Intelligent early warning method and system for rainstorm flood of Yellow River

Abstract

The invention discloses a Yellow River rainstorm flood intelligent early warning method and system, and relates to the technical field of hydrological early warning data processing, and the method comprises the steps: aligning Yellow River multi-source data to generate node time sequence features, and constructing a Yellow River network directed graph. Propagation reasoning is carried out on the Yellow River network directed graph according to flood discharge propagation time lag, and the model is trained through conservation constraint and propagation consistency constraint. Predicting the flood risk index of the Yellow River, performing graded early warning, and updating the time-delay parameters on line according to actual measurement feedback. According to the method, river network topology and flood discharge time lag alignment are taken as a main line, multi-source observation is unified and structured into a node time sequence matrix and a directed river network, conservation and propagation consistency constraints are introduced into space-time diagram reasoning, and a multi-step water level and peak time window for early warning is output; and through risk index grading and actual measurement feedback, time delay is corrected online, and prediction, early warning and calibration closed loop are realized.

Description

A Smart Early Warning Method and System for Rainstorms and Floods in the Yellow River Technical Field

[0001] This invention relates to the field of hydrological early warning data processing technology, specifically to an intelligent early warning method and system for torrential rain and floods in the Yellow River. Background Technology

[0002] In recent years, the technical system for early warning of rainstorms and floods in watersheds has gradually evolved from traditional mechanistic models to a fusion framework integrating multi-source observation, digital processing, and intelligent reasoning. Multi-source data, including rain gauge data, radar echoes, satellite precipitation retrieval, and soil moisture, are incorporated into a unified spatiotemporal benchmark. Through data cleaning, alignment, feature engineering, and model training, rapid prediction of flood events is achieved. Simultaneously, deep learning sequence models and spatiotemporal graph models are used to characterize nonlinear rainfall-runoff response, forming a deployable real-time computing link on a cloud-edge collaborative platform.

[0003] However, existing technologies still face unavoidable bottlenecks in the Yellow River torrential rain and flood scenario. Many methods treat each station as an independent time series or only perform weak coupling fusion, failing to explicitly incorporate the directed connectivity of the Yellow River's main stream and tributaries and the flood propagation time lag into the computational graph. This leads to phase shifts in upstream and downstream responses, inconsistent propagation of tributary inflow effects, and systematic deviations in peak arrival time windows. Multi-source observations suffer from inconsistent sampling frequencies, occlusion and inversion errors, missing data, and anomalous mutations. Without alignment and robust processing mechanisms tailored to the river network structure, models are prone to input noise amplification and prediction drift under extreme conditions. Purely data-driven models often lack conservation constraints and propagation consistency constraints for river processes. Training objectives and inference links struggle to guarantee physical or structural consistency across nodes, resulting in insufficient stability under long prediction steps or strong convective bursts. Furthermore, most systems lack a closed-loop mechanism for online self-calibration of key propagation parameters based on measured feedback, making it difficult to quickly correct accumulated errors during operation. Consequently, they cannot simultaneously meet the engineering requirements of multi-step water level prediction, tiered early warning, and sustainable online calibration. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that existing methods for processing and predicting Yellow River rainstorm and flood warning data have the following problems: they do not explicitly utilize the directional topology of the Yellow River network and the time delay of flood propagation, resulting in upstream and downstream prediction phase deviations; insufficient multi-source observation alignment and handling of abnormal missing data lead to model drift under extreme processes; and the lack of conservation constraints, propagation consistency constraints, and self-calibration mechanisms during operation leads to insufficient prediction stability. The invention also addresses how to achieve multi-step water level prediction based on river network topology and time delay alignment, and how to perform graded early warning and online time delay parameter updates.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a method for intelligent early warning of rainstorms and floods in the Yellow River, comprising aligning the temporal characteristics of multi-source data generation nodes of the Yellow River and constructing a directed graph of the Yellow River network.

[0007] The model is trained by reasoning about the propagation time lag of flood flow on the directed graph of the Yellow River network, and by constraining conservation and propagation consistency.

[0008] Predict the Yellow River flood risk index and issue graded early warnings, and update time-delay parameters online based on actual measurement feedback.

[0009] As a preferred embodiment of the intelligent early warning method for rainstorms and floods in the Yellow River described in this invention, the step of aligning the multi-source data of the Yellow River to generate the time-series features of the nodes includes receiving rainfall from Yellow River rain gauge stations, radar echo statistics, satellite precipitation inversion, soil moisture index, and historical water level sequences of Yellow River stations; unifying the data sources to a preset time granularity and mapping them with Yellow River node numbers; performing amplitude constraints and mutation detection and removal on outliers; interpolating missing data segments according to the maximum tolerance duration; and performing normalization and time window truncation on each dimension of features to form a Yellow River node time-series feature matrix with nodes as indexes and time as sequences.

[0010] As a preferred embodiment of the intelligent early warning method for Yellow River rainstorms and floods described in this invention, the construction of the directed graph of the Yellow River network includes: reading the Yellow River digital river network or connectivity table; determining the node set as Yellow River stations, cross sections, or grid confluence units; and determining the edge set as upstream-downstream connectivity with the edge direction pointing from upstream to downstream. For each edge, edge attributes including channel distance, tributary inflow type, and control section identifier are recorded. A Yellow River network adjacency matrix corresponding one-to-one with the node set and the incoming edge set of each node are generated. The Yellow River network adjacency matrix is ​​then associated and stored with the Yellow River node temporal characteristic matrix as the topological constraint input for flood propagation.

[0011] As a preferred embodiment of the intelligent early warning method for torrential rain and floods in the Yellow River described in this invention, the propagation inference on the directed graph of the Yellow River network based on the flood propagation time delay includes: calculating the flood propagation time delay parameter for each directed edge of the Yellow River network directed graph. The initial value of the flood propagation time delay parameter is obtained by converting the river channel distance in the edge attribute with a preset average flow velocity, and corrected by combining the historical flood peak arrival time difference. The flood propagation time delay parameter is restricted to non-negativity and quantized to a preset time granularity. When the flood propagation time delay parameter is not an integer time step, an interpolation method is used to obtain the alignment state of the upstream node at the current time minus the flood propagation time delay parameter. During propagation inference, the state of the incoming edge node of each node is first aligned in time according to the flood propagation time delay parameter, and then weighted aggregation is performed to obtain the propagation input of the node.

[0012] As a preferred embodiment of the intelligent early warning method for Yellow River torrential rain and floods described in this invention, the training model based on conservation constraints and propagation consistency constraints includes a Yellow River network spatiotemporal graph prediction model. The model takes the temporal feature matrix of Yellow River nodes and the adjacency matrix of the Yellow River network as input, and propagates and infers the future multi-step water levels and hidden node representations of Yellow River nodes on the directed graph of the Yellow River network according to the flood propagation time lag. Simultaneously, the equivalent storage and discharge water level change is output from the output branch of the equivalent storage and discharge water level change of Yellow River nodes. The prediction error is calculated between the future multi-step water levels of Yellow River nodes output by the propagation inference and the measured water levels at the corresponding times, and conservation constraint terms and propagation consistency constraint terms are calculated. The conservation constraint term is obtained by comparing the weighted sum of the predicted water level of the same node and the predicted water level of the node's incoming edges after alignment with the flood propagation time lag, then subtracting the difference after deducting the equivalent storage and discharge water level change, and squaring the result. The propagation consistency constraint is obtained by comparing the difference between the hidden representations of downstream nodes and upstream nodes after alignment with the flood propagation time delay for each directed edge, and then taking the square of the norm. The prediction error, conservation constraint, and propagation consistency constraint are weighted and summed according to the weight coefficients of the Yellow River water level multi-step prediction error, the Yellow River section conservation constraint, and the Yellow River network propagation consistency constraint to form the training objective. Based on the training objective, gradient updates are performed on the parameters of the Yellow River network spatiotemporal map prediction model and the flood propagation time delay parameters, and the process is iterated until the preset number of training rounds is met.

[0013] As a preferred embodiment of the intelligent early warning method for Yellow River torrential rain and floods described in this invention, the method of predicting the Yellow River flood risk index and issuing graded early warnings includes: calculating normalized values ​​of water levels exceeding the Yellow River section's safe water level benchmark, water levels exceeding the Yellow River section's warning water level benchmark, and the predicted water level growth rate based on future multi-step water levels; synthesizing the Yellow River flood risk index by applying weighting coefficients for exceeding the Yellow River section's safe water level, warning water level, and predicted water level growth rate; and generating a blue, yellow, orange, or red early warning based on the segmented interval mapping of the Yellow River torrential rain and flood early warning threshold, and outputting the corresponding Yellow River node identifier and the predicted peak arrival time window.

[0014] As a preferred embodiment of the intelligent early warning method for Yellow River rainstorms and floods described in this invention, the online update of the time delay parameters based on measured feedback includes receiving the latest measured water level at the same Yellow River node and calculating the deviation with the corresponding predicted value. When the deviation exceeds the Yellow River early warning calibration error threshold five times consecutively within a rolling window, an online update is triggered. During the online update, a small-step iterative method is used to correct the flood propagation time delay parameters.

[0015] As a preferred embodiment of the Yellow River rainstorm and flood intelligent early warning system described in this invention, it includes a multi-source alignment river network construction module, a time-delay propagation constraint training module, and a risk early warning feedback update module.

[0016] The multi-source alignment and river network construction module is used to align the temporal characteristics of the multi-source data generation nodes of the Yellow River and construct a directed graph of the Yellow River network.

[0017] The time-delay propagation constraint training module is used to propagate reasoning on the directed graph of the Yellow River network according to the flood propagation time delay, and to train the model with conservation constraints and propagation consistency constraints.

[0018] The risk warning feedback update module is used to predict the Yellow River flood risk index and issue graded warnings, and updates the time-delay parameters online based on actual measurement feedback.

[0019] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program as steps to implement an intelligent early warning method for torrential rain and floods in the Yellow River.

[0020] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a method for intelligent early warning of torrential rains and floods in the Yellow River basin.

[0021] The beneficial effects of this invention are as follows: By generating a time-series feature matrix of Yellow River nodes through multi-source data alignment, abnormal mutation removal, missing data completion, normalization, and time window extraction, and constructing a directed graph and adjacency matrix of the Yellow River network, heterogeneous observations such as rainfall, echo, inversion, soil, and water level are unified into a single time index and node numbering system. Directed connectivity is then used as the topological prior for subsequent calculations. Its function is to provide consistent and reproducible input tensors and upstream / downstream structural constraints for propagation inference, thereby improving data availability and computational link stability under extreme rainstorm scenarios.

[0022] By calculating and correcting the flood propagation time delay parameters for each directed edge of the river network, interpolating and aligning non-integer time delays, and then weighted aggregating propagation inference on the time-delay-aligned incoming edge set, the phase alignment of the upstream and downstream responses of the Yellow River and the structured transmission of the tributary inflow influence were achieved. Simultaneously, conservation constraints, propagation consistency constraints, and prediction errors were used to form the training objective, and time delay parameters were incorporated into gradient updates. This ensures that the model simultaneously satisfies the topological propagation law and cross-node consistency requirements during training and inference, enabling the output of future multi-step water level and peak arrival time series information, thus suppressing long-step prediction drift in pure black-box systems.

[0023] By converting future multi-step water levels into three normalized quantities—exceeding safe water levels, exceeding warning water levels, and water level growth rates—and synthesizing a Yellow River flood risk index sequence with fixed weights, continuous forecast results are mapped into actionable tiered early warning decisions. An early warning is triggered by exceeding a threshold after three consecutive forecast steps, and a peak arrival time window is output to support scheduling and response timing. Simultaneously, online updates are triggered by consecutive exceedances of measured and predicted deviations within a rolling window, using small steps to correct flood propagation time delay parameters. This allows for self-calibration of systematic phase deviations during operation, maintaining long-term consistency in online early warnings. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 is an overall flowchart of an intelligent early warning method for rainstorms and floods in the Yellow River provided in Embodiment 1 of the present invention. Detailed Implementation

[0026] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0027] Example 1, referring to Figure 1, is an embodiment of the present invention, providing a method for intelligent early warning of rainstorms and floods in the Yellow River, including: S1: aligning the temporal characteristics of the multi-source data generation nodes of the Yellow River to construct a directed graph of the Yellow River network.

[0028] The system receives rainfall data from Yellow River rain gauge stations, radar echo statistics, satellite precipitation inversion, soil moisture index, and historical water level sequences from Yellow River stations. It unifies all data sources to a preset time granularity and maps them using Yellow River node numbers. It performs amplitude constraints and abrupt change detection and removal on outliers, interpolates missing data segments according to the maximum tolerance duration, and performs normalization and time window truncation on each dimension of features to form a Yellow River node time series feature matrix with nodes as indexes and time as sequences.

[0029] Furthermore, the preset time granularity is set to 10 minutes. During alignment, rain gauge and satellite precipitation inversion are accumulated over 10 minutes, radar echo statistics are averaged and maximized over 10 minutes, and water levels are averaged over 10 minutes. When the original sampling frequency of a data source is higher than 10 minutes, window aggregation is used; when it is lower than 10 minutes, nearest neighbor preservation is used and a preservation flag is recorded to ensure that the features of each dimension under the same Yellow River node number are aligned at the same time index. The preservation flag and the missing measurement mask are both added as additional feature dimensions to the Yellow River node temporal feature matrix using binary encoding, where 1 indicates that nearest neighbor preservation or missing measurement has occurred, and 0 indicates that it has not occurred.

[0030] The threshold for amplitude constraints is determined based on a physically reasonable range and combined with high-quantile statistics of historical samples from the Yellow River Basin. The threshold values ​​are fixed as follows: 10-minute cumulative rainfall at rain gauge stations is limited to 0–35 mm; radar echo statistics are limited to 0–70 dBZ; 10-minute cumulative rainfall retrieved from satellite precipitation is limited to 0–35 mm; soil moisture index is limited to 0–1; and Yellow River station water levels are limited to the range of the lowest to the highest water level in the past 5 years plus 1.0 m. Samples exceeding these ranges are considered abnormal and removed. Removed samples are marked as missing values ​​in the Yellow River node temporal feature matrix, triggering interpolation completion or missing value masking processes to maintain the continuity of the temporal index at the preset time granularity for each node.

[0031] The threshold for mutation detection is derived from a dual-condition rule based on absolute and relative mutation amounts. The threshold values ​​are fixed: for water levels, a mutation is defined as an absolute difference between two adjacent time points greater than 0.30 m and simultaneously greater than three times the median absolute deviation of the difference sequence over the past 24 hours for that node. For 10-minute cumulative rainfall, a mutation is defined as an absolute difference between two adjacent time points greater than 20 mm. Samples identified as mutations are treated as missing data and enter the interpolation completion process.

[0032] Furthermore, the maximum tolerance period is set to 60 minutes. When the continuous missing data does not exceed 60 minutes, linear interpolation is used for rain gauges, satellite precipitation inversion, and water levels; forward hold interpolation is used for radar echo statistics; and exponential smoothing interpolation is used for the soil moisture index. The smoothing coefficient of exponential smoothing interpolation is fixed at 0.3, and the soil moisture index interpolation value is obtained by weighted recursion of the current observation and the interpolation result of the previous moment. When the continuous missing data exceeds 60 minutes, no numerical interpolation is performed, but a missing data mask is output, and a missing data marker dimension is added to the time series feature matrix of the Yellow River node. The time window truncation length is fixed to the most recent 6 hours, that is, 36 time steps are taken at a 10-minute granularity as the input sequence for a single propagation inference. Normalization adopts the intra-node standardization rule, that is, z-score normalization is performed using the mean and standard deviation of the node over the past 30 days, and the result is clipped to [-5, 5].

[0033] Read the Yellow River digital river network or connectivity table to determine the node set as Yellow River stations, cross sections, or grid confluence units, and determine the edge set as upstream-downstream connectivity with edge direction from upstream to downstream. Record edge attributes for each edge, including channel distance, tributary inflow type, and control section identifier. Generate a Yellow River network adjacency matrix corresponding one-to-one with the node set and the incoming edge set for each node. Link and store the Yellow River network adjacency matrix with the Yellow River node temporal characteristic matrix as topological constraint input for flood propagation.

[0034] Furthermore, the Yellow River node numbering adopts a unique three-segment coding rule: basin coding, river segment coding, and cross-section or station coding. The latitude and longitude of the cross-section are stored in a unified coordinate system and used to verify identical points with different coordinates. When the node set is a raster confluence unit, the raster resolution is fixed at 1km×1km, and the node number is generated based on the raster center point. The Yellow River network adjacency matrix is ​​constructed in the order of node numbers. If multiple tributaries flow into the same downstream node, multiple upstream incoming edges are retained in the adjacency matrix and arranged in ascending order of river distance in the incoming edge set.

[0035] Furthermore, the river channel distances are calculated from the mileage along the central line of the Yellow River digital river network and stored in kilometers. The tributary inflow types are limited to three enumerated values: main stream to tributary, tributary to tributary, and main stream to main stream. Control section identifiers are downstream node numbers. The Yellow River network adjacency matrix and the Yellow River node temporal characteristic matrix are stored together with the same version number. The version number is fixed and updated daily. When the river network connectivity or node set changes, a new version is generated, while historical versions are retained for retrospective reproduction of propagation and inference results.

[0036] It should be noted that this step unifies rainfall, echo, inversion, soil, and water level sequences at a 10-minute granularity, and combines amplitude constraints, abrupt change detection, and missing data masks to form a computable node temporal feature matrix. Simultaneously, the connectivity relationships of the Yellow River's main stream and tributaries are structured into a directed river network adjacency matrix and an incoming edge set, and stored with version numbers, providing a stable topological prior and a reproducible data foundation for subsequent graph reasoning based on flood propagation time delays.

[0037] S2: Based on the propagation time lag of flood flow on the directed graph of the Yellow River network, the model is trained with conservation constraints and propagation consistency constraints.

[0038] For each directed edge in the directed graph of the Yellow River network, the flood propagation time delay parameter is calculated. The initial value of the flood propagation time delay parameter is obtained by converting the river channel distance in the edge attributes with the preset average flow velocity, and corrected by incorporating the historical flood peak arrival time difference. The flood propagation time delay parameter is restricted to non-negativity and quantized to a preset time granularity. When the flood propagation time delay parameter is not an integer time step, interpolation is used to obtain the alignment state of the upstream node at the current time minus the flood propagation time delay parameter. During propagation inference, the state of the incoming edge nodes of each node is first aligned in time according to the flood propagation time delay parameter, and then weighted aggregation is performed to obtain the propagation input of the node.

[0039] Furthermore, the preset average flow velocity is fixed at 2.5 m / s, and its setting rule is to select the statistical average of the flood peak propagation velocity of typical flood events in the Yellow River main stream over the past 5 years. The preset time granularity is 10 minutes, and the initial value of the flood propagation time delay parameter is obtained by dividing the river channel distance by the preset average flow velocity and converting it to a 10-minute time step. When the river channel distance is stored in kilometers, it is first converted to meters before participating in the calculation of the initial value of the flood propagation time delay.

[0040] The initial value of the flood propagation time delay is calculated as follows:

[0041] in, Indicates from the upstream node to downstream nodes The initial value of the flood propagation time lag. Represents a directed edge The distance of the river channel. This indicates the preset average flow velocity of the Yellow River, which is fixed at 2.5 m / s. This indicates the number of seconds corresponding to the preset time granularity.

[0042] Calculate the value of each directed edge first. This serves as a benchmark for subsequent historical flood peak arrival time difference correction and propagation alignment interpolation, enabling different river sections to form a calculable time alignment relationship under the same 10-minute index.

[0043] Furthermore, the rule for correction based on historical flood peak arrival time differences is as follows: Select the historical water level sequences of the two nodes at both ends of each directed edge, extract the flood peak time during the flood rise period and calculate the arrival time difference. If this time difference can be obtained in at least three historical flood events, then take their arithmetic mean as the correction target. When the correction target is consistent with... When the difference exceeds 3 time steps, the flood propagation time delay parameter is gradually corrected with a step size limit of 0.2 time steps per training iteration to avoid the propagation inference being unstable due to sudden changes in the time delay parameter.

[0044] Furthermore, when the flood propagation time delay parameter is not an integer time step, linear interpolation is used. First, the flood propagation time delay parameter is decomposed to obtain the integer part and the fractional part. Then, the water level or hidden representation of the upstream node at two adjacent times is interpolated to obtain the alignment state.

[0045] The linearly aligned interpolation for non-integer time delays is expressed as:

[0046] in, This indicates the water level of the upstream node after alignment with time delay. Indicates upstream node The water level at the corresponding moment. This represents a directed edge. This represents the time delay parameter for flood propagation. Represents the integer part of the time delay. This represents the decimal part of the time delay, with a value range of [0,1].

[0047] when When the time step is a non-integer number, the downstream node At any moment The required upstream input comes from the state between two 10-minute sampling points, obtained through a linear alignment interpolation formula with non-integer time delays. This ensures that the propagation reasoning still has continuous time-delay expression capability at a 10-minute granularity.

[0048] The training model is a spatiotemporal graph prediction model for the Yellow River network. It takes the temporal feature matrix of Yellow River nodes and the adjacency matrix of the Yellow River network as input, and propagates the inference on the directed graph of the Yellow River network according to the flood propagation time lag to obtain the future multi-step water levels and hidden representations of Yellow River nodes. Simultaneously, it outputs the equivalent storage and discharge water level change from the output branch of the Yellow River node. The prediction error is calculated between the future multi-step water levels of Yellow River nodes output by the propagation inference and the measured water levels at the corresponding times. Conservation constraints and propagation consistency constraints are also calculated. The prediction error component for future multi-step water levels is obtained by summing the mean square error of the predicted water levels and measured water levels for the next 6 steps along the node and time dimensions. The conservation constraint is obtained by comparing the weighted sum of the predicted water levels of the same node and the predicted water levels of the node's incoming edges after alignment with the flood propagation time lag, then subtracting the equivalent storage and discharge water level change and squaring the difference. The propagation consistency constraint is obtained by comparing the difference between the hidden representations of downstream nodes and upstream nodes after alignment with the flood propagation time delay for each directed edge, and then taking the square of the norm. The prediction error, conservation constraint, and propagation consistency constraint are weighted and summed according to the weight coefficients of the Yellow River water level multi-step prediction error, the Yellow River section conservation constraint, and the Yellow River network propagation consistency constraint to form the training objective. Based on the training objective, gradient updates are performed on the parameters of the Yellow River network spatiotemporal map prediction model and the flood propagation time delay parameters, and the process is iterated until the preset number of training rounds is met.

[0049] Furthermore, the Yellow River network spatiotemporal map prediction model includes a sequentially connected Yellow River node feature encoding module, a Yellow River network time-delay alignment and propagation module, and a Yellow River multi-step water level prediction output module. The Yellow River node feature encoding module encodes the multi-source feature sequence of each node over the most recent 6 hours into a node hidden representation. The Yellow River network time-delay alignment and propagation module aligns the hidden representations of the upstream edges of each node according to the flood propagation time delay, calculates the edge weights, and aggregates them. The Yellow River multi-step water level prediction output module outputs the node water level prediction for the next 6 steps, i.e., 60 minutes. The Yellow River node equivalent storage and discharge water level change output branch shares the hidden representation output by the same propagation module as the Yellow River multi-step water level prediction output module, only using an independent linear layer at the end to output the equivalent storage and discharge water level change.

[0050] Furthermore, the weighted aggregation in the propagation inference adopts a calculation order of time-delay alignment followed by weighting. The edge weights are jointly determined by the current hidden representation of the downstream node and the aligned hidden representation of the upstream node, and the weights of all incoming edges of the same node are normalized to ensure aggregation stability. The normalized weight of the directed edge at time step is obtained by bilinear scoring of the current hidden representation of the downstream node and the aligned hidden representation of the upstream node, and softmax normalization is performed on the edge scores of all incoming edges of the same downstream node.

[0051] The time-delay aligned directed incoming edge weighted aggregation is represented as:

[0052] in, Indicates downstream node At any moment The propagation input vector. This represents the pointing nodes in the directed graph of the Yellow River network. The set of upstream nodes of the incoming edge. Represents a directed edge At any moment Normalized weights. This represents the transformation matrix of the Yellow River's propagation characteristics. Indicates upstream node The hidden representation is a vector aligned according to the propagation delay of the flood.

[0053] For each downstream node, first obtain the upstream aligned hidden representation according to the time delay, then according to... Aggregation is performed to obtain This is then used to generate multi-step water level predictions for nodes and to output branches for equivalent storage and discharge water level changes, thereby writing the Yellow River network topology and flood discharge time delay into the inference link.

[0054] Furthermore, the preset training rounds are fixed at 50 rounds. During training, a learning rate of 0.001 and a batch size of 64 are used for gradient updates, and an early stopping threshold is set: iteration stops when the training objective on the validation set decreases by less than 0.0001 within 5 consecutive rounds. The weighting rules for the Yellow River water level multi-step prediction error weighting coefficient, the Yellow River section conservation constraint weighting coefficient, and the Yellow River network propagation consistency constraint weighting coefficient are as follows: first, they are initialized with equal weights, then normalized once according to the magnitudes of the three components in the validation set, finally set to fixed values ​​of 0.6, 0.25, and 0.15, respectively.

[0055] The weighted representation of the training objective is as follows:

[0056] in, This represents the training target value for the Yellow River network spatiotemporal map prediction model. This represents the error components of future multi-step water level prediction. The component representing the conservation constraint term is obtained by summing the weighted sum of the predicted water level and the water level aligned with the inlet edge, and subtracting the squared difference between the equivalent storage and release water level changes. The component representing the propagation consistency constraint term is obtained by summing the squared differences in the norms of the downstream hidden representation and the upstream aligned hidden representation. This represents the error weighting coefficient for the multi-step prediction of the Yellow River water level, which is fixed at 0.6. This represents the conservation constraint weighting coefficient for the Yellow River section, which is fixed at 0.25. This represents the weighting coefficient for the consistency constraint of the Yellow River network propagation, which is fixed at 0.15.

[0057] Calculate for each batch during training , , And synthesized according to formula 4 ,by Gradient updates are performed on the model parameters and flood propagation time delay parameters. Training ends when 50 training rounds are completed or the early stopping threshold is triggered, and the obtained flood propagation time delay parameters and model parameters are used for subsequent Yellow River flood risk index prediction and graded early warning.

[0058] It should be noted that this step explicitly incorporates the connectivity of the Yellow River network and the flood propagation time delay into the graph inference link: First, initial edge time delay values ​​are obtained based on channel distance and statistical flow velocity, and progressively corrected using historical flood peak arrival time differences. Then, non-integer time delay linear alignment interpolation is used to align the upstream state to the current downstream time. Subsequently, normalized weights are calculated on the aligned incoming edge set and aggregated to form the propagation input, outputting future multi-step water levels and hidden representations. The model is then jointly trained using a weighted objective of water level prediction error, conservation constraints, and propagation consistency constraints, while simultaneously updating the model parameters and time delay parameters. This differs from conventional models that only perform time-series fitting by binding topology, time delays, and constraints to the training objective and propagation calculation, thus making time delays learnable and inference alignment executable.

[0059] S3: Predict the Yellow River flood risk index and issue graded early warnings, and update time-delay parameters online based on actual measurement feedback.

[0060] Based on future multi-step water levels, normalized values ​​exceeding the Yellow River section's safe water level benchmark, exceeding the Yellow River section's warning water level benchmark, and the predicted water level growth rate are calculated separately. These normalized values ​​are then combined using the Yellow River section's safe water level exceeding limit weighting coefficient, the Yellow River section's warning water level exceeding limit weighting coefficient, and the Yellow River section's predicted water level growth rate weighting coefficient to synthesize the Yellow River flood risk index. When the Yellow River flood risk index exceeds the Yellow River rainstorm and flood warning issuance threshold within three consecutive prediction steps, a blue, yellow, orange, or red warning is generated based on the segmented interval mapping of the Yellow River rainstorm and flood warning issuance threshold, and the corresponding Yellow River node identifier and predicted peak arrival time window are output. The Yellow River flood risk index is calculated separately for each prediction step of the future multi-step water levels to obtain a risk index sequence. When the risk index corresponding to three consecutive adjacent prediction steps in the risk index sequence is greater than the Yellow River rainstorm and flood warning issuance threshold, the triggering condition is met.

[0061] Furthermore, the warning water level benchmark value of the Yellow River section is read from the Yellow River node water level threshold table, and the safe water level benchmark value of the Yellow River section is set according to a fixed rule as the warning water level benchmark value of the Yellow River section minus 0.50m. The Yellow River node water level threshold table is updated annually and stored with the Yellow River node number as an index, so that different sections have a reproducible source of benchmark value.

[0062] The normalized values ​​are calculated as follows: for the values ​​exceeding the Yellow River section's safe water level benchmark and the Yellow River section's warning water level benchmark, the maximum exceedance value among the future multiple water levels is taken, and then divided by 1.00m before being cropped to 0-1. For the predicted water level growth rate, the maximum increment of the adjacent prediction steps is taken, and then divided by 0.10m / 10 minutes before being cropped to 0-1, with the cropping rule being 0 for values ​​less than 0 and 1 for values ​​greater than 1.

[0063] The weighting coefficients for exceeding the safe water level limit at the Yellow River section, the weighting coefficients for exceeding the warning water level limit at the Yellow River section, and the weighting coefficients for the predicted water level growth rate at the Yellow River section are fixed at 0.45, 0.35, and 0.20, respectively. The Yellow River flood risk index is obtained by linearly weighting and summing the three and is limited to 0 to 1 to ensure that the risk index calculation process is definite and executable.

[0064] Furthermore, the threshold for issuing rainstorm and flood warnings for the Yellow River is fixed at 0.65, with segmented interval mapping as follows: 0.65–0.75 generates a blue warning, 0.75–0.85 generates a yellow warning, 0.85–0.93 generates an orange warning, and greater than or equal to 0.93 generates a red warning. The predicted peak arrival time window is determined based on the time corresponding to the prediction step containing the maximum value of the water level over multiple future steps, and is expanded into a time window output with one prediction step before and after it.

[0065] The system receives the latest measured water level at the same Yellow River node and calculates the deviation with the predicted water level at the same moment in the most recent propagation inference output. When the deviation exceeds the Yellow River early warning calibration error threshold five times consecutively within the rolling window, an online update is triggered. During the online update, a small-step iterative method is used to correct the flood propagation time delay parameters.

[0066] Furthermore, the rolling window length is fixed at 60 minutes and updated in 10-minute increments. The deviation is defined as the absolute value of the latest measured water level minus the predicted water level at the same moment. The Yellow River early warning calibration error threshold is fixed at 0.20m. Online updates are triggered when the deviation exceeds 0.20m for five consecutive times; otherwise, only the deviation is recorded for subsequent statistics.

[0067] Furthermore, the small-step iteration method involves correcting the flood propagation time delay parameters edge by edge along the set of incoming edges of the Yellow River node that triggers the update. Each correction increment is fixed at 0.10 time steps, and the flood propagation time delay parameters are limited to the range of 0 to 36 time steps. The correction direction prioritizes updating the incoming edges that have the highest correlation with the upstream aligned water level within the last 60 minutes, and a maximum of 3 incoming edges are updated in one update cycle to avoid parameter oscillation.

[0068] For each correction, the mean deviation over the most recent 60 minutes is calculated for both candidate scenarios of increasing and decreasing the flood propagation time delay parameter by 0.10 time steps. The candidate direction that results in a smaller mean deviation is selected as the correction direction for this time. The range of 0 to 36 time steps is determined by setting the single propagation inference input time window to the most recent 6 hours and the preset time granularity to 10 minutes.

[0069] The correlation was calculated using the Pearson correlation coefficient. The correlation coefficient was calculated by taking the downstream node deviation sequence and the upstream aligned water level sequence of each ingress edge within the most recent 60 minutes. The ingress edge with the largest absolute value of the correlation coefficient was selected as the priority update object. If they were tied, the ingress edge with the shortest river channel distance was selected.

[0070] It should be noted that this step converts future multi-step water levels into a risk index sequence, using a weighted synthesis of three normalized quantities: safe water level, warning water level, and water level growth rate. A tiered early warning and peak time window output are triggered by three consecutive steps exceeding the threshold. Simultaneously, online updates are triggered by consecutive exceedances of the measured and predicted deviations. Through correlation-based edge selection and small-step correction of flood propagation time lag parameters, a closed loop is formed between early warning threshold determination and time lag self-calibration.

[0071] Example 2, an embodiment of the present invention, provides a Yellow River rainstorm and flood intelligent early warning system, including a multi-source alignment to construct a river network module, a time-delay propagation constraint training module, and a risk early warning feedback update module.

[0072] The multi-source alignment module for constructing the river network is used to align the temporal characteristics of the multi-source data generation nodes of the Yellow River and construct a directed graph of the Yellow River network.

[0073] The time-delay propagation constraint training module is used to propagate reasoning on the directed graph of the Yellow River network according to the flood propagation time delay, and to train the model with conservation constraints and propagation consistency constraints.

[0074] The risk warning feedback update module is used to predict the Yellow River flood risk index and issue graded warnings, and updates the time-delay parameters online based on actual measurement feedback.

Claims

1. A method for intelligent early warning of torrential rain and floods in the Yellow River, characterized in that, include: By aligning the temporal characteristics of the nodes generating multi-source data of the Yellow River, a directed graph of the Yellow River network is constructed. The model is trained by reasoning about the propagation time lag of flood flow on the directed graph of the Yellow River network, with conservation constraints and propagation consistency constraints; the Yellow River flood risk index is predicted and graded early warning is issued, and the time lag parameters are updated online based on actual measurement feedback.

2. The intelligent early warning method for torrential rain and floods in the Yellow River as described in claim 1, characterized in that: The process of generating time-series features of Yellow River multi-source data includes receiving rainfall data from Yellow River rain gauge stations, radar echo statistics, satellite precipitation inversion, soil moisture index, and historical water level sequences from Yellow River stations. The data sources are then unified to a preset time granularity and mapped using Yellow River node numbers. Outliers are constrained by amplitude and abruptly detected and removed. Missing data segments are interpolated to fill in the gaps according to the maximum tolerance duration. Normalization and time window truncation are performed on each feature dimension to form a Yellow River node time-series feature matrix indexed by nodes and ordered by time.

3. The intelligent early warning method for torrential rain and floods in the Yellow River as described in claim 2, characterized in that: The construction of the directed graph of the Yellow River network includes: reading the Yellow River digital river network or connectivity table; determining the node set as Yellow River stations, cross sections, or grid confluence units; determining the edge set as upstream-downstream connectivity with the edge direction pointing from upstream to downstream; recording edge attributes for each edge, including river channel distance, tributary inflow type, and control section identifier; generating a Yellow River network adjacency matrix corresponding one-to-one with the node set and the inflow edge set of each node; and storing the Yellow River network adjacency matrix in association with the Yellow River node temporal feature matrix as the topological constraint input for flood propagation.

4. The intelligent early warning method for torrential rain and floods in the Yellow River as described in claim 3, characterized in that: The propagation inference on the directed graph of the Yellow River network based on the flood propagation time delay includes: calculating the flood propagation time delay parameter for each directed edge of the Yellow River network directed graph; the initial value of the flood propagation time delay parameter is obtained by converting the river channel distance in the edge attribute with the preset average flow velocity, and corrected by combining the historical flood peak arrival time difference; restricting the flood propagation time delay parameter to non-negativity and quantizing it to a preset time granularity; when the flood propagation time delay parameter is not an integer time step, interpolation is used to obtain the alignment state of the upstream node at the current time minus the flood propagation time delay parameter; during propagation inference, the state of the incoming edge node of each node is first aligned in time according to the flood propagation time delay parameter, and then weighted aggregation is performed to obtain the propagation input of the node.

5. The intelligent early warning method for torrential rain and floods in the Yellow River as described in claim 4, characterized in that: The training model based on conservation constraints and propagation consistency constraints includes a Yellow River network spatiotemporal graph prediction model. It takes the temporal feature matrix of Yellow River nodes and the adjacency matrix of the Yellow River network as inputs, and propagates the inference on the directed graph of the Yellow River network according to the flood propagation time lag to obtain the future multi-step water levels of Yellow River nodes and the hidden representation of nodes. Simultaneously, it outputs the equivalent storage and discharge water level change from the Yellow River node's equivalent storage and discharge water level change branch. The prediction error is calculated between the future multi-step water levels of Yellow River nodes output by the propagation inference and the measured water levels at the corresponding times. Conservation constraint terms and propagation consistency constraint terms are also calculated. The conservation constraint terms are obtained by comparing the predicted water level of the same node with the predicted water level of the node's incoming edges through the flood propagation... The weighted summation result after aligning the propagation time delay is obtained by subtracting the equivalent change in water level and squaring the difference. The propagation consistency constraint term is obtained by comparing the difference between the hidden representation of the downstream node and the hidden representation of the upstream node after aligning them with the flood propagation time delay for each directed edge and squaring the norm. The prediction error, conservation constraint term and propagation consistency constraint term are weighted and summed according to the weight coefficients of the Yellow River water level multi-step prediction error, the Yellow River section conservation constraint, and the Yellow River network propagation consistency constraint to form the training objective. The parameters of the Yellow River network spatiotemporal map prediction model and the flood propagation time delay parameters are updated by gradient based on the training objective and iterated until the preset training round is met.

6. The intelligent early warning method for torrential rain and floods in the Yellow River as described in claim 5, characterized in that: The method for predicting the Yellow River flood risk index and issuing graded early warnings includes calculating normalized values ​​for water levels exceeding the Yellow River section's safe water level benchmark, water levels exceeding the Yellow River section's warning water level benchmark, and the predicted water level growth rate based on future multi-step water levels. These normalized values ​​are then combined using the Yellow River section's safe water level exceeding weight coefficient, the Yellow River section's warning water level exceeding weight coefficient, and the Yellow River section's predicted water level growth rate weight coefficient to synthesize the Yellow River flood risk index. When the Yellow River flood risk index exceeds the Yellow River rainstorm and flood warning issuance threshold within three consecutive prediction steps, a blue, yellow, orange, or red warning is generated based on the segmented interval mapping of the Yellow River rainstorm and flood warning issuance threshold, and the corresponding Yellow River node identifier and predicted peak arrival time window are output.

7. The intelligent early warning method for torrential rain and floods in the Yellow River as described in claim 6, characterized in that: The online update of time delay parameters based on measured feedback includes receiving the latest measured water level at the same Yellow River node and calculating the deviation with the corresponding predicted value. When the deviation exceeds the Yellow River early warning calibration error threshold five times consecutively within the rolling window, an online update is triggered. During the online update, a small-step iterative method is used to correct the flood propagation time delay parameters.

8. A Yellow River torrential rain and flood intelligent early warning system, employing the Yellow River torrential rain and flood intelligent early warning method as described in any one of claims 1 to 7, characterized in that: This includes a multi-source alignment module for constructing river networks, a time-delay propagation constraint training module, and a risk warning feedback update module. The multi-source alignment and river network construction module is used to align the temporal characteristics of the multi-source data generation nodes of the Yellow River and construct a directed graph of the Yellow River network; the time-delay propagation constraint training module is used to propagate and reason on the directed graph of the Yellow River network according to the flood propagation time delay, and train the model with conservation constraints and propagation consistency constraints; the risk warning feedback update module is used to predict the Yellow River flood risk index and issue graded warnings, and update the time-delay parameters online according to the measured feedback.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the Yellow River rainstorm and flood intelligent early warning method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the Yellow River rainstorm and flood intelligent early warning method as described in any one of claims 1 to 7.

Images