the self-organized critical theory of complex system, the critical precursor data set and the critical state event list are combined, the core analysis dimension is selected, the dimension differentiation ability for the critical state is verified by pre-experiment, the analysis dimension system is formed, and the calculation method of each dimension is marked;

[0027] S32: The potential factors are extracted from the critical precursor data set, the Pearson correlation coefficient of each potential factor and the critical event is calculated, and then combined with the analysis dimension evaluation, an initial screening precursor factor set is formed, and the factor name, correlation degree value and dimension contribution value are output.

[0028] According to the drought-flood sudden change key factor prediction method based on causal inference provided by the application, in step S3, the step of obtaining the candidate factor set comprises:

[0029] S33: Taking the pre-screening precursor factor set as the object, single factor disturbance and multi-factor combined disturbance are combined, the disturbance frequency is set to once a day, the monitoring duration is 15 days after the disturbance, the monitoring index is whether the critical state is triggered, and an experiment scheme containing the disturbance sequence, amplitude and monitoring index is formed;

[0030] S34: Based on the Python SimPy simulation platform, the critical precursor data set is imported, the disturbance is automatically executed according to the experiment scheme, the monitoring index value is calculated after each disturbance, whether the critical state is jumped to, the response delay time and the response intensity are recorded, the experiment is repeated 30 times for each factor, and finally the micro-disturbance response result set is integrated;

[0031] S35: The pseudo-factor determination standard is set, the pre-screening factors are checked one by one according to the micro-disturbance response result set, the factors meeting the pseudo-factor standard are removed, the remaining factors are sorted according to the driving effectiveness score, and the candidate factor set of self-organized critical driving is formed.

[0032] According to the drought-flood sudden change key factor prediction method based on causal inference provided by the application, in step S4, the step of verifying the universality of each factor in the candidate factor set to trigger the system critical phase change comprises:

[0033] S41: According to the classification of hydrological and climatic systems, a preset number of typical river basins are selected in each system domain, the historical critical precursor data set, the historical critical state event list and the geographic parameters of each river basin are collected, and a cross-system domain verification sample set is formed;

[0034] S42: For each verification sample in the cross-system domain verification sample set, the driving effect of each factor in the candidate factor set is tested by using the micro-disturbance experiment method, the critical jump success rate after single factor disturbance and the synergistic driving rate of multi-factor combined disturbance are counted, the effectiveness result and the specific success rate of each factor in the sample are recorded, the data missing sample is replaced by the similar river basin data in the same system domain, and a single-domain verification result table is output;

[0035] S43: The universality evaluation index is defined, the cross-domain passing rate and the coefficient of variation of each candidate factor are calculated based on the single-domain verification result table, and a factor cross-domain universality matrix is formed.

[0036] According to the drought-flood sudden change key factor prediction method based on causal inference provided by the application, in step S4, the step of generating the universality key factor list comprises:

[0037] S44: The universality screening threshold is set, the factors meeting the universality screening threshold are screened from the candidate factor set, the physical reasonableness of the factors is confirmed, and the drought-flood sudden change key factor set is output.

[0038] 45: Integrate the drought and flood sudden change key factor set and the previous data according to the preset format, supplement the monitoring method and data source of the factor, and generate a universal key factor list.

[0039] According to the drought and flood sudden change key factor prediction method based on causal inference provided by the application, in step S5, the step of constructing the critical phase transition deduction model comprises:

[0040] S51: Based on the causal relationship between the key factor and the critical phase transition of the system, a causal inference neural network is selected as the core model, including an output layer, a hidden layer and an output layer, an embedded sequence parameter contribution degree calculation module, a network structure is optimized by referring to an existing drought and flood prediction model, and a model architecture diagram is drawn;

[0041] S52: The model is trained using historical critical event data, and a trained model, a parameter configuration table and an error evaluation report are output.

[0042] According to the drought and flood sudden change key factor prediction method based on causal inference provided by the application, in step S5, the step of outputting the sequence parameter contribution degree of each key factor in promoting the critical phase transition comprises:

[0043] S53: Based on non-invasive multi-field coupling perception, real-time data is automatically collected every day, and the real-time data is preprocessed to output a standardized real-time input data set;

[0044] S54: The standardized real-time input data set is imported into the trained model, and the future 1-15 day system evolution trajectory is simulated, when the critical transition probability of any day is greater than a preset probability threshold, it is marked as a high-risk transition day, and the sequence parameter contribution degree of each key factor is calculated through the SHAP value algorithm, and a 1-15 day critical transition prediction result table and a contribution degree matrix are output;

[0045] S55: According to the structure of background, method, prediction result, key factor contribution analysis and risk prompt, insert the prediction probability graph and the contribution degree pie chart, supplement the model error range and data source description, and output the prediction report according to the standard format of the hydrological department.

[0046] The application also provides a drought and flood sudden change key factor prediction system based on causal inference, comprising:

[0047] A multi-field coupling perception module is used to collect multi-dimensional data related to the drought and flood system in real time;

[0048] A data construction module is used to construct a critical precursor data set based on multi-dimensional data;

[0049] A critical starting point determination module is used to locate the critical starting point of the drought and flood sudden change through fingerprint matching technology to form a traceable critical state event list.

[0050] a factor screening module configured to screen precursor factors capable of driving critical phase transition of the system from the critical precursor dataset, eliminate pseudo factors, and form a candidate factor set;

[0051] a key factor determination module configured to verify the universality of the candidate factors in different hydrological and climatic systems based on the candidate factor set, and determine a drought-flood rapid transition key factor;

[0052] a model construction module configured to construct a causal inference neural network model based on the causal relationship between the key factor and the critical phase transition, import real-time data into the model, simulate a system evolution trajectory, calculate the order parameter contribution degree of each key factor through a SHAP value algorithm, and output a prediction result table and a contribution degree matrix;

[0053] a report generation module configured to generate a prediction report based on the prediction result table and the contribution degree matrix.

[0054] The drought-flood rapid transition key factor prediction method and system based on causal inference provided by the application can accurately determine the critical starting point of drought-flood rapid transition by extracting multi-dimensional core fingerprint features through the critical state fingerprint matching method and verifying the threshold value through the ROC curve, generate a complete critical state event list associated with geographical environmental information, and improve the accuracy of critical event identification. In the factor screening, pseudo factors that only accompany critical states but have no driving effect are effectively eliminated through Pearson correlation analysis, single and multi-factor disturbance experiments, and micro-disturbance tests on the Python SimPy simulation platform, and candidate factors with self-organized critical driving effect are screened out, thereby ensuring the effectiveness and pertinence of the factors. The universality of the key factor in different hydrological and climatic systems is ensured through the universality verification of cross-system domain multi-basin samples combined with the universality index calculation, thereby expanding the application range of the method. The critical phase transition deduction model constructed based on the causal inference neural network can realize timely prediction of the system critical jump in the future 1-15 days, and the order parameter contribution degree of each key factor is output through the SHAP value algorithm to generate a prediction report that meets the standards of the hydrological department, contains visual charts and error explanations, and significantly improves the scientificity, accuracy and practicability of drought-flood rapid transition prediction, thereby providing clear factor action basis for early warning and response decision-making of drought-flood rapid transition disasters. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0056] Fig. 1is a flowchart of a drought-flood sudden change key factor prediction method based on causal inference provided by an embodiment of the present application;

[0057] Fig. 2 is a flowchart of determining a critical starting point of drought-flood sudden change in the drought-flood sudden change key factor prediction method based on causal inference provided by an embodiment of the present application;

[0058] Fig. 3 is a structural schematic diagram of a drought-flood sudden change key factor prediction system based on causal inference provided by an embodiment of the present application. DETAILED DESCRIPTION

[0059] To make the objectives, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described below in detail with reference to the drawings in the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0060] Embodiment one: the following will be described in combination with Figs. 1-3 The drought-flood sudden change key factor prediction method and system based on causal inference of the present application are described.

[0061] As shown in Figs. 1-2 The drought-flood sudden change key factor prediction method based on causal inference provided by an embodiment of the present application comprises the following steps.

[0062] S1: capture the critical phase change precursor of a complex system, adopt non-invasive multi-field coupling perception, obtain the critical state data of the drought-flood system when it transits from a metastable state to a sudden change state, and form a critical precursor data set. Referring to the research on the interaction mechanism of the drought-flood system, such as the “hydrology-weather-soil-vegetation” feedback chain, the core domain sensitive to the drought-flood conversion is preferentially selected, for example, the combination of precipitation-soil moisture-runoff-NDVI, and the data availability of the research area is combined, for example, the precipitation and NDVI domain monitored by the satellite is preferentially retained in remote areas, and finally 3-5 key coupling perception domains are determined through expert demonstration to form a list.

[0063] The weather domain adopts Fengyun-3 satellite (precipitation data, spatial resolution 10km, daily scale) and ground wireless weather station (air temperature, wind speed, hourly scale). The soil domain is deployed with TDR soil moisture sensor (buried depth 10 / 30cm, collected every 30 minutes). The hydrology domain uses radar water level gauge (runoff data, collected every 15 minutes). The vegetation domain uses Landsat-8 satellite NDVI data (8-day scale), and the equipment is deployed in layers according to “basin outlet-middle reaches-upper reaches”, and the time synchronization (unified UTC time) of each equipment is ensured during debugging. The output system configuration scheme includes equipment model, deployment coordinates and collection frequency.

[0064] According to the configuration scheme, remote sensing data is acquired through satellite data, and ground equipment transmits the data to a data center through LoRa wireless transmission. Daily automatic triggering of data collection tasks is performed, and missing transmission data (such as sensor failure) is marked as “to be completed”. Finally, the original multi-field data set containing time stamp, spatial coordinates, and index value is integrated.

[0065] Using the Python pandas library, the original data is first removed by the 3σ principle, then linear interpolation is used to complete the short-time missing values, Markov chain is used to complete the long-time missing values, and finally Z-score standardization is used, which is represented as: , Y is the standardized original data value, B is the original data value, μ is the mean, and σ is the standard deviation. Eliminate the dimension, output the preprocessed data set, and ensure that the mean of each index is 0 and the standard deviation is 1.

[0066] In step S1, the step of obtaining the critical state data of the drought and flood system when it transitions from a metastable state to a sudden state includes:

[0067] S11: Historical drought and flood sudden change events, the sliding window method is used to analyze the runoff mutation point, and the window size is set to 2 times the event period, for example, if the historical event lasts for 15 days, then the window is set to 30 days. Combined with the Pettitt mutation test, the Pettitt mutation test is a non-parametric test method that identifies the mutation point of the runoff sequence. The 30 days before the mutation point to the day of the mutation point are defined as the transition stage, which covers the energy accumulation period from stability to instability.

[0068] S12: Based on the transition stage, combined with the soil moisture drop amplitude to verify the reasonableness of the stage, calculate the difference between the minimum value of 10cm soil moisture in the transition stage and the average value of the previous 30 days. If the drop amplitude is > 20%, it is considered that the stage division is reasonable, and random fluctuations are avoided to be misjudged as a transition stage. Comprehensive runoff mutation point and soil moisture response, output the transition stage data segment labeled from the start time of the metastable state (30 days before the mutation point) to the trigger time of the sudden state (the day of the mutation point).

[0069] S13: Based on the transition stage data segment, filter the key indicators of each sensing domain. The meteorological domain selects the daily rainfall increase, which is the ratio of the daily rainfall to the average rainfall of the previous 30 days. The soil domain selects the 10cm soil moisture daily drop, which is the difference between the current day and the previous day's humidity value. The hydrological domain selects the runoff-precipitation response lag time, which is the time difference between the runoff peak and the precipitation peak. The vegetation domain selects the NDVI spatial heterogeneity, which is the standard deviation of NDVI, reflecting the uniformity of vegetation coverage, ensuring that the indicators are directly related to the critical phase change.

[0070] S14: Structure integration according to time sequence and spatial distribution, spatial fusion of different source data through Kriging interpolation method; supplement data sources (including Fengyun-3 satellite, ground station, etc.), collection equipment (such as TDR sensor number), quality control (such as missing value has been interpolated) and other metadata, finally form a standardized critical precursor data set, containing four-dimensional information of time, space, index value and metadata.

[0071] S2: Using critical state fingerprint matching method, by comparing the "characteristic fingerprint" of the current state and the historical critical state, the critical starting point of drought-flood rapid transition is determined, the event type is marked, such as rainstorm type, drought type, and the critical state event list containing event metadata is formed.

[0072] In step S2, the step of determining the critical starting point of drought-flood rapid transition includes:

[0073] S21: Taking the system equilibrium state reference value as the reference, analyze the characteristics deviating from the reference in the critical precursor data, extract the core fingerprint characteristics including daily rainfall increase, 10cm soil moisture daily decrease, runoff-precipitation response lag time, NDVI spatial heterogeneity, form a fingerprint characteristic type library.

[0074] S22: Based on the critical precursor data set, combined with the fingerprint characteristic type library, calculate the value of the core fingerprint characteristics, generate the quantized fingerprint characteristic matrix.

[0075] S23: According to the historical drought-flood rapid transition event, extract the characteristic values of 7 days (15 days in total) before and after the corresponding starting time point in the quantized fingerprint characteristic matrix, such as event starting day T, then extract the characteristic data from T-7 to T+7, form a critical sample set, containing event period and non-event period data. The percentile method is used to determine the threshold value of each characteristic, that is, the 95% percentile of the daily rainfall increase in the critical sample set is taken as the supercritical threshold value. Through the ROC curve, true positive rate vs false positive rate, verify the effectiveness of the threshold value, output the critical threshold table.

[0076] S24: Check the quantized fingerprint characteristic matrix by time and space unit (traverse each grid cell and each hour data), when a cell meets the following conditions at the same time, it is marked as a potential critical starting point: daily rainfall increase > threshold value; 10cm soil moisture daily decrease > threshold value; runoff-precipitation response lag time < threshold value; NDVI spatial heterogeneity > threshold value; and compared with historical confirmed events, exclude false positives caused by abnormal values due to equipment failure, finally output the critical starting point list containing starting time, spatial coordinates and event type.

[0077] S25: Assign a unique event ID to each potential critical onset point, associate the corresponding precursor data segments (30 days before onset to 7 days after onset data), supplement the geographical environment information at the time of event occurrence, including terrain slope, land use type, soil texture, and previous drought days, and generate a critical state event list containing 12 fields (field examples: event ID, onset time, latitude, longitude, event type, precursor data path, terrain slope, land use type, soil texture, previous drought days, and verification status).

[0078] S3: Based on the critical precursor data set and the critical state event list, from the self-organized criticality of complex systems, screen the precursor factors of order parameters that can drive the system from metastable state to critical state, such as specific frequency band atmospheric turbulence energy input, critical combination of soil particle size distribution, and threshold value of synergistic effect of vegetation root water absorption. Through system perturbation experiment, introduce small turbulence energy, observe whether it accelerates the arrival of critical state, exclude pseudo factors that only accompany critical state but have no driving effect, and obtain the candidate factor set of self-organized critical driving.

[0079] In step S3, the step of screening the precursor factors of order parameters that can drive the system from metastable state to critical state includes:

[0080] S31: According to the self-organized critical theory of complex systems, through slow accumulation, eventually reach the critical state, small perturbation may trigger a chain reaction, combined with the critical precursor data set and the critical state event list, select the core analysis dimensions, including: energy accumulation dimension, soil moisture deficit accumulation; coupling strength dimension, precipitation-runoff correlation coefficient; response sensitivity dimension, derivative of NDVI to precipitation. Through pre-experiment, i.e. select 10 typical basins, calculate the mutual information value of each dimension and critical event, verify the dimension's ability to distinguish critical state (mutual information > 0.3 is considered as effective dimension), form an analysis dimension system containing 3 dimensions, and mark the calculation method of each dimension, such as energy accumulation dimension = Σ(field water capacity-soil moisture) × time step.

[0081] S32: Extract potential factors from the critical precursor data set, including meteorological factors: daily rainfall, air temperature; soil factors: 10 cm soil moisture, soil temperature; hydrological factors: runoff, water level; vegetation factors: NDVI, leaf area index, a total of 12 factors. Calculate the Pearson correlation coefficient of each potential factor and critical event, and then evaluate it based on the analysis dimension, which is expressed as: energy accumulation dimension contribution value = correlation coefficient of factor and energy accumulation dimension × 0.5 + correlation coefficient of factor and critical event × 0.5, screen out factors with absolute value of correlation coefficient > 0.5 and dimension contribution value > 0.3, form the initial screening precursor factor set, and output factor name, correlation value, dimension contribution value, and screening basis. The calculation method of Pearson correlation coefficient is represented as:

[0082]

[0083] Wherein, X is a factor value, Y is an event label, a value of 1 indicates occurrence, and a value of 0 indicates non-occurrence.

[0084] In step S3, the step of obtaining the candidate factor set comprises:

[0085] S33: Design a perturbation experiment scheme for the pre-screening precursor factor set: perturbation type: single factor perturbation (change a factor value), multi-factor combination perturbation (change two or more factor values); perturbation frequency: once a day (simulate daily weather / human intervention); perturbation amplitude: set according to gradient (such as 5%, 10%, 15% increase in precipitation, 5%, 10%, 15% decrease in soil moisture); monitoring duration: 15 days after perturbation (covering the key window period of critical phase change); monitoring indicators: whether the critical state is triggered (yes / no), response delay time (number of days from perturbation to first triggering of critical state), response intensity (NDVI drop amplitude or runoff increase after triggering).

[0086] S34: Based on the Python SimPy simulation platform, i.e. a discrete event simulation library, import the critical precursor data set as baseline scene data, automatically perform perturbation according to the experiment scheme, initialize the system state, i.e. load the soil moisture, runoff, etc. data of the baseline scene; perform single factor perturbation for each factor, such as increasing the daily precipitation by 10%; run the simulation for 15 days after the perturbation, record whether the critical state is transitioned to, the response delay time, and the response intensity; repeat the experiment for each factor 30 times to reduce random errors. Calculate the monitoring indicator value after each perturbation, record whether the critical state is transitioned to, the response delay time, and the response intensity, repeat the experiment for each factor 30 times, and finally integrate into a micro-perturbation response result set.

[0087] S35: Set the pseudo-factor determination standard, and check the pre-screening factors one by one according to the micro-perturbation response result set, eliminate the factors meeting the pseudo-factor standard, and sort the remaining factors according to the driving effectiveness score to form a candidate factor set for self-organized critical driving. The scoring formula is expressed as:

[0088]

[0089] Wherein, A is the score value, a1 is the triggering success rate, a2 is the average response intensity, t YC is the average delay time.

[0090] The pseudo-factor determination standard includes: triggering success rate < 10% (small perturbation hardly triggers transition); average response delay time > 30 days (exceeding the critical phase change key window period); and no linear correlation between response intensity and perturbation amplitude.

[0091] S4: Conduct cross-system domain universality verification. In different hydrological and climatic systems (including monsoon / non-monsoon basins, humid / arid regions), verify the universality of each factor in the candidate factor set triggering the critical phase transition of the system, i.e., when the factor reaches the critical value, it triggers the critical transition of drought and flood regardless of the system background. Select factors with cross-system domain universality driving and determine them as key factors of drought and flood rapid transition to form a list of universal key factors.

[0092] In step S4, the step of verifying the universality of each factor in the candidate factor set triggering the critical phase transition of the system includes:

[0093] S41: Classify by hydrological and climatic system, i.e., based on the Koppen climate classification, divide into humid regions (C), semi-arid regions (BS), and arid regions (BW). Select a predetermined number (e.g., 3) of typical basins in each system domain, which must meet the following criteria: area ≥ 1000 km², complete observation data, historical drought and flood rapid transition events, collect historical critical precursor data sets (2015-2024), historical critical state event lists, and geographic parameters (area, average slope, soil type, vegetation coverage), form a cross-system domain verification sample set containing 3 system domains x 3 basins = 9 verification samples.

[0094] S42: For each verification sample in the cross-system domain verification sample set, use the perturbation experiment method to test the driving effect of each factor in the candidate factor set, including: single factor perturbation: apply a 10% increase perturbation to factor A, repeat 30 times, and calculate the trigger success rate; multi-factor combination perturbation: apply a 10% increase perturbation to factors A+B, repeat 30 times, and calculate the cooperative driving rate; record the effectiveness results (trigger success rate, cooperative driving rate) and specific success rate of each factor in the sample; for samples with missing data, replace them with similar data from the same system domain, output a single-domain verification result table containing sample ID, factor name, single-factor trigger success rate, multi-factor cooperative driving rate, and data missing situation.

[0095] S43: Define universality evaluation indicators, including: cross-domain pass rate = number of domains passing effectiveness verification in N system domains / N, pass standard is single-factor trigger success rate > 70%; coefficient of variation = standard deviation / mean of factor trigger success rate in each domain, reflecting the stability of the effect, coefficient of variation < 0.2 is considered stable. Based on the single-domain verification result table, calculate the cross-domain pass rate and coefficient of variation of each candidate factor to form a factor cross-domain universality matrix.

[0096] In step S4, the step of generating the list of universal key factors includes:

[0097] S44: Set the threshold for universal screening, cross-domain pass rate ≥ 80%, coefficient of variation ≤ 0.2, screen the factors that meet the threshold from the candidate factor set, confirm the physical reasonableness of the factors (for example, factor A is "soil moisture deficit accumulation", its increase will lead to decreased infiltration and increased surface runoff, which is consistent with the physical mechanism of sudden drought and flood), and output the set of key factors for sudden drought and flood.

[0098] 45: Integrate the set of key factors for sudden drought and flood with the previous data in a pre-set format (JSON or Excel), including critical precursor data, critical state event list, supplementary factors monitoring methods such as "soil moisture using TDR sensor, buried depth 10cm and 30cm, collecting every 30 minutes", data sources (usually meteorological data from China Meteorological Administration, soil data from basin monitoring station), generate a universal key factor list containing 15 fields (field examples: factor name, physical meaning, monitoring method, data source, cross-domain pass rate, coefficient of variation, dominant mechanism).

[0099] S5: Based on the causal relationship between the key factors for sudden drought and flood and the self-organized critical phase transition of the system, build a critical phase transition deduction model. Based on the critical phase transition deduction model, input real-time critical precursor data and key factor state, simulate the evolution trajectory of the system from metastable state to critical state, predict whether the system will undergo critical transition in the next 1-15 days, output the order parameter contribution of each key factor in promoting critical phase transition, and generate a prediction report.

[0100] In step S5, the steps of building the critical phase transition deduction model include:

[0101] S51: Based on the causal relationship between the key factors and the critical phase transition of the system, select a causal inference neural network as the core model, combining the causal graph structure and the nonlinear fitting ability of the neural network. The model architecture includes: input layer, node number = number of key factors, such as 2 factors, then the input layer has 2 nodes; hidden layer 1, 64 nodes, ReLU activation function, extract the nonlinear relationship between factors; hidden layer 2, 32 nodes, tanh activation function, capture causal dependence; output layer, 1 node, Sigmoid activation function, output critical transition probability; embedded order parameter contribution calculation module, based on SHAP value decomposition of output layer input, calculate the marginal contribution of each factor to the output; refer to existing drought and flood prediction models, optimize network structure, adjust hidden layer node number, learning rate, draw model architecture diagram, label each layer type, node number, activation function, connection method.

[0102] S52: Train the model using historical critical event data, steps include: data division: divide into training set, validation set, test set according to 7:2:1; set the loss function, expressed as binary cross-entropy loss, formula is:

[0103]

[0104] wherein, L denotes binary cross-entropy loss, y is the real label, is the model prediction probability.

[0105] Optimization algorithm: Adam optimizer (learning rate is usually 0.001); training process: 50 iterations, batch size 32, stop when the validation set loss does not decrease for 10 consecutive rounds; output the trained model, parameter configuration table (including learning rate, batch size, number of iterations), error evaluation report (including accuracy, precision, recall, F1 score, etc.).

[0106] In step S5, the step of outputting the order parameter contribution degree of each key factor in promoting the critical phase transition comprises:

[0107] S53: Based on non-invasive multi-field coupling perception, real-time data is automatically collected every day, the acquisition methods include satellite remote sensing data and ground station observation data, and the real-time data is preprocessed, including outlier elimination, missing value completion, Z-score standardization, etc., to output the standardized real-time input data set, the time window is the data of the past 30 days, which is used to predict the future 15 days.

[0108] S54: The standardized real-time input data set is imported into the trained model to simulate the system evolution trajectory in the future 1-15 days. When the critical transition probability of any day is greater than the preset probability threshold, it is marked as a high-risk transition day. At the same time, the SHAP value algorithm is used to calculate the order parameter contribution degree of each key factor, and the critical transition prediction result table and the contribution degree matrix of 1-15 days are output.

[0109] S55: According to the structure of background, method, prediction result, key factor contribution analysis and risk prompt, insert the prediction probability graph and contribution degree pie chart, supplement the model error range and data source explanation, and typeset according to the standard format of the hydrological department, and output the prediction report. Specifically: background, briefly describe the general situation of the research area (basin area, climate type, drought and flood events in recent years); method, summarize the key factor selection process and model construction method; prediction result, list the high-risk transition days in the future 1-15 days; key factor contribution analysis, insert the contribution degree pie chart (showing the contribution proportion of each factor) and the text explanation (such as soil moisture deficit cumulative amount is the main driving factor of this flood); risk prompt, put forward suggestions (such as high-risk days need to strengthen reservoir regulation to prevent floods); supplement the model error range (such as the prediction error of transition probability is ±5%) and the data source explanation (such as meteorological data from China Meteorological Administration, soil data from basin monitoring station), and typeset according to the standard format of the hydrological department.

[0110] In summary, the method and system for predicting key factors of sudden change from drought to flood based on causal inference provided by the application can accurately capture the critical precursor data of the transition of the drought-flood system from metastable state to sudden change state, providing reliable data support for subsequent analysis, through non-intrusive multi-field coupling perception combined with sliding window method and soil moisture verification; With the help of critical state fingerprint matching method, the multi-dimensional core fingerprint characteristics are extracted and verified by ROC curve threshold, which can accurately determine the critical starting point of sudden change from drought to flood, generate a complete list of critical state events related to geographical environment information, and improve the accuracy of critical event identification. In factor screening, through Pearson correlation analysis, single and multi-factor disturbance experiment and micro-disturbance test on Python SimPy simulation platform, pseudo factors that only accompany critical state but have no driving effect are effectively eliminated, and candidate factors with self-organized critical driving effect are screened out, ensuring the effectiveness and pertinence of the factors. Through the universality verification of cross-system domain multi-basin samples, combined with the universality index calculation, the applicability of the key factors in different hydrological climate systems is ensured, and the application range of the method is expanded; the critical phase transition deduction model based on causal inference neural network can realize the timely prediction of the system critical jump in the future 1-15 days, and then output the sequence parameter contribution degree of each key factor through the SHAP value algorithm, generate a prediction report conforming to the standard of hydrological department, containing visual charts and error explanation, significantly improve the scientificity, accuracy and practicability of the prediction of sudden change from drought to flood, and provide clear factor action basis for early warning and decision-making of sudden change from drought to flood disaster.

[0111] As shown in Fig. 3 The application also provides a key factor prediction system for sudden change from drought to flood based on causal inference, which comprises a multi-field coupling perception module, a data construction module, a critical starting point determination module, a factor screening module, a key factor determination module, a model construction module and a report generation module.

[0112] The multi-field coupling perception module is used for real-time acquisition of multi-dimensional data related to the drought-flood system.

[0113] The data construction module is used for constructing a critical precursor data set based on the multi-dimensional data.

[0114] The critical starting point determination module is used for locating the critical starting point of sudden change from drought to flood through fingerprint matching technology, and forming a traceable critical state event list.

[0115] The factor screening module is used for screening precursor factors capable of driving the critical phase transition of the system from the critical precursor data set, eliminating pseudo factors, and forming a candidate factor set.

[0116] The key factor determination module is used for verifying the universality of the candidate factors in different hydrological climate systems based on the candidate factor set, and determining the key factors of sudden change from drought to flood.

[0117] The model construction module is configured to construct a causal inference neural network model based on the causal relationship between the key factors and the critical phase change. Real-time data is imported into the model to simulate the evolution trajectory of the system. The SHAP value algorithm is used to calculate the contribution degree of the order parameter of each key factor, and a prediction result table and a contribution degree matrix are output.

[0118] The report generation module is configured to generate a prediction report based on the prediction result table and the contribution degree matrix.

[0119] Through the description of the above embodiments, those skilled in the art can clearly understand that the embodiments can be implemented by means of software and the necessary general hardware platform, and of course, can also be implemented by hardware. Based on such understanding, the above technical solutions can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the method described in each embodiment or some parts of the embodiment.

[0120] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting the key factor of sudden change from drought to flood based on causal inference, characterized in that, The method comprises the following steps: S1: capturing the precursors of critical phase transition of complex systems, using non-invasive multi-field coupling sensing to obtain critical state data of the drought and flood system when it transits from metastable state to critical state, and forming a critical precursor data set; S2: using critical state fingerprint matching method to determine the critical starting point of drought and flood sudden change, mark the event type, and form a critical state event list; In step S2, the step of determining the critical starting point of drought and flood sudden change comprises: S21: taking the system equilibrium state reference value as a reference, analyzing the characteristics deviating from the reference in the critical precursor data, extracting the core fingerprint characteristics including daily rainfall increase, 10cm soil moisture daily decrease, runoff-precipitation response lag time, and NDVI spatial heterogeneity, and forming a fingerprint characteristic type library; S22: based on the critical precursor data set, combining the fingerprint characteristic type library, calculating the value of the core fingerprint characteristics, and generating a quantized fingerprint characteristic matrix; S23: according to historical drought and flood sudden change events, the feature values of the quantized fingerprint characteristic matrix corresponding to the starting time point and 7 days before and after the starting time point are extracted as a critical sample set, the percentage method is used to determine the threshold value of each feature, the threshold value effectiveness is verified through ROC curve, and a critical threshold table is output; S24: checking the quantized fingerprint characteristic matrix by time and space unit, when a unit simultaneously meets the supercritical threshold value of the fingerprint characteristics, it is marked as a potential critical starting point, and then compared with the historical drought and flood sudden change events, the critical starting point list containing the starting time and spatial coordinates is output, and the event type is marked; S25: assigning a unique event ID to each potential critical starting point, associating the corresponding precursor data segment, supplementing the geographical environment information at the time of event occurrence, and generating the critical state event list; S3: based on the critical precursor data set and the critical state event list, screening the precursor factors of order parameters that can drive the system to transit from metastable state to critical state; excluding pseudo factors that only accompany critical state but have no driving effect, to obtain a candidate factor set driven by self-organized criticality; S4: in different hydrological climate systems, verify the universality of each factor in the candidate factor set in triggering system critical phase transition; screen the factors with cross-system domain universal driving property, determine as the drought and flood sudden change key factor, and form a universal key factor list; S5: based on the causal relationship between the drought and flood sudden change key factor and the system self-organized critical phase transition, a critical phase transition deduction model is constructed; based on the critical phase transition deduction model, whether the system will occur critical transition in the future 1-15 days is predicted, the order parameter contribution degree of each key factor in promoting critical phase transition is output, and a prediction report is generated; In step S5, the step of constructing the critical phase transition deduction model comprises: S51: based on the causal relationship between the key factor and the system critical phase transition, a causal inference neural network is selected as the core model, including an output layer, a hidden layer and an output layer, an order parameter contribution degree calculation module is embedded, the network structure is optimized by referring to the existing drought prediction model, and a model architecture diagram is drawn; S52: using historical critical event data to train the model, outputting the trained model, parameter configuration table and error evaluation report; In step S5, the step of outputting the order parameter contribution degree of each key factor in promoting critical phase transition comprises: S53: Based on the non-invasive multi-field coupling perception, real-time data is automatically collected every day, and the real-time data is preprocessed to output standardized real-time input data set; S54: The standardized real-time input data set is imported into the trained model to simulate the future 1-15 day system evolution track. When the critical transition probability of any day is greater than the preset probability threshold, it is marked as a high-risk transition day. At the same time, the SHAP value algorithm is used to calculate the contribution degree of each key factor order parameter, and the 1-15 day critical transition prediction result table and contribution matrix are output; S55: According to the structure of background, method, prediction result, key factor contribution analysis and risk prompt, insert prediction probability graph and contribution pie chart, supplement model error range and data source explanation, and output the prediction report according to the standard format of hydrological department.

2. The method of claim 1, wherein the method is characterized by, In step S1, the step of obtaining the critical state data of the drought and flood system from the metastable state to the sudden transition state includes: S11: For the historical drought and flood sudden transition event, the sliding window method is used to analyze the runoff mutation point, and the 30 days before the mutation point to the day of the mutation point are defined as the transition stage; S12: Based on the transition stage, the soil moisture drop is verified to output the transition stage data segment from the metastable state starting time to the sudden transition state triggering time; S13: Based on the transition stage data segment, the key indicators of each sensing domain are screened; S14: The structure is integrated according to the time sequence and spatial distribution, and the metadata of the data source and the collection equipment is supplemented to form the critical precursor data set. 3.The method of claim 1, wherein, In step S3, the step of screening the order parameter precursor factor that can drive the system from the metastable state to the critical state includes: S31: According to the self-organized critical theory of complex system, the core analysis dimension is selected combined with the critical precursor data set and the critical state event list, the dimension discrimination ability for critical state is verified through pre-experiment, the analysis dimension system is formed, and the calculation method of each dimension is labeled; S32: The potential factors are extracted from the critical precursor data set, the Pearson correlation coefficient of each potential factor and the critical event is calculated, and the initial screening precursor factor set is formed combined with the analysis dimension evaluation, and the factor name, correlation value and dimension contribution value are output.

4. The method of claim 3, wherein the method is characterized by, In step S3, the step of obtaining the candidate factor set includes: S33: Taking the initial screening precursor factor set as the object, single factor disturbance and multi-factor combined disturbance are used, the disturbance frequency is set to once a day, the monitoring duration is 15 days after disturbance, and the monitoring index is whether to trigger the critical state, and the experimental scheme including disturbance sequence, amplitude and monitoring index is formed; S34: Based on Python SimPy simulation platform, the critical precursor data set is imported, the disturbance is automatically executed according to the experimental scheme, the monitoring index value is calculated after each disturbance, whether to jump to the critical state, response delay time and response intensity are recorded, and each factor is repeated for 30 times, and finally integrated into the micro-disturbance response result set; S35: The pseudo factor judgment standard is set, the initial screening factor is checked one by one according to the micro-disturbance response result set, the factor meeting the pseudo factor standard is removed, the remaining factors are sorted according to the driving effectiveness score, and the candidate factor set driven by self-organized criticality is formed.

5. The method of claim 1, wherein the method is characterized by, In step S4, the step of verifying the universality of each factor in the candidate factor set triggering the critical phase transition of the system includes: S41: According to the classification of hydrological and climatic systems, a preset number of typical river basins are selected in each system domain, historical critical precursor data sets, historical critical state event lists and geographical parameters of each river basin are collected to form a cross-system domain verification sample set; S42: For each verification sample in the cross-system domain verification sample set, the driving effect of each factor in the candidate factor set is tested using the perturbation experiment method, the critical transition success rate after single factor perturbation and the synergistic driving rate of multi-factor combination perturbation are counted, the effectiveness result and specific success rate of each factor in the sample are recorded, the data missing sample is replaced with similar river basin data in the same system domain, and a single domain verification result table is output; S43: Define a universality evaluation index, calculate the cross-domain passing rate and coefficient of variation of each candidate factor based on the single domain verification result table, and form a factor cross-domain universality matrix.

6. The method of claim 5, wherein the method is characterized by, In step S4, the step of generating the list of universal key factors includes: S44: Set a universality screening threshold, screen factors that meet the universality screening threshold from the candidate factor set, confirm the physical reasonableness of the factors, and output a set of drought-flood sudden change key factors; S45: Integrate the set of drought-flood sudden change key factors and the previous data in a preset format, supplement the monitoring method and data source of the factors, and generate the list of universal key factors.

7. A system for predicting a key factor of abrupt transition from drought to flood based on causal inference, which employs the method for predicting a key factor of abrupt transition from drought to flood based on causal inference according to any one of claims 1 to 6, characterized in that, It includes: A multi-field coupling perception module for real-time acquisition of multi-dimensional data related to the drought-flood system; A data construction module for constructing a critical precursor data set based on the multi-dimensional data; A critical starting point determination module for locating the critical starting point of drought-flood sudden change through fingerprint matching technology to form a traceable critical state event list; A factor screening module for screening precursor factors that can drive the critical phase transition of the system from the critical precursor data set and eliminating pseudo factors to form a candidate factor set; A key factor determination module for verifying the universality of candidate factors in different hydrological and climatic systems based on the candidate factor set and determining drought-flood sudden change key factors; A model construction module for constructing a causal inference neural network model based on the causal relationship between key factors and critical phase transition, importing real-time data into the model, simulating system evolution trajectories, calculating the contribution degree of each key factor through SHAP value algorithm, and outputting a prediction result table and a contribution degree matrix; A report generation module for generating a prediction report based on the prediction result table and the contribution degree matrix.

Images