Dynamic discrimination method for post-flood season based on water regime factor stratification screening and multi-condition combination

Abstract

The application discloses a post-flood season dynamic discrimination method based on water regime factor hierarchical screening and multi-condition combination, comprising obtaining daily water regime data of a target river basin and preconfigured discrimination parameters; based on the water regime data, multi-dimensional attributes are extracted and importance grading is performed, thereby obtaining a first layer core factor set and a second layer auxiliary factor set; current monitoring values of each factor set are obtained; based on the monitoring values and the discrimination parameters, verification is performed, if the factors in the first layer set all meet core discrimination conditions, and the number of factors in the second layer set meeting auxiliary discrimination conditions reaches an optimal satisfaction number, then it is determined that the target river basin enters the post-flood season. The application constructs a hierarchical heterogeneous discrimination network, anchors the flood control safety bottom line, improves the fault tolerance of the system to complex water regimes, alleviates the technical contradiction that the traditional unified rule discrimination is prone to be early or late, and takes into account the reservoir regulation safety and the comprehensive utilization benefit of water resources at the end of the flood season.

Description

Technical Field

[0001] This invention relates to data processing technology in the field of flood control scheduling and water resources management, and in particular to a post-flood season dynamic identification method based on hydrological data mining and multi-condition combination. Background Technology

[0002] Currently, for determining the flood season and flood status in a river basin, the industry mostly adopts fixed threshold methods based on historical hydrological statistical analysis or simply relying on instantaneous hydrological indicators. In terms of feature construction, conventional methods typically extract data from hydrological stations with high correlation coefficients directly as input for discrimination, without fully assessing the independent information carrying capacity of the indicators.

[0003] In terms of the discrimination logic, the system mostly uses a globally unified Boolean rule to perform multi-factor verification: if the global logic AND is used, it requires that the water level or flow rate of all observation points be lower than the set threshold at the same time, and the state switching conditions of the system are extremely harsh; if the global logic OR is used, the switching of the system is too aggressive.

[0004] In terms of assessing water level decline, existing technologies mostly use the instantaneous drop of water level at a specific cross-section below the warning line as the sole triggering condition, lacking consideration of water level fluctuation trends.

[0005] In summary, existing discrimination methods struggle to effectively isolate redundant signals and highlight core security lines when dealing with multi-source and dense hydrological information. Furthermore, when faced with complex conditions of alternating flood peaks and overlapping local disturbances, the uniform rigid rules and transient judgment mechanisms often make it difficult to balance safety risks and scheduling benefits. The robustness of the system's discrimination conclusions and the safety of flood control remain challenging.

[0006] Therefore, there is an urgent need to study a comprehensive assessment method that can improve the adaptive capability and spatiotemporal robustness of flood season status identification while ensuring the safety of the flood control baseline. Summary of the Invention

[0007] The purpose of this invention is to provide a post-flood season dynamic discrimination method based on hierarchical screening of hydrological factors and combination of multiple conditions, in order to solve the above-mentioned problems in the prior art.

[0008] Technical solution: A post-flood season dynamic discrimination method based on hierarchical screening of hydrological factors and combination of multiple conditions, including:

[0009] Acquire daily hydrological data and pre-configured discrimination parameters for the target watershed;

[0010] Multidimensional attributes of hydrological factors are extracted based on daily hydrological data, and importance classification is performed based on the multidimensional attributes to obtain a first-level core factor set and a second-level auxiliary factor set.

[0011] Obtain the current monitoring value of each factor in the first-level core factor set and the second-level auxiliary factor set;

[0012] Within the window period during which the timing constraints are met on the current date, verification is performed based on the current monitoring values ​​and discrimination parameters. If all factors in the first-layer core factor set meet the core discrimination conditions, and the number of factors in the second-layer auxiliary factor set that meet the auxiliary discrimination conditions reaches the pre-configured optimal number of conditions met, then the target watershed is determined to have entered the post-flood season, and the dynamic discrimination result is output.

[0013] Beneficial effects:

[0014] By using a multi-attribute cascaded screening and a multi-dimensional security contribution scoring mechanism, the characteristics of hydrological stations are adaptively reduced in dimensionality and divided into a mandatory interception layer and an elastic voting layer. This helps to filter out irrelevant information disturbances and is conducive to maintaining the physical security baseline of water conservancy control sections.

[0015] By utilizing threshold margin, receding slope, and reverse rise constraint to perform product coupling, a fuzzy evaluation mechanism for receding credibility in continuous space is constructed. This improves the system's robustness in temporal identification under extreme abrupt changes.

[0016] A heterogeneous condition combination network was constructed. The underlying core factors maintain the flood control safety threshold, while the peripheral auxiliary factors cooperate with the elastic satisfaction number release space after asymmetric risk optimization. Under the premise of ensuring the safety of dam and downstream flood control, the water resources at the end of the flood season are captured in advance to the maximum extent, which is conducive to balancing the flood control scheduling safety of reservoir group and the economic benefits of water storage. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the post-flood season dynamic discrimination method based on hierarchical screening of hydrological factors and combination of multiple conditions, as proposed in this invention.

[0018] Figure 2 This is a schematic diagram of the process for extracting multidimensional attributes of hydrological factors according to the present invention.

[0019] Figure 3 This is a schematic diagram of the joint evaluation process of the present invention.

[0020] Figure 4 This is a schematic diagram of the importance classification process based on multi-dimensional attributes in this invention.

[0021] Figure 5 This is a schematic diagram of the dynamic truncation and partitioning process based on security contribution scores and pre-configured capacity constraints according to the present invention. Detailed Implementation

[0022] Example 1: This example provides a post-flood season dynamic discrimination method based on hierarchical screening of hydrological factors and combination of multiple conditions, such as... Figure 1 As shown, the method includes the following steps:

[0023] Step 101: Obtain daily hydrological data and pre-configured discrimination parameters for the target watershed;

[0024] The daily hydrological data is retrieved in real time or loaded in batches through sensor networks and hydrological station databases; it includes daily water level records and daily flow records of each preset hydrological control station within the target watershed.

[0025] After obtaining the daily hydrological data, outliers are filtered out using quality control rules, and dimensionless mapping is performed on the historical data of each control station using the range normalization method to eliminate the dimensional differences between water level and flow rate, thereby obtaining preprocessed standardized data.

[0026] The pre-configured discrimination parameters are multi-dimensional control hyperparameters that are stored in advance in the scheduling center database. Specifically, they may include adaptive discrimination thresholds, receding trend judgment windows, time-series prerequisite boundaries, and optimal satisfaction numbers used to assist network decision-making.

[0027] Step 102: Extract multidimensional attributes of hydrological factors based on daily hydrological data, and perform importance classification according to the multidimensional attributes to obtain the first layer of core factor set and the second layer of auxiliary factor set;

[0028] Specifically, multidimensional attributes refer to technical quantities that characterize hydrological features, such as temporal morphological characteristics, information carrying capacity, seasonality index, and physical security correlation. By performing cascaded dimensionality reduction calculations on normalized daily hydrological data, the information entropy index, seasonality index, and independence index obtained through maximum uncorrelation analysis of each factor are extracted.

[0029] Based on this, the safety contribution of each hydrological factor is comprehensively quantified and scored through weighted calculation or statistical ranking. Combined with the upper and lower limits of the pre-configured factor library capacity, the traditional mechanical division model within a category is broken. Characteristic factors that exhibit strong physical control attributes and are highly correlated with the target flood control section are classified into the first-level core factor set; at the same time, other representative factors that indirectly or auxiliaryly reflect the overall flood attenuation situation of the basin are classified into the second-level auxiliary factor set.

[0030] By grading the importance of spatial dimensions, the control factor matrix is ​​compressed and decoupled into two heterogeneous logical levels.

[0031] Step 103: Obtain the current monitoring value of each factor in the first-layer core factor set and the second-layer auxiliary factor set;

[0032] Specifically, after completing the spatial dimension factor hierarchy division, the flood control dispatch center system moved from the offline configuration stage to the real-time on-site online monitoring stage.

[0033] The system uses a watershed telemetry system, satellite hydrological relay nodes, and an automated hydrological station network to collect real-time hydrological feedback from control stations corresponding to each specific factor in the first-layer core factor set and the second-layer auxiliary factor set at preset time steps.

[0034] Specifically, the current monitoring value is represented by the measured water level or the measured composite flow rate of the river on the same day. Real-time acquisition of the latest dynamic feedback of each factor layer enables subsequent logical condition verification work to be based on the latest field conditions, improving the real-time response efficiency of the entire judgment process.

[0035] Step 104: Within the window period during which the timing constraints are met on the current date, verification is performed based on the current monitoring values ​​and discrimination parameters. If all factors in the first-layer core factor set meet the core discrimination conditions, and the number of factors in the second-layer auxiliary factor set that meet the auxiliary discrimination conditions reaches the pre-configured optimal number of conditions met, then the target watershed is determined to have entered the post-flood season, and the dynamic discrimination result is output. If both conditions are not met simultaneously, the current flood season status is maintained, and the verification is repeated in the next discrimination cycle.

[0036] This step utilizes a heterogeneous conditional combination discriminant network to perform real-time rolling exercises on the collected data.

[0037] During the verification process, we first check whether the current real-time date falls within the time-series constraints jointly constructed by the external end date of the plum rain season and the preset earliest judgment deadline.

[0038] After this timing condition is met, a two-layer parallel decision matrix is ​​enabled.

[0039] For the first-level core factor set, the system executes strict AND logic for all phases, requiring that the current monitoring values ​​of all core stations must be lower than the corresponding discrimination thresholds, and that the real-time receding trend confirmation results corresponding to the water level factors must all be in an effective receding state. If any core factor is still at a high level or there is a false receding signal, it will directly prevent the determination of the post-flood season, thereby anchoring the safety bottom line of flood control.

[0040] For the second-level auxiliary factor set, the system executes a flexible voting logic that satisfies K conditions. As long as the number of auxiliary conditions in the qualified state is greater than or equal to the optimal number of conditions satisfied, the second-level verification is deemed to have passed.

[0041] When both layers of verification pass simultaneously, the system automatically determines that the target basin has officially entered the post-flood season on that day, and outputs the corresponding dynamic judgment result to the control network to automatically switch the operation control mode of the flood control dispatch system.

[0042] In some alternative implementations, the backflow trend verification method within the core and auxiliary discrimination conditions can adopt a multi-alternative architecture.

[0043] For example, the dual-threshold Boolean hard judgment based on the slope of univariate linear regression and the maximum daily increase limit can be replaced by a continuous spatial receding credibility index judgment based on the product coupling of the threshold margin index, the receding slope index and the anti-increase constraint index.

[0044] By introducing continuous fuzzy control logic, the system's recognition reliability on complex flood peak abrupt boundary can be further improved.

[0045] In addition, the dynamic discrimination results can not only be used for the conclusion presentation on the large screen at the site, but also as the input of the downstream flow control operator to directly drive the dam gate to perform cascade water storage.

[0046] Example 2: Based on Example 1 above, this example further details the specific steps for extracting multi-dimensional attributes of hydrological factors. Specifically, this example discloses a cascaded screening process for multi-attribute hydrological factors in a spatial dimension, used to extract a set of preferred hydrological factors with high independence and high temporal discriminative power from massive amounts of original hydrological station network observation characteristics.

[0047] In one possible implementation, such as Figure 2 and Figure 3 As shown, it includes the following steps:

[0048] Step 201: Perform hierarchical clustering on each hydrological factor based on daily hydrological data to obtain the factor clustering results;

[0049] Specifically, the multi-year daily average sequence of each hydrological factor in the acquired daily hydrological data is used as a clustering sample, where the time span of the multi-year daily average sequence corresponds to the preset flood season window.

[0050] In this embodiment, the total number of days in the flood season window can be set to 153 days. Fully linked hierarchical clustering is used to spatially group all original hydrological factors within the target watershed according to their temporal sequence.

[0051] In the calculation process, the Euclidean distance between the multi-year daily average series of any two hydrological factors is first calculated to characterize the morphological differences between different factors.

[0052] Next, a fully linked measure is introduced to define the inter-cluster distance between different clusters. The inter-cluster distance is limited to the Euclidean distance between the furthest sample pairs in two clusters. By constructing a cluster dendrogram and combining it with the balance principle of having similar numbers of samples within each cluster, the optimal number of clusters is adaptively determined, and the factor clustering results are finally output.

[0053] The purpose of performing morphological clustering is to group similar control stations with dense spatial distribution and highly similar process line morphology into a group, thereby effectively reducing the computational overlap and redundancy of subsequent multidimensional feature analysis.

[0054] Those skilled in the art can determine the appropriate cutoff point by observing the jump points of the distance threshold in the clustering dendrogram and combining the balance of sample sizes in each category.

[0055] Before conducting the joint assessment, a preliminary validity check was performed on each hydrological factor: factors whose multi-year daily average series remained constant within the flood season window were directly identified as non-informative factors and removed from the assessment because their information entropy was always zero. During the probability discretization process of calculating information entropy, for intervals with a probability of zero, the contribution of that term to the information entropy was set to zero.

[0056] Step 202: Within each category of the factor clustering results, a joint evaluation of information content and seasonality index is performed to screen out candidate representative factors.

[0057] Specifically, joint evaluation refers to breaking away from the traditional single statistical screening criteria and conducting a dual assessment of various factors within the same cluster from both the information theory dimension and the meteorological and hydrological period identification dimension.

[0058] Parallel closed-loop verification is performed on each category in the factor clustering results. The information entropy index and seasonality index of each factor are calculated in combination to quantitatively evaluate the richness of independent information it carries and the sensitivity of time identification of the transition boundary between the main flood season and the post-flood season.

[0059] Through joint evaluation, weak feature factors that change slowly within a category and lack temporal representativeness can be eliminated, thus locking in the core candidate pool for the next step of global redundancy elimination across categories.

[0060] Step 203: Based on circular statistics, the flood season sequence is converted into directional angles to determine the seasonality index, and information entropy is calculated based on probability discretization;

[0061] Calculate the intra-class mean of the seasonality index and the intra-class mean of the information entropy within each category, and establish them as the corresponding intra-class benchmarks;

[0062] Factors whose seasonality index and information entropy both meet the corresponding intra-class benchmarks are retained as candidate representative factors; specifically, factors whose seasonality index is not lower than the mean of the corresponding intra-class seasonality index and whose information entropy is not lower than the mean of the corresponding intra-class information entropy are retained as candidate representative factors.

[0063] If there is no factor in a category whose information entropy and seasonality index are both not lower than the mean of the corresponding category, then the factor with the largest sum after normalization of the two indicators within the category is established as the candidate representative factor of the category, ensuring that each category retains at least one representative factor and avoiding the overall loss of hydrological information of that category.

[0064] Specifically, when calculating information entropy, the system first performs equal-width interval probability discretization on the multi-year daily average sequence of the factor.

[0065] To eliminate the subjective arbitrariness of manually setting the number of intervals, the classic histogram interval number determination rule is introduced, and the discretized interval number is adaptively calculated by the following formula: B=1+log2(L);

[0066] Where B is the number of discretized intervals, L is the total number of days in the flood season window, and log2 is the logarithmic function with base 2.

[0067] It should be noted that the calculated number of discretized intervals B is rounded up. When the total number of days L in the flood season window is constant at 153 days, the number of discretized intervals B determined by the formula and rounded up is 9.

[0068] The range of the multi-year daily average series of each hydrological factor is divided into 9 equally wide intervals. The proportion of the actual number of days that the series data falls into each interval is used as the corresponding probability of occurrence.

[0069] Based on this, the system calculates the information entropy using the probability discretization formula:

[0070] Hj=-1*sum(P ξ *log2(P ξ ));

[0071] Where Hj is the information entropy of the j-th hydrological factor, sum is the function that sums up all terms, and P ξ Let be the probability that the range of the multi-year daily average sequence of the j-th hydrological factor falls within the ξ-th equal-width interval, and log2 be the logarithmic function with base 2.

[0072] The higher the value of information entropy, the higher the uncertainty of the hydrological factor, and the richer the amount of physical hydrological information it carries.

[0073] Meanwhile, when calculating the seasonal index, the principle of circular statistics is introduced to convert the daily data of each level within the flood season window into unit observation vectors in a two-dimensional circular space.

[0074] For day t, its position on the time axis is converted into a circular spatial direction angle using the following formula: φ t =2*π*(t-1) / L; where, φ tLet t be the direction angle corresponding to the t-th day in the circular statistical space, π be the value of pi, t be the rank of that day within the flood season window, and L be the total number of days in the flood season window.

[0075] The multi-year daily average value of each factor, after range normalization, is taken as the length of the corresponding daily vector. Orthogonal decomposition and summation are then performed in a rectangular coordinate system. The specific operational logic is characterized by the following formulas for cosine synthesis of opposing components and sine synthesis of opposing components: C j =sum(x j_t_prime *cos(φt));

[0076] Where Cj is the cosine composite opposing component of the j-th hydrological factor, sum is the summation function, and x j_t_prime Let φt be the normalized multi-year daily average value of the j-th hydrological factor on day t, cos be the cosine function, and φt be the direction angle corresponding to day t.

[0077] S j * =sum(x j_t_prime *sin(φt)); where S j * Let x be the sinusoidal composite opposing component of the j-th hydrological factor, and sum be the summation function. j_t_prime Let φt be the normalized multi-year daily average value of the j-th hydrological factor on day t, sin is the sine function, and φt is the direction angle corresponding to day t.

[0078] The modulus of the composite vector and the total modulus of the vector lengths for each day are calculated using the following formulas:

[0079] The magnitude R of the composite vector j =sqrt(C j 2 +S j * 2 );

[0080] Where sqrt is the square root function, Cj is the cosine composite opposing component of the j-th hydrological factor, Sj* is the sine composite opposing component of the j-th hydrological factor, and Cj² and Sj*² represent the squares of the corresponding components, respectively.

[0081] M j =sum(x j_t_prime );

[0082] Among them, M j Let x be the total modulus of the vector length for each day, and sum be the summation function. j_t_prime Let be the normalized multi-year daily average value of the j-th hydrological factor on day t.

[0083] The seasonality index SI is determined by the ratio of the magnitude of the composite vector to the total modulus. j =R j / M j ;

[0084] Among them, SI j R is the seasonal index of the j-th hydrological factor. j M is the magnitude of the composite vector. j This represents the total modulus of the vector length for each day.

[0085] The seasonality index is strictly limited to a dimensionless range of 0 to 1. When the hydrological changes of a factor are highly concentrated in a certain period during the flood season, the geometric direction of the daily vectors tends to be consistent, and the calculated seasonality index approaches 1, indicating that the factor has a strong ability to identify the time period of the flood season. After calculating the indices of all factors within a category, the average information entropy and the average seasonality index within that category are calculated and established as the corresponding intra-category benchmarks. Only high-quality factors with information entropy and seasonality indices not lower than the intra-category mean are retained as candidate representative factors, and the remaining factors are eliminated as redundant features within the category.

[0086] Step 204: Perform inter-class redundancy elimination on the candidate representative factors to obtain the preferred set of hydrological factors;

[0087] Since there may still be cross-category information overlap between different clustering categories due to the synchronicity of rainstorm weather in the basin, it is necessary to perform a global cross-category maximum irrelevance exercise.

[0088] By performing inter-class redundancy resolution on all remaining candidate representative factors, it is possible to ensure that the final selected factor set has information complementarity and extremely high independence on the whole basin spatial scale, thereby narrowing the input feature space of the subsequent heterogeneous discrimination network without losing key flood control and water storage technology features.

[0089] Step 205: Calculate the multiple correlation coefficient between each candidate representative factor and the remaining candidate representative factors; remove factors whose multiple correlation coefficients exceed the pre-configured correlation threshold to obtain the preferred set of hydrological factors.

[0090] During the redundancy elimination iteration process, each candidate representative factor is examined one by one.

[0091] Using the multi-year daily average series of the target candidate representative factor as the dependent variable and the multi-year daily average series of all other remaining candidate representative factors as independent variables, a multiple linear regression model is performed to calculate the multiple correlation coefficient of the target factor.

[0092] The elimination decision is performed by comparing the following inequality constraints: R j_star >Z;

[0093] Among them, R j_star Z is the multiple correlation coefficient of the j-th candidate representative factor relative to all other candidate representative factors, and Z is the pre-configured correlation threshold.

[0094] Furthermore, the pre-configured correlation threshold Z possesses spatial adaptive adjustment characteristics: when the observation stations within the basin are densely distributed and the hydrological correlations among stations are generally high, the pre-configured correlation threshold Z is relaxed to 0.90 to prevent excessive elimination leading to insufficient representativeness; when the observation stations within the basin are sparsely distributed and the factors themselves exhibit significant differences, the pre-configured correlation threshold Z is tightened to 0.80 to enhance the interception of redundant signals. By repeatedly executing the above iterative calculation and comparison elimination process until the multiple correlation coefficients of all remaining factors in the pool do not exceed the pre-configured correlation threshold Z, the final set of remaining factors is established as the preferred set of hydrological factors.

[0095] Example 3: This example discloses a preferred implementation method for factor hierarchical allocation based on safety contribution and capacity constraints, which is used to solve the technical defects of traditional feature classification mechanism that fails to effectively integrate the flood control business attributes of specific control sections on site and that the boundary division is too mechanical, resulting in the core necessary condition set being too strict or too broad in subsequent heterogeneous discrimination.

[0096] In one possible implementation, such as Figure 4 and Figure 5 As shown, it includes the following steps:

[0097] Step 301: Obtain the flood control safety correlation between each factor in the preferred set of hydrological factors and the target flood control section;

[0098] Specifically, the flood control safety correlation reflects the intensity of hydrodynamic transmission or statistical significance of correlation between specific site factors and flood control control sections of the entire basin.

[0099] As one approach, if there is a lack of historical long-term synchronous hydrological observation sequences within the current target watershed, flood control safety correlation can be qualitatively configured based on topological location networks.

[0100] For example, if the hydrological control station corresponding to the factor is located on a key node of the main stream directly upstream of the target flood control section and directly participates in the determination of flood control control rules, its flood control safety correlation degree is pre-configured as a (e.g., 1.0); if the station is located on a general node of a tributary, its flood control safety correlation degree is configured as b (e.g., 0.5).

[0101] As one implementation method, when the system has a complete historical synchronous measured hydrological sequence, this indicator is extracted through quantitative statistical calculation, and its specific formula is as follows: A j =abs(corr(xj ,x c ));

[0102] Among them, A j Let x be the flood control safety correlation degree of the j-th preferred hydrological factor, abs be the function for calculating the absolute value, corr be the function for calculating the correlation coefficient between the two sequences, and x be the value of x. j For the multi-year daily hydrological sequence of the j-th preferred hydrological factor, x c The multi-year daily hydrological sequence of the target flood control section.

[0103] This process enables the flood control safety correlation to adaptively reflect the statistical mapping of the physical laws governing the natural regulation and storage of rivers and the evolution of flood propagation.

[0104] Step 302: Calculate the safety contribution score of each factor based on the seasonality index, independence, and flood control safety correlation.

[0105] Step 303: The safety contribution score is obtained by weighted summation of the seasonality index, the independence index representing independence, and the flood control safety correlation.

[0106] Specifically, after extracting the correlation degree of flood control safety, the multidimensional attributes of the factors in the spatial dimension should be organically integrated.

[0107] Before fusion, in order to eliminate the dimensional conflicts between different statistical indicators, the multiple correlation coefficients obtained from the previous cascade resolution are first converted to calculate the independence index.

[0108] Independence index I of the j-th preferred hydrological factor j =1-R j *; where R j * represents the multiple correlation coefficient of the j-th preferred hydrological factor relative to other candidate factors.

[0109] Based on pre-configured weighting coefficients, the seasonality index representing time discernment, the independence index representing spatial independence, and the flood control safety correlation degree representing physical business attributes are subjected to multi-dimensional linear weighting to quantify the comprehensive safety contribution score of each factor.

[0110] The safety contribution score G of the j-th preferred hydrological factor j =ω s *SI j +ω i *I j +ω a *A j ;

[0111] Where, ω s SI is a preset seasonal index weighting parameter. jLet I be the seasonal index of the j-th preferred hydrological factor. j ω is the independence index of the j-th preferred hydrological factor; i ω is the preset independence weight parameter. a As a preset security correlation weight parameter, A j Let be the flood control safety correlation degree of the j-th preferred hydrological factor.

[0112] The weight parameters satisfy the following normalization constraint: ω s +ω i +ω a =1;

[0113] Where, ω s ω is the preset seasonal index weighting parameter. i ω is the preset independence weight parameter. a These are preset security correlation weight parameters.

[0114] In this embodiment, the weight parameters are adjusted in real time according to the focus of the actual flood control operations in the basin. For example, when the system pays more attention to the spatially independent defense capability of factors, the independence weight parameter can be manually increased; when the system pays more attention to the physical coupling safety of key flood control sections under the control of dam storage and discharge, the safety correlation weight parameter can be actively increased. Through multi-dimensional weighted scoring, a dimensionality reduction mapping from a high-dimensional feature space to a one-dimensional comprehensive safety contribution scoring space is achieved.

[0115] The weight parameters can be determined by those skilled in the art based on the flood control priorities of the target watershed, using conventional multi-criteria decision-making methods such as the analytic hierarchy process or expert scoring.

[0116] Step 304: Based on the security contribution score and the pre-configured capacity constraints, perform dynamic truncation and partitioning to obtain the first-layer core factor set and the second-layer auxiliary factor set;

[0117] Step 305: Factors whose safety contribution score and flood control safety correlation both meet the corresponding grading thresholds are included in the first-level core factor set; that is, factors whose safety contribution score is not lower than the pre-configured core contribution threshold and whose flood control safety correlation is not lower than the pre-configured safety correlation threshold are included in the first-level core factor set.

[0118] If the number of factors that meet the grading threshold exceeds the capacity limit, only the factors with the highest security contribution score and whose number equals the capacity limit will be retained.

[0119] If the number of factors that meet the grading threshold is less than the lower limit of capacity, then the remaining factors that have not met the grading threshold are added in descending order of safety contribution score until the number reaches the lower limit of capacity; and the remaining factors are assigned to the second-level auxiliary factor set.

[0120] Specifically, traditional hierarchical division often mechanically divides clusters into two categories based on statistical ranking, which can easily lead to some key factors closely related to flood control safety being wrongly downgraded to the auxiliary condition layer due to their lower overall ranking within the category, or cause the base number of the necessary factors in the first layer to be too large, making the logical network of the full AND too demanding, thus missing out on considerable economic benefits of water storage.

[0121] By introducing global capacity constraints and multiple service-level thresholds, a dynamic truncation and adaptive replenishment allocation mechanism is constructed. The capacity constraint includes a pre-configured capacity limit M. max With the pre-configured lower capacity limit M min The tiered thresholds include a pre-configured core contribution threshold G0 and a pre-configured security association threshold A0.

[0122] All retained factors in the preferred hydrological factor set are globally sorted in descending order according to their calculated safety contribution scores. The entire sorted factor sequence is then traversed from beginning to end. Only factors that simultaneously satisfy the safety contribution score G are considered retained. j ≥G0 and flood control safety correlation degree A j Only those business indicators ≥A0 that are considered excellent are allowed to be candidates for the first-level core factors.

[0123] During the inclusion process, the current cumulative number of the first-level core factor set is monitored in real time. If the total number of factors that fully meet the double threshold conditions exceeds the capacity limit M, the inclusion process is initiated. max To prevent excessively long full-phase and logic conditions from causing excessively delayed subsequent decision-making, dynamic truncation is implemented to automatically cut off the ingress process, prioritizing the retention of the top M with the highest security contribution scores. max Each factor enters the first-level core factor set.

[0124] Conversely, if the total number of factors that fully satisfy the double threshold condition is less than the pre-configured capacity lower limit M... min To prevent the risk of missed flood warnings at the end of the main flood season due to an overly weak set of necessary conditions in the first layer, an adaptive replenishment operator is automatically activated. From the remaining factors that have not met the criteria, the first-layer core factor set is forcibly replenished in descending order of safety contribution score until the actual number of core factors reaches the capacity lower limit M. min .

[0125] All other preferred hydrological factors not included in the first-level core factor set are uniformly classified into the second-level auxiliary factor set. This flexible boundary global capacity control ensures that factors directly related to flood control core security are not overlooked, while also limiting the reasonable size of the first-level phase and logic network.

[0126] In one alternative implementation, the capacity limit M maxWith lower capacity limit M min It can be determined based on the total number of control stations within the basin and flood control management experience.

[0127] Example 4: This example discloses a multi-mechanism coupled implementation method for confirming the credibility of real-time flood recession trends. In actual flood evolution, an instantaneous water level below a threshold does not necessarily equate to entering a true recession phase; it may only be a temporary low water level between two flood peaks. This example constructs a dynamic verification mechanism to filter false recession signals by introducing a time-dimensional sequence determination.

[0128] In one possible implementation, the following steps are included:

[0129] Step 401: For water situation factors that are water level-type factors, determine the real-time receding trend confirmation results of each factor, which can be achieved through any of the following methods:

[0130] Before verification based on current monitoring values ​​and discrimination parameters, the process also includes: for water level-type factors in the first-layer core factor set and the second-layer auxiliary factor set, determining the real-time receding trend confirmation results of each factor based on a pre-configured historical monitoring window, which can be achieved through any of the following methods:

[0131] Based on the rate of change of water level and the maximum daily fluctuation amplitude within the historical monitoring window, Boolean judgment is performed to obtain the real-time recession trend confirmation result; alternatively, based on multi-dimensional hydrological control indicators, continuous spatial mapping and coupling are performed to calculate the recession credibility, and when the recession credibility reaches the pre-configured credibility standard, the real-time recession trend confirmation result is marked as a valid recession state. The pre-configured credibility standard can be determined according to the flood control safety requirements of the actual watershed by evaluating the balance point between the missed detection rate and the false detection rate under different credibility standards on the historical validation set.

[0132] Two alternative mechanisms for receding water identification are provided, which are mutually evolving.

[0133] As a basic implementation method, a pre-configured historical monitoring window is used to extract water level sequences of water level-related factors. The preferred duration of the historical monitoring window is 5 days. The system uses a univariate linear least squares fitting method to perform regression fitting on the extracted water level sequences, and calculates the linear regression slope as the rate of water level change. Simultaneously, the maximum daily fluctuation increase within the window period is calculated. A Boolean value of "true" is output only if the rate of water level change is less than 0 and the maximum daily fluctuation is strictly less than the preset maximum allowable daily increase threshold, indicating a valid receding trend.

[0134] As an improvement, since hard Boolean logic is prone to state oscillations at threshold boundaries, this invention preferably employs a method based on continuous space mapping and coupling. By calculating the discharge reliability, which characterizes the degree of discharge reliability, the determination of hard conditions is transformed into a continuous evaluation in fuzzy space. Only when the calculated discharge reliability is greater than or equal to the pre-configured reliability standard does the system confirm and output a command indicating a valid discharge state.

[0135] Step 402, when obtaining the real-time water level trend confirmation result by calculating the water level receding confidence, specifically includes: calculating the threshold margin index, which characterizes the sufficiency of the water level being below the threshold, based on the current monitoring value of the water level factor and the corresponding discrimination threshold; calculating the water level receding slope index based on the regression state of the monitoring sequence within the historical monitoring window; calculating the anti-rise constraint index based on the maximum single-day rise within the historical monitoring window; and multiplying and coupling the threshold margin index, the water level receding slope index, and the anti-rise constraint index to obtain the water level receding confidence.

[0136] To accurately identify high-risk boundary situations where the water level is just below the threshold but the receding trend is unstable, the threshold margin index is first calculated. This index represents the physical depth of the current monitored water level deviating downward from the judgment threshold.

[0137] The threshold margin index is calculated using the following formula: D j,t =max(0,min(1,(θ j -h j,t ) / Δh j ));

[0138] Among them, D j,t Let θ be the threshold margin index of the j-th water level factor on day t, where max is the function to maximize the value and min is the function to minimize the value. j h is the discrimination threshold for the j-th water level factor. j,t Let Δh be the current monitoring value of the j-th water level factor on day t. j is the threshold buffer scale for the j-th water level factor.

[0139] The system evaluates the rate characteristics of water receding. It extracts water level sequences within a preset historical monitoring window, calculates the corresponding linear regression slope, and compares it with a benchmark receding slope scale representing the standard receding rate.

[0140] The following formula is used to calculate the receding slope index: W j,t =max(0,min(1,-b j,t / b 0,j ));

[0141] Among them, W j,tLet b be the recession slope index of the j-th water level factor on day t, where max is the function to maximize the value and min is the function to minimize the value. j,t Let b be the recession slope of the j-th water level factor within the historical monitoring window corresponding to day t. 0,j is the benchmark receding slope scale for the j-th water level type factor.

[0142] Furthermore, the system checks whether there are any precursors to a secondary flood peak within the monitoring window. It searches and calculates the maximum daily flood level within the window and introduces a maximum permissible daily flood level threshold for dimensionless mapping.

[0143] The physical meaning of the maximum permissible daily rise threshold is that it allows for slight intraday fluctuations caused by rainfall or the inflow of local tributaries, but does not allow for significant rises in water levels. The anti-rise constraint index is calculated using the following formula: F j,t =max(0,min(1,1-u j,t / δ j ));

[0144] Among them, F j,t Let be the anti-rise constraint index for the j-th water level factor on day t, where max is the function to maximize the value and min is the function to minimize the value. j,t δ represents the maximum daily water level rise of the j-th water level factor within the historical monitoring window corresponding to day t. j This represents the maximum permissible daily increase threshold for the j-th water level factor.

[0145] After obtaining the three local indicators mentioned above, they are substituted into the calculation logic for the reliability of flood discharge. Since any of the following situations—insufficient threshold margin, slow decline of the flood discharge slope, or sudden and significant rise in water level—will directly undermine the physical reliability of flood control and flood discharge, a multiplication operator is selected to perform multi-indicator coupling.

[0146] Refund Credibility CR j,t =D j,t *W j,t *F j,t ;in,

[0147] CR j,t Let D be the receding water confidence level of the j-th water level factor on day t. j,t W is the threshold margin indicator. j,t F is an index of the slope of the receding water level. j,t This is an indicator for resisting price increases.

[0148] As the actual rise approaches the permissible critical limit, the anti-rise constraint indicator decreases monotonically. Once an abnormal rise occurs that exceeds the maximum permissible daily rise threshold, the anti-rise constraint indicator is truncated by the minimum function and forcibly assigned a value of 0. This triggers a global chain reaction by utilizing the product coupling characteristic, causing the final receding water level to drop directly to zero, thereby executing the interception of the water safety bottom line.

[0149] In this embodiment, the threshold buffer scale Δh j , benchmark receding slope scale b 0,j And the maximum permissible daily price increase threshold δ j These are all engineering-scale configuration parameters.

[0150] According to one aspect of this application, the pre-configured credibility standard and the maximum allowable daily increase threshold to be invoked for calculating the credibility of the water level drop are determined through a two-stage offline calibration step driven by negative samples, and this calibration is performed after the discrimination thresholds of each factor are determined; specifically including:

[0151] From the false negative samples of water receding from historical years, select the phased decline segments within the segment where at least some local monitoring values ​​are lower than the discrimination threshold as valid negative samples;

[0152] The empirical cumulative distribution of daily increase values ​​within the historical real receding positive sample segment and the empirical cumulative distribution of daily increase values ​​within the monitoring window before the second reversal and rise of the effective negative sample are respectively statistically analyzed. The daily change value corresponding to the position where the difference between the two empirical distributions reaches its maximum is selected and established as the maximum allowable daily increase threshold.

[0153] Using the historical real positive samples and valid negative samples, the receding reliability sequence of various water conditions factors is calculated day by day;

[0154] Construct an asymmetric weighted loss assessment function that includes false confirmation rate and late confirmation rate, wherein the weight coefficient assigned to false confirmation rate is strictly greater than the weight coefficient assigned to late confirmation rate;

[0155] Within the candidate confidence parameter range, the value that minimizes the result of the asymmetric weighted loss evaluation function is selected and established as the pre-configured confidence standard.

[0156] Pre-configured credibility standard C0 and maximum allowable daily increase threshold δ j The following offline calibration steps were used to determine the discrimination threshold θ, and this calibration was completed in Example 7. j Execution is performed after calibration; the threshold margin metric depends on θ. j Two-level calibration with unidirectional dependency and no circularity:

[0157] The first step is effective negative sample screening. Only retain segments that satisfy h at least once within a given time period. j,t <θj The false water level drop segments are used as valid negative samples. False water level drop segments where the water level is consistently above the discrimination threshold have been naturally blocked by the threshold criterion and provide no information for credibility calibration; C0 targets high-risk segments where the water level has fallen below the threshold but is about to rise again. If the number of valid negative sample segments is less than the preset minimum calibration segment number (e.g., 5 segments), it indicates that the discrimination threshold has blocked the vast majority of false water level drop segments. In this case, C0 is configured to a preset lower limit value (e.g., 0.3), and only the late confirmation rate from the third step is used for verification.

[0158] Step 2, δ j Separation point calibration. Empirical distribution F of daily rise within the actual receding water section is statistically analyzed. j true (δ) and the empirical distribution of the daily increase within the window before the effective false receding section rises again F j fake (δ), take the point with the largest separation between the two distributions:

[0159]

[0160] When multiple maxima exist, take the smaller one.

[0161] The third step is to optimize the asymmetric cost of C0. For the negative sample library N_f segment and the positive sample library N_r segment, the confidence sequence C of the water level drop is calculated daily using the same formula as online calculation. j,t ret For the candidate criterion set C, for example, with a step size of 0.05 from 0.30 to 0.80, calculate the false confirmation rate and the late confirmation rate:

[0162]

[0163]

[0164] Where t s For the actual start and end date of the s-th actual receding section, d c To confirm the timeframe (which can be consistent with historical monitoring windows, using 5 days), an asymmetric weighted loss is used.

[0165]

[0166] False confirmation (confirming false water loss as valid) corresponds to flood control risk, while delayed confirmation only corresponds to loss of water storage benefits. Therefore, λ_f is strictly greater than λ_m (e.g., λ_f=2.0, λ_m=1.0), which is consistent with the asymmetric risk concept of early penalty being strictly greater than late penalty in Example 6; when multiple C0 losses are the same, the larger one is taken.

[0167] Optionally, b 0,j With δ jWater level segmentation configuration: High, medium and low water level zones are divided with warning water level and discrimination threshold as the breakpoints. The above method is used to statistically analyze the actual water receding section samples in each zone to adapt to the physical difference of slow water receding in high water level zones and fast water receding in low water level zones.

[0168] Example 5 provides a detailed description of the heterogeneous condition combination and online timing prerequisite constraint collaborative control link of the post-flood season dynamic discrimination network integration layer.

[0169] Specifically, this embodiment integrates the spatial dimension hierarchical factors and the time dimension receding water identification logic in the real-time operation layer to vertically synthesize multiple physical factors and construct the overall boundary for judgment execution.

[0170] In one possible implementation, the following steps are included:

[0171] Step 501: Before performing the verification, check whether the current date meets the timing prerequisite constraints; the timing prerequisite constraints include the end date condition and the earliest allowed judgment date condition; when verifying whether the timing prerequisite constraints are met, specifically, it includes: determining the end date of the plum rain season issued by the external meteorological department as the end date of the plum rain season, and determining the period after the end date of the plum rain season and no earlier than the preset earliest allowed judgment date as the window period for allowing the dynamic judgment of the post-flood season; the verification is performed only when the current date is within the window period.

[0172] Before starting the spatial-level threshold comparison network, enforce global time-limited locking verification.

[0173] The end of the plum rain season, as an external meteorological input signal, marks the end of a large-scale, continuous period of heavy rainfall. While the plum rain season is still ongoing, the basin remains in a period of high probability of heavy precipitation, and the system disables all bottom-level control logic. Simultaneously, the earliest permissible discrimination date is set based on long-term climate statistical patterns. Taking specific basin statistical characteristics as an example, the earliest permissible discrimination date is typically set to August 1st.

[0174] Due to the high density of statistical peak values ​​during the main flood season, flood peaks occurring before August 1st still have a high probability of being surpassed later.

[0175] In practical engineering deployments, an adaptive fine-tuning configuration interface is provided, allowing flood control personnel to dynamically update the earliest permissible judgment date within the time interval of July 15th to August 15th, based on the year's abnormal weather conditions. This time-truncation mechanism prevents erroneous flood season assessments from being initiated during the peak flood season.

[0176] Step 502: The pre-configured discrimination parameters include discrimination thresholds for each factor. For factors exhibiting flood season peak characteristics, the discrimination thresholds include time-period thresholds. The time-period thresholds are determined based on the peak dates in historical hydrological statistics.

[0177] Obtain the current identification date. If the current identification date is earlier than the peak date of historical hydrological statistics, use the lower first threshold (i.e., a relatively conservative low level threshold).

[0178] When the current determination date is equal to or later than the historical peak date of hydrological statistics, a second threshold with a higher value (i.e., a relatively lenient high-level threshold) is used. "Lower" and "higher" are relative terms, meaning the second threshold is greater than the first threshold.

[0179] Specifically, to balance the flood control safety of flood-control reservoirs with the benefits of water storage and power generation, the system is equipped with the ability to adaptively switch the time dimension of the discrimination threshold. The system calls upon the historical hydrological database to extract the frequency distribution of the annual peak dates for the target factor, setting the date with the highest probability density as the historical hydrological statistical peak date. During real-time discrimination, the system extracts the current calendar time and performs segmented assignment using conditional constraints. The execution logic is as follows:

[0180] When the current determination date is earlier than the peak date in historical hydrological statistics, the first logic is applied: θ j,t =θ j_pre ;

[0181] Where, θ j,t Let θ be the threshold for determining the effectiveness of the j-th hydrological factor on the current day t. j_pre This is the pre-configured low-level peak day threshold.

[0182] When the current determination date is equal to or later than the peak date of historical hydrological statistics, the second logic is applied: θ j,t =θ j_post ;

[0183] Where θj,t is the threshold for determining the effectiveness of the j-th hydrological factor on the current day t, θ j_post The pre-configured high-level peak day threshold.

[0184] By adaptively switching between time periods, the system maintains high vigilance in the early stages when the flood risk has not significantly decreased, and automatically releases the threshold control range after the risk enters the low-frequency range, thereby improving the efficiency of comprehensive water resource utilization.

[0185] Step 503: The hydrological factors in the first-layer core factor set and the second-layer auxiliary factor set include water level factors and flow rate factors;

[0186] Before performing the verification, for water level factors, the real-time receding trend confirmation results of each water level factor are determined based on the monitoring sequence within the pre-configured historical monitoring window;

[0187] For water level factors, the corresponding core and auxiliary discrimination conditions are: the current monitoring value is lower than the corresponding discrimination threshold, and the corresponding real-time receding trend confirmation result is in a valid confirmation state; for flow rate factors, the corresponding core and auxiliary discrimination conditions are: the synthetic flow rate generated based on the current monitoring value is lower than the corresponding discrimination threshold.

[0188] Hydrological factors are classified into water level patterns and flow patterns based on their different physical observation objects. For example, parallel decision-making protocols can be loaded based on differences in their physical properties.

[0189] For water level-related factors, due to the interference of multiple dynamic factors such as river channel storage and downstream backwater, the system forcibly superimposes threshold verification and state verification. Only when the water level drops below the control line, and the multi-dimensional credibility verification described in the aforementioned embodiments confirms that the water level is indeed in the recession channel, can the condition for this factor be determined to be met.

[0190] Conversely, flow-type factors directly reflect the upstream basin's water inflow process and total runoff, and are less affected by transient disturbances from local river topography. Their instantaneous values ​​are already highly representative, so the system only performs simple threshold limit verification on their synthetic flow data, abandoning the branch for verifying the receding water trend, thus improving computational efficiency.

[0191] Step 504, the process of generating the synthetic flow includes: acquiring the flow scale data of each flow-type factor participating in the synthesis, and performing a weighted summation of the flow scale data of each factor based on the pre-configured synthesis weights to obtain the synthetic flow; wherein, the flow scale data is limited to measured flow data or normalized flow data; when performing the comparison between the synthetic flow and the corresponding discrimination threshold, when the corresponding discrimination threshold is generated based on historical measured flow data, the measured flow data is forcibly used for each flow scale data; when the corresponding discrimination threshold is generated based on historical normalized flow data, the normalized flow data is forcibly used for each flow scale data.

[0192] Furthermore, when performing integrated determination of flow-type factors, to avoid local random interference from single-station flow observations, the system performs multi-source fusion on flow data from multiple stations above a specific flood control section. The system reads the weight percentage of each flow station, i.e., the composite flow Q on day t. _syn_t =sum(μ _j *Q _j,t );

[0193] Among them, Q _syn_t Let μ be the composite flow rate on day t, and let sum be the summation function. _j Q is the composite weight of the j-th flow factor. _j,t Here is the flow-scale data for the j-th flow factor on day t.

[0194] Furthermore, to ensure absolute security and isolation of the underlying data stream, a robust physical dimension error-proofing mechanism is deployed within the computing engine. Since the preceding data base contains both measured sequences carrying raw hydrological units and normalized sequences in the zero-to-one range, a hard-binding attribute checker was designed.

[0195] Before initiating the numerical comparison operation, the system extracts metadata tags for the threshold. If the tag indicates that the generation benchmark is a physically measured value, the system intercepts all normalization input requests entering the synthesis formula and forces the data bus to provide only measured flow data with accompanying flow units. Through this dimensional error prevention and forced mapping interception, the risk of type penetration and state misjudgment that may occur under multi-dimensional data concurrent reading can be reduced.

[0196] Example 6: In one possible implementation, the pre-configured optimal satisfaction number is pre-constructed through the following offline optimization steps:

[0197] Step 601: Divide the historical year data into a training year set and a validation year set;

[0198] Before performing offline hyperparameter optimization, time-series segmentation or random sampling is performed on the long-sequence historical year data extracted from the hydrological database.

[0199] To ensure that the optimized parameters have robust generalization capabilities when facing unknown future water condition variations, the system acquires a complete flood season observation sequence over a preset time span and divides it into two mutually exclusive subsets according to a ratio.

[0200] The training year set is used to expose historical hydrological characteristics to fit the model parameters, while the validation year set is used to perform independent cross-validation for the early risk of the fitted output candidate parameters. This separation mechanism eliminates the risk of overfitting caused by using the entire historical data for training.

[0201] Step 602: In the training year set, extract the dates with no flood risk defined by historical measured water data, and calculate the asymmetric risk loss and discrimination stability index under each candidate satisfaction number. Combine them to construct a comprehensive optimization target to screen out the target candidate satisfaction number.

[0202] In this step, the first step is to establish a real-world label reference for offline supervised learning. Historical measured hydrological data from the training year set are scanned year by year, and the earliest calendar time when all control stations simultaneously recede below their respective warning water levels, and no further instances of water levels exceeding the warning level occur in subsequent periods, is defined as the flood-free date for that year.

[0203] This date serves as an objective physical benchmark for ensuring the absolute safety of water storage scheduling for that year. Subsequently, the system extracts the total number of factors from the second-layer auxiliary factor set as the upper limit of the traversal space, and sets an integer starting from 1 and incrementing as the current number of candidate satisfactions. Using this number of candidate satisfactions, the factor sequence of historical years is substituted into the two-layer heterogeneous discriminant network of Example 1 to calculate the date of entering the post-flood season as determined by the model year by year.

[0204] Step 603: Calculate the deviation of the judgment date corresponding to the number of candidate satisfactions being earlier or later than the date without flood risk.

[0205] Specifically, for each candidate satisfaction number traversed, the difference between the judgment date triggered in each year in the training year set and the corresponding date with no flood risk is calculated.

[0206] When the determination date is earlier than the date without flood risk in the calendar sequence, the system calculates the early deviation representing the early relaxation of scheduling; when the determination date is later than the date without flood risk in the calendar sequence, the system calculates the late deviation representing the delayed relaxation of scheduling.

[0207] Step 604: Apply an early penalty coefficient to the deviation representing early behavior and an late penalty coefficient to the deviation representing late behavior, wherein the early penalty coefficient is strictly greater than the late penalty coefficient.

[0208] When calculating the cost of systematic errors, an asymmetric penalty mechanism from operations research is introduced. Early decision-making means lifting safety controls when there is still flood risk in the basin, resulting in extremely high safety costs; while delayed decision-making only causes flood control reservoirs to miss some of the power generation or water resource allocation benefits from end-of-flood storage, without involving systemic physical dam failure risks.

[0209] Therefore, asymmetric risk loss can be calculated using the following formula:

[0210] J _K =sum(λ _e *max(T _y_0 -T _y_K,0 )+λ_l*max(T _y_K -T _y_0 ,0));

[0211] Among them, J _K For asymmetric risk loss, sum is the summation function over the training year set, and λ _e The early penalty coefficient is T, and max is the function that maximizes the value of T. _y_0 For the date with no flood risk in year y, T _y_K Let λ be the discrimination date in year y when the number of candidates satisfies K. _l This represents the penalty coefficient for partial delay.

[0212] When configuring the penalty matrix, the underlying constraint that the early penalty coefficient is strictly greater than the late penalty coefficient is enforced. In a typical application example, the system reads the configuration file, assigns a value of 2.0 to the early penalty coefficient, and assigns a value of 1.0 to the late penalty coefficient. This weighting bias causes the optimization algorithm to converge towards the safer side during the global search.

[0213] Step 605: The stability index is characterized based on the variance characteristics of the discrimination date deviation for each year in the training year set.

[0214] To avoid extreme fluctuations in parameter performance under different climatic years, stability indicators are calculated and evaluated simultaneously.

[0215] The error fluctuation of the same candidate satisfaction number is extracted from the training year set for each year and calculated as follows:

[0216] Discriminating stability index V _K =(1 / N _tr )*sum((ΔT _y_K -ΔT _mean_K )^2);

[0217] Where, N _tr Let T be the total number of samples in the training year set, SUM be the summation function, and ΔT be the summation value. _y_K Let ΔT be the deviation in year y. _mean_K This represents the average deviation across the set of training years.

[0218] The asymmetric risk loss is added to the stability index to obtain the comprehensive optimization objective. The fusion logic is as follows:

[0219] Q _K =J _K +η*V _K ;

[0220] Among them, Q _K To comprehensively optimize the objectives, J _K For asymmetric risk loss, η is a pre-configured stability penalty weight, which is also used for absorbing dimensions, V _K To determine stability indicators.

[0221] In an alternative embodiment, σk = sqrt(V) can also be used. _K (Use L* instead) _K =(1 / N)J _K .

[0222] By comparing the comprehensive optimization target value calculated from all candidate satisfaction numbers, a specific candidate satisfaction number that minimizes this value is selected and established as the target candidate satisfaction number. If multiple candidate satisfaction numbers result in the same comprehensive optimization target value, the one with the larger value is extracted first to ensure that the number of auxiliary condition constraints ultimately put into operation is maximized, thereby enhancing the flood control safety margin.

[0223] The stability penalty weight η is used to balance the relative importance between asymmetric risk loss and the stability of the judgment. Those skilled in the art can determine its specific value through cross-validation on the training year set, so that the comprehensive optimization objective presents a clear optimal point on multiple candidate values.

[0224] Step 606: Apply the target candidate satisfaction number to the set of verification years to perform early risk test and calculate the verification early rate. When the verification early rate is not higher than the pre-configured early rate upper limit, the target candidate satisfaction number is established as the optimal satisfaction number. When the verification early rate is higher than the early rate upper limit, the target candidate satisfaction number is incremented and the early risk test is repeated until the verification early rate meets the constraint of the early rate upper limit.

[0225] If the number of candidates that meet the requirements increases to the total number of auxiliary factors in the second layer, it is still not possible to make the early verification rate meet the upper limit constraint. The system outputs a parameter alarm and executes at least one of the conservative backoff strategies: uniformly switching the discrimination threshold of each factor to the low level threshold before the peak date, or temporarily upgrading the auxiliary layer decision to full phase AND logic, and at the same time transferring to the manual review process.

[0226] Specifically, after obtaining the target candidate satisfaction number, it is not directly approved, but instead, it enters a separate set of verification years for one-way stress testing. The absolute number of years in the verification year set where using the target candidate satisfaction number leads to an earlier judgment date is counted.

[0227] Calculate the ratio: Verify the premature rate re=N _e / N _va ; where N _e To verify the total number of years in the year set where the discrimination result occurs too early, N _va To verify the total number of samples in the year set.

[0228] Obtain the pre-configured upper limit for the premature rate. When the calculated premature rate is not higher than this upper limit, it is confirmed that the parameter has a good generalization boundary, and it is locked and updated in the control database as the optimal satisfaction number required for Implementation Example 1.

[0229] When the premature verification rate exceeds the upper limit, an automatic correction loop is activated, incrementing the target candidate satisfaction number by 1, and the validation test program is restarted until the premature verification rate meets the safety limit or the candidate parameters reach the upper limit of the total number of second-layer auxiliary factors. This dual offline closed-loop system based on training loss and validation interception eliminates the reliance on subjective experience in configuring on-site operating parameters.

[0230] Example 7: This example provides a construction and implementation method for classifying historical flood and recession patterns and adaptively calibrating thresholds.

[0231] For example, one possible implementation includes the following steps:

[0232] Step 701: Extract the historical process lines of each hydrological control station in the target basin year by year, and divide the historical years into years exceeding the warning level and years not exceeding the warning level according to whether the annual highest water level exceeds the warning level.

[0233] Specifically, by connecting to the historical hydrological database, water level and flow sequences of all control stations within the target basin during the annual flood season are extracted in batches to reconstruct historical hydrographs. Based on this, the statutory flood control warning water level parameters corresponding to each control station are obtained and used as the primary physical benchmark for classifying the time series hydrographs.

[0234] By iterating through the extreme value data for each year, years in which the highest annual water level exceeded the flood control warning level were grouped into a subset of years exceeding the warning level, and years in which the highest annual water level did not reach the flood control warning level were grouped into a subset of years not exceeding the warning level. Through this dimensional division, the sample set that actually experienced flood control pressure can be selectively extracted during subsequent feature extraction.

[0235] Step 702: Within the years exceeding the warning level and the years not exceeding the warning level, based on the temporal relationship between the rise and fall of the water level, further divide them into the rising-receding pattern after the flood peak has passed, and the receding-rising pattern in which the water level falls for a preset number of days before reaching the annual maximum value and then rises again, and extract the corresponding historical start and receding value samples accordingly.

[0236] Specifically, a second-level refinement classification is performed on the water level evolution time sequence patterns within the two subsets mentioned above.

[0237] The inflection point of the water level sequence is identified using a first-order difference algorithm. When the water level rises continuously to the absolute highest value of the year and then enters a monotonic or fluctuating downward channel, and no water level exceeding this highest value is recorded in subsequent periods until the end of the flood season, that year is marked as a rising-falling pattern. Physically, this indicates that the main flood peak has safely passed, and the flood control pressure in the subsequent basin has entered a stage of substantial reduction, which is a favorable signal combination for judging the entry into the post-flood season.

[0238] On the other hand, when the system identifies that the water level has experienced a phased decline in the time series before reaching the highest extreme date of the year, and that the decline process meets the strict constraint of a continuous decrease for a preset number of days, and then the water level reverses and rises again to surpass the high point of the previous decline, the year is marked as a receding-rising water pattern.

[0239] Here, the preset number of days is determined based on the statistical patterns of typical flood peak receding durations in the middle reaches of the Yangtze River. In some scenarios, a preferred configuration value of 10 days is selected, while modification permissions of 5 to 15 days are allowed for different tributary basins to make hydrodynamic adaptive adjustments. The successful extraction of the receding-rising flood pattern confirms that the basin may face the threat of multiple superimposed flood peaks, which has key defensive value for preventing flood risks caused by false receding floodwaters.

[0240] According to one aspect of this application, differential sample extraction and dual-sample library construction are performed based on the pattern classification results. For flood-receding years, the moment when the derivative reverses after the annual highest water level extreme point is taken as the true start and end time of receding. The water level sequence segment from this moment until the end of the flood season window and before the water level exceeds the warning level again is entered into the true receding positive sample library.

[0241] For years with receding and rising water levels, in addition to extracting the actual receding segment after the annual maximum value according to the aforementioned rules, the phased decline segment before the annual maximum value that meets the constraint of continuously decreasing preset number of days is marked as a false receding segment and entered into the false receding negative sample library. At the same time, when this type of year is used to extract the starting and receding value samples in step 703, only the actual starting and receding time after the annual maximum value is used to exclude the contamination of the threshold calibration sample by the starting and receding points of the false receding segment.

[0242] According to one aspect of this application, the pre-configured discrimination parameters include discrimination thresholds for each factor, which are obtained through pattern segmentation and offline calibration of historical year data, specifically including:

[0243] Extract the daily hydrological sequence of the target watershed for historical years. When the monitored value of hydrological factors in a certain historical year continues to rise to the highest extreme value of the year and then enters a monotonic or fluctuating decline channel, and no limit-breaking records occur again until the end of the flood season, the historical year is classified as a rising-falling water pattern.

[0244] The moment when the derivative reverses and begins to decline after the highest extreme point of the year under this model is established as the true start and end time, and the corresponding true positive samples of water receding are entered in the corresponding historical start and end value interval.

[0245] When the monitored value of water factors in a certain historical year experiences a phased decline that meets the constraint of continuously decreasing days before reaching the highest extreme value of the year, and then the monitored value reverses upward again and surpasses the previous decline starting point, the historical year is classified as a receding-rising water type.

[0246] Under this model, the phased decline segment that meets the continuous decline constraint before the highest extreme value is extracted and marked as a false negative sample of water receding;

[0247] When extracting the starting and ending samples for threshold calibration, the starting and ending data points of the time period to which the false negative water receding samples belong are removed in time sequence, and only the actual starting and ending time data after the highest extreme value day of the year are used.

[0248] Step 703: Obtain historical starting and receding value samples of each hydrological factor at the moment when the receding water begins after the historical highest flood level, and divide the sample distribution into multiple starting and receding value intervals.

[0249] Feature vectors are extracted based on the above pattern classification results. The exact time points at which each hydrological factor appears at the historical highest water level extreme point and the derivative reverses, initiating the water recession, are located. The corresponding absolute water level or flow rate values ​​at that moment are captured and defined as historical start and recession value samples.

[0250] After completing a global traversal of historical data, a raw one-dimensional sample space containing discrete hydrological values ​​is obtained. In order to map the discrete values ​​to frequency boundaries that can be statistically analyzed, the system uses an equal-width partitioning algorithm or a quantile statistical algorithm to divide the range of values ​​in this one-dimensional sample space into several consecutive candidate start and end value intervals.

[0251] Step 704: When the number of historical year samples in any set-off interval is less than the preset minimum number of samples, perform an adaptive merging operation with the adjacent set-off interval until the number of samples in each independently running set-off interval meets the minimum number of samples constraint.

[0252] Specifically, due to the scarcity of historical extreme flood years, the starting and receding value ranges in some high-water-level sections may only fall into a very small number of year samples.

[0253] When the starting and ending value interval with insufficient samples is the highest value interval and there is no adjacent next interval, it is merged with the previous interval that is adjacent to the value. When there are multiple starting and ending value intervals that meet the safety level constraints, the largest continuous segment that continuously meets the safety constraints from the lowest value end is selected, and the upper boundary of the continuous segment is established as the discrimination threshold to prevent isolated 'pseudo-safe' intervals at the high value end from raising the threshold due to sample merging.

[0254] Directly utilizing such sparse data for probability statistics can easily lead to statistical frequency distortion due to the failure of the law of large numbers. Therefore, the system extracts the minimum sample size preset in the configuration table as a baseline defense constraint. The number of historical years falling within the r-th starting and ending value interval is calculated and compared with the minimum sample size.

[0255] When it is determined that the number of historical years is less than the minimum number of samples, the system activates the interval merging engine and forces the structural merging of the starting and ending value interval with insufficient samples with the next starting and ending value interval adjacent to its value, thereby generating a new widened starting and ending value interval.

[0256] The number of historical years aggregated within the expanded start and end value intervals is recounted, and the aforementioned comparison and merging logic is continuously executed in a loop until the sample size of all existing independent start and end value intervals in the entire domain is greater than or equal to the minimum sample size.

[0257] Step 705: Calculate the subsequent alarm frequency of each start-stop value interval after the merging operation, and select the interval boundary value of the start-stop value interval where the subsequent alarm frequency is not higher than the preset safety level, and establish it as the corresponding discrimination threshold.

[0258] Furthermore, after obtaining statistically significant effective intervals, the system performs frequency conversion calibration operations. For specific hydrological factors within specific start and end value intervals, the system extracts the number of years in which subsequent periods within that interval show water levels exceeding flood control warning levels again.

[0259] p _j_r =m _j_r / n _j_r ; where p _j_r For the j-th hydrological factor, calculate the subsequent over-alert frequency within the r-th starting and retreating value interval, m _j_r Let n be the number of years in which the warning level will exceed the warning level again within the interval. _j_r This represents the number of historical years included in the aggregation within the interval.

[0260] After calculating the frequency distribution of all valid intervals, a preset safety level parameter is introduced to perform hard truncation screening.

[0261] Taking a specific engineering requirement as an example, the preset safety level parameter is configured as 5%, or a tolerance of 0.05. The system iterates through and extracts all safety start-up and de-starting value ranges that satisfy the core inequality condition of subsequent over-alarm frequency ≤ 0.05. The upper boundary value of this safety range is extracted, converted from the characteristic parameters of the data processing space, and output as the physical calibration threshold directly called by the field operation system. Through this offline calibration process, an objective calibration procedure is established that strictly back-calculates the adaptive threshold from the distribution of historical measured data and the safety tolerance frequency.

[0262] Example 8: This example further illustrates the industrial application of switching the control mode of the flood control dispatch system and dynamically adjusting the discharge flow of the reservoir group after outputting the dynamic discrimination result.

[0263] In one possible implementation, after determining that the target watershed has entered the post-flood season and outputting the dynamic discrimination result, the following steps are included:

[0264] Step 801: Based on the dynamic discrimination results, the flood control dispatch system in the target watershed is switched from flood control mode to water storage benefit optimization mode. Based on the difference space between the current monitoring value and the discrimination threshold of each factor, the outflow of the target reservoir group is dynamically adjusted to raise the reservoir water level in stages to carry out cascade water storage.

[0265] Specifically, the dynamic discrimination result is represented by a Boolean-type conclusion signal containing true or false values, output by the core control center. Flood control mode refers to the on-site operational state of the reservoir group scheduling system with the single constraint of ensuring the flood control safety of the dam and downstream areas. In this state, the reservoir group maintains a preset limit water level, freeing up a large amount of storage capacity to intercept potentially sudden main flood season peaks. Water storage benefit optimization mode refers to the on-site operational state of the reservoir group with the comprehensive constraint of improving the overall water resource utilization rate at the end of the flood season, increasing the power generation head, ensuring winter and spring water supply, and ecological backfilling.

[0266] Within the software architecture of the control system, a hard-triggered listener is deployed on the control bus to continuously monitor the level transitions of the dynamic judgment results. When a truth signal indicating the start of the post-flood season is received from the heterogeneous judgment network of Implementation Example 1, a high-priority control switching interruption protection is automatically triggered, shutting down the original flood control scheduling algorithm operator and batch loading the multi-objective optimization control operator under the water storage benefit optimization mode.

[0267] After completing the mode switch, the current monitoring value of each factor and the discrimination threshold dynamically determined in Example 5 are extracted and sent to the difference operation unit.

[0268] The difference space represents the safe net depth of the current actual water level from the allowable extreme value boundary at the end of the flood season. The difference space for each specific station is quantitatively calculated using the following formula: S _space_j,t =θ j,t -h j,t ;

[0269] Among them, S _space_j,t Let θ be the difference space of the j-th factor on day t. j,t h is the discrimination threshold for the j-th factor on day t. j,t Let be the current monitoring value of the j-th factor on day t.

[0270] Based on the difference space generated by various hydrological stations throughout the entire basin, the discharge restriction coefficient of each cascade hydropower station or flow control gate in the target reservoir group is dynamically calculated.

[0271] During the calculation process, the discharge restriction coefficient shows a monotonically positive correlation with the spatial difference between various stations. That is, the more significantly the actual monitored water level or flow rate is below the discrimination threshold, the higher the flood safety margin of the basin, and the smaller the discharge restriction coefficient calculated by the system. Based on this, the control network dynamically reduces the gate opening commands of each hub in the target reservoir group to implement flow restriction and water storage.

[0272] To mitigate the risks of ecological disruption or bank collapse caused by a sudden drop in downstream river levels due to a one-time excessive reduction in discharge flow, a time-step control operator is introduced for discrete gradient processing when executing the control output. The system gradually and smoothly reduces the flow in stages according to the preset time gradient.

[0273] In one possible implementation, the specific verification process includes the following steps:

[0274] The system acquires all the discrimination criteria parameters completed in the aforementioned offline calibration, including the membership of the first-layer core factor set and the second-layer auxiliary factor set, the time-segmented discrimination thresholds, the number of days in the sliding monitoring window, the maximum allowable daily increase, and the optimal number of satisfactions. The system loads these parameters as a unified configuration instruction set into the historical backtracking simulation engine. Simultaneously, the system retrieves the original matrix of daily hydrological factors over many years, which completely contains the measured daily water level and flow observation sequences for dozens of consecutive historical years, serving as an unbiased historical scenario base for panoramic verification.

[0275] The retrospective simulation engine uses historical years as the basic loop unit to simulate the state machine behavior of the online operating system. Within the flood season window of each reviewed year, the system reads historical hydrological data daily, starting from the set allowable judgment start date, and applies a heterogeneous condition combination discrimination network for logical judgment. The system records the specific calendar time when the two-layer discrimination conditions are first fully triggered in each year, defining it as the simulated judgment date for that historical year. By traversing all historical samples, the system generates a continuous time series set consisting of the simulated judgment dates for each year.

[0276] A true value for panoramic verification is introduced. The system extracts the annual sequence of flood-risk-free dates established by the water level safety boundaries of each control station during the aforementioned offline construction process. The data processing unit compares the simulated judgment date of each year with the corresponding flood-risk-free date in a time series, and calculates the time evolution deviation of each historical year under the current judgment standard.

[0277] The absolute deviation for each historical year is quantified using the following formula:

[0278] E _day_y =abs(T _sim_y -T _real_y );

[0279] Among them, E_day_y Let T be the absolute deviation for the y-th historical year, and abs be the absolute value function. _sim_y T is the simulated discrimination date output by the system in the y-th historical year. _real_y Let y be the date with no flood risk corresponding to the y-th historical year.

[0280] The calculation results for all years are aggregated and summarized, and a comparison table of year-by-year discrimination results is automatically generated.

[0281] After generating a yearly comparison table of discrimination results, the boundary anomaly interceptor is activated to retrieve and extract specific abnormal years where the absolute deviation exceeds a preset tolerance threshold. In this embodiment, the preset tolerance threshold is strictly limited to 15 days. The system isolates samples with an absolute deviation greater than 15 days and classifies them into a high-dimensional root cause diagnosis pool, triggering a multi-factor hydrological root cause diagnosis and tracking program.

[0282] When performing multi-factor root cause diagnosis, the data analysis module performs refined slice analysis of the complete flood season hydrological process line for abnormal years. For example, if a year exhibits a specific deviation where the system's simulated judgment date is more than 15 days earlier than expected, the system automatically retrieves all water level and flow records after the simulated judgment date for that year, identifying situations where specific stations, such as the Shashi hydrological station, experience water levels exceeding warning levels but not exceeding the hydrodynamic boundary of the flood control guarantee level. The system quantitatively assesses whether this earlier-than-expected phenomenon poses a substantial threat to the actual flood control safety at the time. In another example, if a year exhibits a loss of benefits due to a simulated judgment date more than 15 days later than expected, the system uses time-series derivative analysis to detect whether the water levels of specific hydrological factors have remained within a narrow oscillation range slightly above the judgment threshold for several weeks, thus preventing the heterogeneous network from forming a closed loop for an extended period.

[0283] The data analysis module integrates the physical and statistical mechanism diagnostic conclusions of each abnormal year, and outputs a multi-dimensional integrated deviation year diagnostic report. Based on the bottleneck characteristics disclosed in the deviation year diagnostic report, the system issues parameter fine-tuning feedback instructions to the user.

[0284] On-site maintenance personnel can adjust the threshold buffer scale of Implementation Example 4 or appropriately widen the historical monitoring window days accordingly, thereby eliminating specific disturbances in extreme years. This panoramic retrospective and diagnostic feedback mechanism helps improve the discrimination criteria to have complete anti-attack capability and global generalization capability before delivery to industrial-grade field control networks.

[0285] According to one aspect of this application, after obtaining the current monitoring values ​​of each factor in the first-layer core factor set and the second-layer auxiliary factor set, and before performing the verification, the method further includes performing a health check and same-cluster substitution operation on the factors in the first-layer core factor set:

[0286] The monitoring data status of each factor in the first-level core factor set is checked one by one. When the data missing time of the target core factor exceeds the preset allowable time, the daily value change exceeds the preset limit change, or the continuous monitoring data remains unchanged for a preset number of days, the target core factor is determined to be in an invalid state.

[0287] The category affiliation relationship of the factor hierarchical clustering is used as a substitute mapping table. From the second-level auxiliary factor set, all auxiliary factors belonging to the same cluster category as the target core factor in the failed state are selected, and the auxiliary factor with the highest security contribution score is selected to be promoted to the first-level core factor set for replacement.

[0288] During the process of selecting auxiliary factors to replace the first-level core factor set, a threshold tightening penalty is applied to the upgraded auxiliary factors. Specifically, this includes deducting a certain percentage of the threshold buffer scale as the new threshold based on the original threshold corresponding to the upgraded auxiliary factor.

[0289] After the replacement execution, if the original core discrimination conditions for the target core factor in the first-level core factor set are replaced, and the total number of remaining auxiliary factors in the second-level auxiliary factor set is less than the pre-configured optimal number of satisfactions due to the removal of members, then the effective auxiliary discrimination condition requirement to meet the standard will be adaptively modified to be equal to the total number of the remaining auxiliary factors, and the downgrade record in the running log will be triggered at the same time.

[0290] When the monitoring data of the target core factor that was in a failed state returns to normal and the health check is continuously met for a predetermined number of historical days, the original discrimination status of the core factor in the first-level core factor set is restored, and the upgraded auxiliary factor is reset to the second-level auxiliary factor set.

[0291] Wherein, if all factors in the first-layer core factor set satisfy the core discrimination condition, and the number of factors in the second-layer auxiliary factor set satisfying the auxiliary discrimination condition reaches the pre-configured optimal satisfaction number, then the target watershed is determined to have entered the post-flood season, replaced by:

[0292] If all factors in the first-layer core factor set meet the core discrimination conditions, and the ratio of the sum of the safety contribution scores of each factor in the second-layer auxiliary factor set that meets the auxiliary discrimination conditions to the sum of the total safety contribution scores of all factors in the second-layer auxiliary factor set reaches the pre-configured weighted satisfaction rate threshold, then the target watershed is determined to have entered the post-flood season.

[0293] It should be noted that the following parameters are standard engineering configurations.

[0294] Core contribution threshold: The arithmetic mean of the safety contribution scores of the preferred set of hydrological factors.

[0295] Safety correlation threshold: 0.6 to 0.8 for quantitative mode; value of key node level for qualitative mode.

[0296] Capacity upper and lower limits: 10% to 30% of the total number of control stations;

[0297] The upper limit of the premature rate is generally 10% (at most 1 year out of 10 in the validation set is premature).

[0298] Threshold buffer scale: 5% to 10% of the annual average annual water level variation (annual highest minus annual lowest) of the station.

[0299] Baseline receding slope scale: the median of the absolute value of the regression slope within the 5-day window of the historical receding section of the station.

[0300] Maximum permissible daily increase: 90th percentile of the daily increase during the historical actual recession period.

[0301] Composite weight: based on the ratio of the average annual flow of each station to the total flow of the stations participating in the composite, or based on the ratio of the controlled catchment area.

[0302] In case of missing end-of-plum rain signal: When there is no external end-of-plum rain signal, the end of the plum rain season shall be deemed to have occurred when the rainfall in the basin is less than 5 mm for 5 consecutive days and the water level at the main control station is continuously receding.

[0303] Mapping of discharge restriction coefficient: Take the minimum value s* of the normalized difference space of each station, and the discharge restriction coefficient c = 1 - (1 - c) min min(1,s*), c min This is the minimum guarantee factor for leakage.

Claims

1. A post-flood season dynamic discrimination method based on hierarchical screening of hydrological factors and combination of multiple conditions, characterized in that, include: Acquire daily hydrological data and pre-configured discrimination parameters for the target watershed; Multidimensional attributes of hydrological factors are extracted based on daily hydrological data, and importance classification is performed based on the multidimensional attributes to obtain a first-level core factor set and a second-level auxiliary factor set. Obtain the current monitoring value of each factor in the first-level core factor set and the second-level auxiliary factor set; Within the window period during which the timing constraints are met on the current date, verification is performed based on the current monitoring values ​​and discrimination parameters. If all factors in the first-layer core factor set meet the core discrimination conditions, and the number of factors in the second-layer auxiliary factor set that meet the auxiliary discrimination conditions reaches the pre-configured optimal number of conditions met, then the target watershed is determined to have entered the post-flood season, and the dynamic discrimination result is output. The pre-configured optimal satisfaction number is pre-constructed through the following offline optimization steps: The historical year data is divided into a training year set and a validation year set; In the training year set, dates with no flood risk defined by historical measured water conditions are extracted, and the asymmetric risk loss and discrimination stability index under each candidate satisfaction number are calculated respectively. The comprehensive optimization objective is then integrated to screen out the target candidate satisfaction number. The target candidate satisfaction number is applied to the set of verification years to perform a premature risk test and calculate the premature verification rate. When the premature verification rate is not higher than the pre-configured upper limit of the premature rate, the target candidate satisfaction number is established as the optimal satisfaction number. The calculation process for asymmetric risk loss includes: Calculate the deviation of the judgment date corresponding to the number of candidates that meet the criteria from the date of no flood risk, whether the date is earlier or later than the date of no flood risk. An early penalty coefficient is applied to the deviation amount representing early behavior, and a late penalty coefficient is applied to the deviation amount representing late behavior, wherein the early penalty coefficient is strictly greater than the late penalty coefficient. The discrimination stability index is characterized based on the variance characteristics of the discrimination date deviation for each year in the training year set; Among them, the hydrological factors in the first-layer core factor set and the second-layer auxiliary factor set include water level-type factors and flow-type factors; For water level-type factors, the corresponding core and auxiliary discrimination conditions are: the current monitoring value is lower than the corresponding discrimination threshold, and the corresponding real-time water receding trend confirmation result is in a valid confirmation state. For flow-type factors, the corresponding core and auxiliary discrimination conditions are both: the synthetic flow generated based on the current monitoring value is lower than the corresponding discrimination threshold.

2. The method according to claim 1, characterized in that, Extracting multidimensional attributes of hydrological factors, including: Hierarchical clustering of various hydrological factors was performed based on daily hydrological data to obtain factor clustering results; Within each category of the factor clustering results, a joint evaluation of information content and seasonality index is performed to screen out candidate representative factors. Inter-class redundancy elimination is performed on the candidate representative factors to obtain the target hydrological factor set; Among them, the multidimensional attributes include seasonality index and independence based on inter-class redundancy resolution representation.

3. The method according to claim 2, characterized in that, The joint assessment includes: Based on circular statistics, the flood season sequence is converted into directional angles to determine the seasonality index, and information entropy is calculated based on probability discretization; Factors whose seasonality index and information entropy both satisfy the corresponding intra-class benchmark are retained as candidate representative factors. Inter-class redundancy resolution includes: Calculate the multiple correlation coefficient between each candidate representative factor and the remaining candidate representative factors; Factors with multiple correlation coefficients exceeding the pre-configured correlation threshold are removed to obtain the target hydrological factor set.

4. The method according to claim 2, characterized in that, Based on the importance ranking of multidimensional attributes, a first-level core factor set and a second-level auxiliary factor set are obtained, including: Obtain the correlation degree between each factor in the target hydrological factor set and the flood control safety of the target flood control section; Based on seasonality index, independence, and correlation with flood control safety, the safety contribution score of each factor is calculated; Based on the security contribution score and the pre-configured capacity constraints, dynamic truncation and partitioning are performed to obtain the first-layer core factor set and the second-layer auxiliary factor set.

5. The method according to claim 4, characterized in that, The safety contribution score is obtained by weighted summation of seasonality index, independence indicator, and flood control safety correlation. Dynamic truncation and partitioning are performed based on security contribution scores and pre-configured capacity constraints, including: Sort the factors in the target hydrological factor set in descending order of their safety contribution scores; Factors whose safety contribution score and flood control safety correlation both meet the corresponding grading thresholds are classified into the first-level core factor set. If the number of factors that meet the grading threshold exceeds the capacity limit, only the factors with the highest security contribution score and whose number equals the capacity limit will be retained. If the number of factors that meet the grading threshold is less than the lower limit of capacity, then the remaining factors that have not met the grading threshold are added in descending order of safety contribution score until the number reaches the lower limit of capacity; and the remaining factors are assigned to the second-level auxiliary factor set.

6. The method according to claim 1, characterized in that, When hydrological factors exhibit water level characteristics, the real-time receding trend of each factor can be determined through any of the following methods: Based on the rate of change of water level and the maximum daily fluctuation range within the historical monitoring window, Boolean judgment is performed to obtain the real-time confirmation result of the receding water trend. Alternatively, based on multidimensional hydrological control indicators, continuous spatial mapping and coupling can be performed to calculate the water discharge confidence level. When the water discharge confidence level reaches the pre-configured confidence level standard, the real-time water discharge trend confirmation result can be marked as an effective water discharge status.

7. The method according to claim 6, characterized in that, The confidence level of the receding water level is calculated to obtain the real-time receding water trend confirmation result, which specifically includes: Based on the current monitoring value of water level factors and the corresponding discrimination threshold, calculate the threshold margin index that characterizes the degree to which the water level is below the threshold. The receding slope index is calculated based on the regression status of the monitoring sequence within the historical monitoring window. The anti-rise constraint index is calculated based on the maximum single-day rise within the historical monitoring window. The confidence level of the receding water level is obtained by multiplying and coupling the threshold margin index, the receding water slope index, and the reverse rise constraint index.

8. The method according to claim 1, characterized in that, The pre-configured discrimination parameters include discrimination thresholds for each factor, and for factors with obvious flood season peak characteristics, the discrimination thresholds include time-period thresholds; The time-period thresholds are determined based on the peak dates of historical hydrological statistics: when the current determination date is earlier than the peak date of historical hydrological statistics, the first threshold is used; When the current determination date is equal to or later than the peak date of historical hydrological statistics, a second threshold is used; wherein the second threshold is greater than the first threshold. The pre-configured discrimination parameters include discrimination thresholds for each hydrological factor; these discrimination thresholds are pre-constructed through the following offline calibration steps: Historical values ​​of the start and end of the recession for each hydrological factor at the moment when the recession begins after the highest historical flood level are obtained, and multiple ranges of start and end values ​​are obtained according to the sample distribution. If the number of historical year samples in any of the aforementioned start-to-end value intervals is less than the preset minimum number of samples, an adaptive merging operation with the adjacent start-to-end value intervals is performed until the number of samples in each independently running start-to-end value interval meets the constraint condition of the minimum number of samples. After the merging operation, the subsequent alarm frequency of each of the starting and ending value intervals is statistically analyzed, and the interval boundary value of the starting and ending value interval where the subsequent alarm frequency is not higher than the preset safety level is selected and established as the corresponding discrimination threshold. Before obtaining the historical start and end value samples, the following steps are also included: Historical timelines of each hydrological control station are extracted year by year, and historical years are divided into years exceeding the warning level and years not exceeding the warning level according to whether the annual highest water level exceeds the warning level. Within the years exceeding the warning level and the years not exceeding the warning level, based on the temporal relationship between the rise and fall of the water level, it is further divided into the rising-receding pattern after the flood peak has passed, and the receding-rising pattern in which the water level falls for a preset number of days before reaching the annual maximum value and then rises again. Based on this classification, the corresponding historical starting and receding value samples are extracted. For years with rising and receding water patterns, the moment when the derivative reverses after the highest annual water level extreme is taken as the actual start and receding time. The water level sequence from this moment until the end of the flood season window and before the water level exceeds the warning level again is recorded in the real receding positive sample database. For years with receding and rising water levels, in addition to extracting the actual receding segment after the annual peak value according to the rules, the phased decline segment before the annual peak value that meets the constraint of continuously decreasing number of days is marked as a false receding segment and entered into the false receding negative sample library.

9. The method according to claim 1, characterized in that, Before performing the verification, check whether the current date satisfies the timing prerequisite constraints; the timing prerequisite constraints include the end of the plum rain season date condition and the earliest allowed judgment date condition; When verifying whether the timing constraints are met, the specific steps include: determining the end date of the plum rain season issued by the external meteorological department as the end date of the plum rain season, and determining the period after the end date of the plum rain season but not earlier than the preset earliest allowed judgment date as the window period for allowing the dynamic judgment of the post-flood season.

Images