A method and system for forecasting and early warning of floods in a hilly plain mixed area

Abstract

The application discloses a kind of research hilly plain mixed area flood's forecast early warning method and system, belong to flood early warning technical field, comprising: S1 multi-source data fusion and adaptive preprocessing, S2 light-weight data preprocessing, S3 low-dimensional feature engineering, S4 light-weight dynamic model construction, S5 dynamic early warning threshold generation, S6 rapid early warning and feedback optimization and S7 verification and applicability expansion.The application can input low parameter requirements on the basis of realizing the forecast early warning of flood, with short time consumption, precision and efficiency are considered.

Description

Technical Field

[0001] This invention relates to the field of flood early warning technology, and more specifically, to a method and system for forecasting and early warning of floods in mixed hilly and plain areas. Background Technology

[0002] Floods are natural disasters caused by heavy rains, torrential rains, or continuous rainfall, which inundate and flood low-lying areas, posing a serious threat to human life, property, and the socio-economic development. Mixed hilly and plain areas are a typical topographical feature in my country. For example, the Qinhuai River basin in Nanjing, Jiangsu Province, located in the lower reaches of the Yangtze River, has hilly and mountainous areas accounting for 74% of its area, while low-lying plains account for 26%. Surrounded by mountains on all sides, with a low-lying center, it forms a complete intermontane basin, a typical mixed hilly and plain area. In such areas, the short distance between upstream and downstream sections and the large elevation difference result in short, rapid flood sources, limited flood storage capacity in the upper and middle reaches, rapid flood rise, and frequent flood peaks. Downstream, the floodwaters flow into the Yangtze River, and the flood level is affected by the backwater effect of the downstream Yangtze River floods, thus posing a high risk of flooding.

[0003] In order to effectively prevent and mitigate the impact of floods in such areas, it is necessary to establish scientific and reasonable flood forecasting and early warning methods and systems, provide timely, accurate and reliable flood disaster information, and guide flood prevention and disaster relief work.

[0004] Traditional flood forecasting and early warning methods are mainly based on hydrological and hydrodynamic models. By simulating regional hydrological and hydrodynamic processes, they predict hydrological and hydrodynamic elements such as water level, flow rate, and velocity at various control nodes of rivers, thereby determining the likelihood and severity of floods. However, this method has some limitations, such as difficulty in determining model parameters, high requirements for input data, large computational load, long computation time, and low efficiency.

[0005] Therefore, there is a need for a flood forecasting and early warning method and system that has low requirements for input parameters, short computation time, and balances accuracy and efficiency, and can be applied to mixed hilly and plain areas. Summary of the Invention

[0006] 1. Technical problems to be solved

[0007] To address the problems existing in the prior art, the purpose of this invention is to provide a method and system for forecasting and early warning of floods in mixed hilly and plain areas. In addition to achieving flood forecasting and early warning, this invention also allows for low parameter input requirements, short calculation time, and balances accuracy and efficiency.

[0008] 2. Technical Solution

[0009] To solve the above problems, the present invention adopts the following technical solution:

[0010] A method for forecasting and early warning of floods in mixed hilly and plain areas includes:

[0011] S1. Multi-source data fusion and adaptive preprocessing;

[0012] S2, Lightweight Data Preprocessing;

[0013] S3, Low-dimensional feature engineering;

[0014] S4, lightweight dynamic model construction;

[0015] S5. Dynamic early warning threshold generation;

[0016] S6, Optimized rapid early warning and feedback;

[0017] S7, Validation and Applicability Extension.

[0018] As a preferred embodiment of the present invention, step S1 includes:

[0019] S101, Low-parameter data acquisition: Integrate rainfall data from meteorological stations, river water level monitoring, soil moisture sensor data, and satellite remote sensing data. Eliminate dimensional differences through standardization processing and use KrigingGIS spatial interpolation technology to supplement missing values, reducing dependence on single parameters.

[0020] S102. Terrain Adaptive Modeling: Based on the DEM digital elevation model, the grid is dynamically divided to identify the river network and sub-basin boundaries. The mixed terrain is processed through adaptive resolution technology to optimize computational efficiency.

[0021] As a preferred embodiment of the present invention, step S2 includes:

[0022] S201, Simplified missing value imputation: Use global constants or attribute mean values ​​to imput missing values, avoiding complex regression or decision tree methods;

[0023] S202. Noise Smoothing and Simplification: The binning method is used to replace the computationally intensive regression or clustering noise reduction technique. The binning method is to smooth the data values ​​by examining the range of data values.

[0024] As a preferred embodiment of the present invention, step S3 includes:

[0025] S301. Redundancy Feature Screening: Redundancy attributes are quickly identified by the correlation coefficient of the Pearson product-moment function, rather than by the chi-square test or covariance analysis.

[0026] S302, Attribute Subset Selection: Retain key influencing factors and actively remove weakly correlated parameters using data reduction methods.

[0027] As a preferred embodiment of the present invention, the data reduction method in step S302 includes: data cube aggregation, data attribute subset selection, dimension reduction, numerical reduction, and discretization;

[0028] Data cube aggregation is used for data in a data cube structure. The data cube stores multidimensional aggregation information. Each cell stores an aggregation value, which corresponds to a data point in a multidimensional space. Each attribute may have conceptual hierarchies, allowing data analysis to be performed at multiple levels of abstraction.

[0029] Attribute subset selection is used when the dataset to be analyzed contains a large number of attributes, most of which are irrelevant or redundant to the mining task. Attribute subset selection can detect and remove irrelevant, redundant or weakly related attributes or dimensions. Its goal is to find the smallest set of attributes so that the probability distribution of the data class is as close as possible to the original distribution obtained by using all attributes.

[0030] Dimension reduction uses data encoding or transformation to obtain a reduced or compressed representation of the original data, reducing the number of random variables or attributes considered. If the reduced data can only reconstruct an approximate representation of the original data, then the data reduction is lossy; if it can reconstruct the original data without losing any information, then it is lossless.

[0031] Numerical reduction reduces the amount of data by choosing alternative data representations, replacing or estimating data with smaller data representations;

[0032] Data discretization divides the attribute value range into intervals to reduce the number of given continuous attribute values. The interval labels can replace the actual data values, replacing the numerical values ​​of continuous attributes with a few interval labels.

[0033] As a preferred embodiment of the present invention, step S4 includes:

[0034] S401. Selection of Hybrid Hydrological Model: Combining distributed hydrological models with simplified physical mechanisms, an improved model is adopted. The model is defined as follows: f and B are both reciprocals of soil permeability, and f and B are used to dynamically adjust runoff calculations.

[0035] S402. Machine Learning Model Optimization: Lightweight models such as Random Forest or LSTM networks are used to train time series prediction models with rainfall, water level, and soil moisture as inputs, and redundant parameters are reduced by LASSO regression feature selection.

[0036] As a preferred embodiment of the present invention, step S5 includes:

[0037] S501, Multi-scenario Simulation and Threshold Determination: Based on historical rainfall data, random rainfall scenarios are generated, and the runoff generation and confluence process is simulated by combining a distributed hydrological model. The critical rainfall and water level thresholds for different regions are dynamically calculated. At the same time, the Seq2Seq model is used to predict disaster processes and generate dynamic thresholds by taking the combination of rainfall, water level and rise rate as input.

[0038] S502 Soil-topography coupling analysis: In plains areas, soil saturation and runoff generation are considered, while in mountainous areas, a direct runoff generation mechanism is introduced when rainfall intensity exceeds infiltration rate. The runoff inflow time is corrected by combining vegetation cover and surface evaporation.

[0039] As a preferred embodiment of the present invention, step S6 includes:

[0040] S601. Real-time dynamic early warning release: The early warning results are integrated through the GIS platform and displayed in the form of risk heat maps to classify high, medium and low risk areas. Early warning information is released through multiple channels such as SMS, broadcast and App.

[0041] S602, Model Iteration and Optimization: Utilize cross-validation and online learning mechanisms, and dynamically update model parameters based on real-time monitoring data. Adjust the hyperparameters of the Seq2Seq model through K-fold cross-validation to ensure generalization ability.

[0042] As a preferred embodiment of the present invention, step S7 includes:

[0043] S701, Multi-case validation: Select a typical hilly transition zone, compare historical disaster data with model prediction results, and evaluate the accuracy F1 score > 0.85 and the calculation time per simulation < 10 minutes;

[0044] S702, Modular Expansion Design: The model is encapsulated as an independent module, supporting quick switching between different regional parameters such as permeability coefficient library in plains areas and slope in mountainous areas.

[0045] A forecasting and early warning system for flooding in mixed hilly and plain areas includes: a lightweight data acquisition and fusion module, an adaptive cleaning module, a hybrid model inference core, a spatiotemporal dimension compression module, a database module, a hydrological forecasting module, an early warning application module, a hierarchical early warning decision-making module, a three-dimensional spatial early warning mapping module, and a crowdsourced data feedback channel.

[0046] The lightweight data acquisition and fusion module integrates data from multiple sensor types, including water level stations, rain gauge stations, and meteorological radar. It supports hourly or minute-level data stream access, uses a sliding window mean method to replace complex parameter inputs, and automatically calibrates key thresholds using historical data to reduce manual intervention. The adaptive cleaning module incorporates a dynamic missing value filling strategy and automatically identifies outliers using covariance analysis. The spatiotemporal dimension compression module uses the DBA average sequence algorithm and DTW dynamic time warping technology to perform spatiotemporal alignment of multi-site data, reducing data redundancy. The database module provides sufficient training samples for hydrological forecasting and discriminative data support for risk zoning early warning. The database module includes a historical hydrological monitoring database module and current real-time water conservancy project information. The system includes a database module and a future hypothetical scenario database module. The hydrological forecasting module selects input parameters based on the database module and establishes regional hydrological forecasts through LSTM deep learning. The early warning application module generates graded early warning labels based on the water level values ​​of control nodes predicted by the hydrological forecasting module and compares them with the corresponding warning water levels for decision-making and command reference. The graded early warning decision module evaluates the reliability of the model in real time based on the Nash coefficient, triggering a model self-check when NSE < 0.95. The three-dimensional spatial early warning mapping module is used to combine the DEM digital elevation model to map the predicted water level values ​​to the hill-to-plain transition zone, generating a heat map of inundation depth. The crowdsourced data feedback channel is used to access flood images from social media and uses CV technology to assist in verifying the accuracy of the early warning area.

[0047] The historical hydrological monitoring database module includes historical rainfall and evaporation information from various rain gauge stations within the region, historical water level and flow information from control stations, and tidal information along the Yangtze River at the regional outlet. The current real-time water conservancy project information database module includes information on the length, direction, bottom elevation, bottom width, and slope ratio of the main flood channels within the region, as well as information on the scheduling and operation rules of water conservancy projects such as reservoirs, flood storage areas, polder areas, and sluice gates within the region. The future hypothetical scenario database module, based on historical rainfall data from various rain gauge stations within the region, calculates rainfall at different frequencies and rainfall distribution through frequency sorting. It also calculates tidal patterns at different frequencies using historical Yangtze River tidal data and simulates the water level, flow, and velocity of the main control stations under different future flood and tide combination scenarios based on hydrological and hydrodynamic models, thus forming the future hypothetical scenario database.

[0048] 3. Beneficial effects

[0049] Compared with the prior art, the advantages of this invention are:

[0050] (1) This invention requires only historical water level and rain gauge time series data, avoiding complex parameter calibration, and is suitable for mixed terrain areas with scarce data. Through pattern library and DTW distance matching, it adaptively identifies the sudden flood in hilly areas and the slow rise process in plain areas. The ASCS algorithm reduces the time spent on LSTM parameter optimization by 40%, and combined with the attention mechanism to filter redundant information, the time spent on a single prediction is less than 3 seconds.

[0051] (2) This invention uses a lightweight model with a MAE ≤ 0.017m within a 1-3 hour forecast period to meet the real-time early warning requirements. The Nash coefficient reaches 0.992, which is 15% higher than the traditional model. The false negative rate of extreme events is reduced by 22%. The dynamic time-curved average sequence reduces pattern redundancy and keeps the peak prediction error within ±5%. It supports multi-site data fusion and quickly matches similar flood patterns through symbolic distance hierarchical clustering to achieve basin-level linkage early warning. It takes into account both the short-duration early warning needs of hilly areas and the long-duration scheduling needs of plain areas. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating a method for forecasting and early warning of flooding in mixed hilly and plain areas according to the present invention.

[0053] Figure 2 This is a module diagram of a flood forecasting and early warning system for mixed hilly and plain areas according to the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0055] Example:

[0056] Please see Figure 1 A method for forecasting and early warning of floods in mixed hilly and plain areas includes:

[0057] S1. Multi-source data fusion and adaptive preprocessing;

[0058] S2, Lightweight Data Preprocessing;

[0059] S3, Low-dimensional feature engineering;

[0060] S4, lightweight dynamic model construction;

[0061] S5. Dynamic early warning threshold generation;

[0062] S6, Optimized rapid early warning and feedback;

[0063] S7, Validation and Applicability Extension.

[0064] In a specific embodiment of the present invention, only historical water level and rain gauge time-series data are required, avoiding complex parameter calibration. This makes it suitable for mixed terrain areas with scarce data. Through pattern library and DTW distance matching, it adaptively identifies sudden floods in hilly areas and slow flooding processes in plains. The ASCS algorithm reduces the time spent on LSTM parameter optimization by 40%. Combined with the attention mechanism to filter redundant information, the time for a single prediction is less than 3 seconds. The lightweight model achieves a MAE ≤ 0.017m within a 1-3 hour lead time, meeting the requirements for real-time early warning. The Nash coefficient reaches 0.992, which is 15% higher than the traditional model. The false negative rate of extreme events decreases by 22%. The dynamic time-warped average sequence reduces pattern redundancy, keeping the peak prediction error stable within ±5%. It supports multi-site data fusion and quickly matches similar flood patterns through symbolic distance hierarchical clustering, realizing basin-level linkage early warning. It takes into account both short-duration early warning in hilly areas and long-duration scheduling in plains.

[0065] Specifically, step S1 includes:

[0066] S101, Low-parameter data acquisition: Integrate rainfall data from meteorological stations, river water level monitoring, soil moisture sensor data, and satellite remote sensing data. Eliminate dimensional differences through standardization processing and use KrigingGIS spatial interpolation technology to supplement missing values, reducing dependence on single parameters.

[0067] S102. Terrain Adaptive Modeling: Based on the DEM digital elevation model, the grid is dynamically divided to identify the river network and sub-basin boundaries. Adaptive resolution technology is used to process mixed terrain, such as using coarse-grained grids in plains and fine-grained grids in mountainous areas to optimize computational efficiency.

[0068] Specifically, step S2 includes:

[0069] S201, Simplified missing value imputation: Use global constants or attribute mean values ​​to imput missing values, avoiding complex regression or decision tree methods;

[0070] S202. Noise Smoothing and Simplification: Binning is used to replace computationally intensive regression or clustering noise reduction with local smoothing techniques. Binning smooths the stored data by examining the data values ​​around the data. The stored values ​​are divided into bins. Since only the nearest neighbor values ​​are examined, the binning method performs local smoothing.

[0071] Specifically, step S3 includes:

[0072] S301. Redundancy Feature Screening: Redundant attributes are quickly identified using the correlation coefficient of the Pearson product-moment function, rather than the chi-square test or covariance analysis. The correlation coefficient r and g between numerical attributes A and B, estimated by the Pearson product-moment function, is defined as Equation 1. Equation 1 is:

[0073]

[0074] Where N is the number of tuples (samples), a i and b i These are the values ​​of tuple (sample) i in attributes A and B, respectively. and These are the means of attributes A and B, respectively, and σ A and σ B These are the standard deviations of attributes A and B over the N samples. It is the inner product of the values ​​of attributes A and B in these N samples, -1≤r A,B ≤1, if r A，B A value greater than 0 indicates that the value of attribute A increases as the value of attribute B increases. A，B The larger the value of r, the stronger the correlation, meaning the greater the probability that each attribute implies another attribute. Therefore, a larger r... A，B The value indicates that attribute A or B can be removed as redundant. If r A，B If the value is 0, then A and B are independent and uncorrelated. If the value is less than 0, then A and B are negatively correlated, indicating that each attribute prevents the other from occurring.

[0075] Specifically, the data reduction methods in step S302 include: data cube aggregation, data attribute subset selection, dimension reduction, numerical reduction, and discretization;

[0076] Data cube aggregation is used for data in a data cube structure. The data cube stores multidimensional aggregation information. Each cell stores an aggregation value, which corresponds to a data point in a multidimensional space. Each attribute may have conceptual hierarchies, allowing data analysis to be performed at multiple levels of abstraction.

[0077] When the dataset to be analyzed contains a large number of attributes, most of which are irrelevant or redundant to the mining task, attribute subset selection can detect and remove irrelevant, redundant, or weakly related attributes or dimensions. Its goal is to find the minimum set of attributes that makes the probability distribution of the data classes as close as possible to the original distribution obtained by using all attributes. Its advantage is that it reduces the number of attributes that appear in the discovery pattern, making the pattern easier to understand. For attribute subset selection, exhaustive search to find the best attribute subset may be impractical. Therefore, heuristic algorithms that compress the search space are often used.

[0078] Dimension reduction uses data encoding or transformation to obtain a reduced or compressed representation of the original data, reducing the number of random variables or attributes considered. If the reduced data can only reconstruct an approximate representation of the original data, then the data reduction is lossy; if it can reconstruct the original data without losing any information, then it is lossless.

[0079] Numerical reduction reduces the amount of data by choosing alternative data representations, replacing or estimating data with smaller data representations;

[0080] Data discretization divides the attribute value range into intervals to reduce the number of given continuous attribute values. The interval labels can replace the actual data values. By replacing the values ​​of continuous attributes with a few interval labels, the original data is reduced and simplified, making the data analysis results of unsupervised learning concise, easy to use, and with knowledge-level representation.

[0081] Specifically, step S4 includes:

[0082] S401. Selection of Hybrid Hydrological Model: Combining distributed hydrological models with simplified physical mechanisms, an improved model is adopted. The model is defined as follows: f and B are both reciprocals of soil permeability, and f and B are used to dynamically adjust runoff calculations.

[0083] S402. Machine Learning Model Optimization: Lightweight models such as Random Forest or LSTM networks are used to train time series prediction models with rainfall, water level, and soil moisture as inputs. LASSO regression feature selection is used to reduce redundant parameters and improve computation speed.

[0084] Specifically, step S5 includes:

[0085] S501, Multi-scenario Simulation and Threshold Determination: Based on historical rainfall data, random rainfall scenarios are generated, and the runoff generation and confluence process is simulated by combining a distributed hydrological model. The critical rainfall and water level thresholds for different regions are dynamically calculated. At the same time, the Seq2Seq model is used to predict disaster processes and generate dynamic thresholds by taking the combination of rainfall, water level and rise rate as input.

[0086] S502 Soil-topography coupling analysis: In plains areas, soil saturation and runoff generation are considered, while in mountainous areas, a direct runoff generation mechanism is introduced when rainfall intensity exceeds infiltration rate. The runoff inflow time is corrected by combining vegetation cover and surface evaporation.

[0087] Specifically, step S6 includes:

[0088] S601. Real-time dynamic early warning release: The early warning results are integrated through the GIS platform and displayed in the form of risk heat maps to classify high, medium and low risk areas. Early warning information is released through multiple channels such as SMS, broadcast and App to shorten the response time.

[0089] S602, Model Iteration and Optimization: Utilize cross-validation and online learning mechanisms, and dynamically update model parameters based on real-time monitoring data. Adjust the hyperparameters of the Seq2Seq model through K-fold cross-validation to ensure generalization ability.

[0090] Specifically, step S7 includes:

[0091] S701, Multi-case validation: Select a typical hilly transition zone, compare historical disaster data with model prediction results, and evaluate the accuracy F1 score > 0.85 and the calculation time per simulation < 10 minutes;

[0092] S702, Modular Expansion Design: The model is encapsulated as an independent module, supporting rapid switching between different regional parameters such as permeability coefficient library in plains areas and slope in mountainous areas, thereby improving the universality of the method.

[0093] Please see Figure 2 A forecasting and early warning system for flooding in mixed hilly and plain areas includes:

[0094] The system includes a lightweight data acquisition and fusion module, an adaptive cleaning module, a hybrid model inference core, a spatiotemporal dimension compression module, a database module, a hydrological forecasting module, an early warning application module, a graded early warning decision-making module, a three-dimensional spatial early warning mapping module, and a crowdsourced data feedback channel.

[0095] The lightweight data acquisition and fusion module integrates data from multiple types of sensors, such as water level stations, rain gauge stations, and weather radar. It supports hourly or minute-level data stream access, uses the sliding window mean method to replace complex parameter input, and automatically calibrates key thresholds through historical data to reduce manual intervention.

[0096] The adaptive cleaning module has a built-in dynamic missing value filling strategy and automatically identifies outliers by combining covariance analysis.

[0097] The spatiotemporal dimension compression module uses the DBA average sequence algorithm and DTW dynamic time warping technology to perform spatiotemporal alignment on multi-site data, reducing data redundancy;

[0098] The database module is used to provide sufficient training samples for hydrological forecasting and to provide discriminative data support for risk zoning early warning. The database module includes a historical hydrological monitoring database module, a current real-time water conservancy project information database module, and a future hypothetical scenario database module.

[0099] The hydrological forecasting module selects input parameters based on the database module and establishes regional hydrological forecasts through LSTM deep learning.

[0100] The early warning application module generates graded early warning labels based on the water level values ​​of control nodes predicted by the hydrological forecasting module and the corresponding warning water levels, for reference in decision-making and command.

[0101] For example, the warning water level at Dongshan Station in the Qinhuai River Basin is 8.8m. When the forecast water level is in the range of 8.8 to 9.5m, a blue warning is issued; when the forecast water level is in the range of 9.5 to 10.5m, a yellow warning is issued; when the forecast water level is in the range of 10.5 to 11.5m, an orange warning is issued; and when the forecast water level exceeds 1.5m, a red warning is issued.

[0102] The graded early warning decision module evaluates the model reliability in real time based on the Nash coefficient, and triggers model self-check when NSE < 0.95;

[0103] The 3D spatial early warning mapping module is used to combine the DEM digital elevation model to map the predicted water level to the transition zone from hills to plains, generating a heat map of inundation depth.

[0104] The crowdsourced data feedback channel is used to access flood images from social media, and computer vision (CV) technology is used to help verify the accuracy of the early warning areas.

[0105] The historical hydrological monitoring database module includes historical rainfall and evaporation information from various rain gauge stations within the region, historical water level and flow information from control stations, and tidal information along the river at the regional outlet.

[0106] The current real-time water conservancy project information database module includes information on the length, direction, bottom elevation, bottom width, and slope ratio of the main flood channels in the region, as well as information on the scheduling and operation rules of water conservancy projects such as reservoirs, flood storage and detention areas, polder areas, and sluice gates in the region;

[0107] The future hypothetical scenario database module uses historical rainfall data from various rain gauge stations within the region to calculate rainfall and distribution at different frequencies, such as 5-year, 10-year, 20-year, 50-year, and 100-year return periods. It also uses historical Yangtze River tide data to calculate tide patterns at different frequencies, such as 2-year, 5-year, 10-year, 20-year, and 50-year return periods. Finally, it uses hydrological and hydrodynamic models to simulate and calculate information such as water level, flow rate, and velocity at key control stations under different future flood and tide combination scenarios, thus forming the future hypothetical scenario database.

[0108] The usage of the hydrological forecasting module includes:

[0109] The input parameters are selected based on the current water level and flow rate within the region, as well as the rainfall in the previous n hours and the next n hours, where n is the forecast period;

[0110] After preparing the input data, the data is normalized and divided into training and test sets. 80% of the data is used as the training set and 20% of the data is used as the test set.

[0111] The training data is input into the LSTM model, and the parameters are adjusted according to the model characteristics to achieve the optimal model. The ReLU activation function is selected.

[0112] After the model training is completed, the model is used to predict the test samples and then inversely normalized. The results are then compared with the actual results to complete the model validation.

[0113] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. A method for forecasting and early warning of floods in a hilly plain mixed area, characterized in that, The method comprises the following steps: S1, multi-source data fusion and adaptive preprocessing; S1 comprises the following steps: S101, low-parameter data acquisition: integrate meteorological station rainfall data, river water level monitoring, soil moisture sensor and satellite remote sensing data, eliminate dimensional differences through standardization processing, supplement missing values by using Kriging GIS spatial interpolation technology, and reduce dependence on a single parameter; S102, terrain adaptive modeling: dynamically divide grids based on DEM digital elevation model, identify river network and sub-basin boundary, process mixed terrain through adaptive resolution technology, and optimize calculation efficiency; S2, lightweight data preprocessing; S2 comprises the following steps: S201, missing value simplified filling: use global constant or attribute mean value filling to avoid complex regression or decision tree method; S202, noise smoothing simplification: replace the regression or clustering denoising with large calculation amount by using the local smoothing technology of the binning method, wherein the binning method is to smooth the stored data value by investigating the data value of the range of the data; S3, low-dimensional feature engineering; S3 comprises the following steps: S301, redundant feature screening: quickly identify redundant attributes by using the correlation coefficient of Pearson product moment function instead of chi-square test or covariance analysis; S302, attribute subset selection: retain key influencing factors and actively remove weakly correlated parameters by using data reduction method; S4, lightweight dynamic model construction; S4 comprises the following steps: S401、Hydrological model selection: combined with distributed hydrological model and simplified physical mechanism, improved model is adopted, wherein, and are reciprocal of soil permeability; the and are used for dynamic adjustment of runoff calculation; S402, machine learning model optimization: use a random forest or LSTM network lightweight model to take rainfall, water level and soil moisture as input, train a time series prediction model, and reduce redundant parameters through LASSO regression feature selection; S5, dynamic early warning threshold generation; S5 comprises the following steps: S501, multi-scenario simulation and threshold setting: generate random rainfall scenarios based on historical rainfall data, simulate the runoff and confluence process by combining a distributed hydrological model, dynamically calculate the critical rainfall and water level thresholds of different regions, and input the rainfall, water level and rise speed combination into a Seq2Seq model to predict the disaster process and generate dynamic thresholds; S502, soil-terrain coupling analysis: consider soil storage runoff in plain areas, introduce direct runoff mechanism when the rainfall intensity exceeds the infiltration rate in mountainous areas, and correct the runoff inflow time by combining vegetation coverage and surface evaporation; S6, rapid early warning and feedback optimization; S6 comprises the following steps: S601, real-time dynamic early warning release: release early warning information through a GIS platform integrated early warning results, display high, medium and low risk areas in the form of risk heat map, and release early warning information through multiple channels such as short message, broadcast and App; S602, model iteration optimization: dynamically update model parameters by using cross-validation and online learning mechanism in combination with real-time monitoring data, adjust the hyperparameters of the Seq2Seq model through K-fold cross-validation, and ensure the generalization ability; S7, verification and applicability expansion. 2.The method of claim 1, wherein, The data reduction method in S302 comprises data cube aggregation, data attribute subset selection, dimension reduction, numerical reduction and discretization. Data cube aggregation is used for data in data cube structure, data cube stores multi-dimensional aggregation information, each cell stores an aggregation value, corresponding to a number point in multi-dimensional space, each attribute may exist conceptual hierarchy, allowing data analysis at multiple abstraction layers; Attribute subset selection can detect and remove irrelevant, redundant or weakly related attributes or dimensions when the data set to be analyzed contains a large number of attributes, most of which are irrelevant or redundant to the mining task, the goal of attribute subset selection is to find the smallest attribute set, so that the probability distribution of the data class is as close as possible to the original distribution obtained using all attributes, for attribute subset selection; Dimension reduction uses data encoding or transformation to obtain original data reduction or compressed representation, reduces the number of random variables or attributes considered, if the reduced data can only reconstruct the approximate representation of the original data, then the data reduction is lossy, if the original data can be constructed without losing any information, it is lossless; Numerical reduction reduces the amount of data by selecting alternative data representation, replacing or estimating data with smaller data representation; Data discretization divides the attribute value domain into intervals to reduce the number of given continuous attribute values, the interval label can replace the actual data value, and a small number of interval labels replace the numerical value of continuous attributes.

3. The method of claim 2, wherein the method is characterized by, The step S7 comprises: S701, multi-case verification: select typical hilly transition zone, compare historical disaster data and model prediction results, evaluate precision F1 score> 0.85 and calculate time-consuming single simulation< 10 minutes; S702, modular extension design: encapsulate the model as an independent module, support fast switching of different regional parameter plain area permeability coefficient library and mountain area slope.

4. A forecasting and early warning system for flood in a hilly plain mixed region, applied to the forecasting and early warning method for flood in a hilly plain mixed region according to any one of claims 1-3, characterized in that, Comprise: Lightweight data acquisition and fusion module, adaptive cleaning module, hybrid model inference core, space-time dimension compression module, database module, hydrological forecasting module, early warning application module, hierarchical early warning decision module, three-dimensional space early warning mapping module and crowdsourcing data feedback channel; The lightweight data acquisition fusion module integrates water level stations, rainfall stations, and weather radar multi-type sensor data, supports hour-level or minute-level data stream access, uses a sliding window mean method instead of complex parameter input, automatically calibrates key thresholds through historical data, reduces manual intervention, the adaptive cleaning module has a built-in missing value dynamic filling strategy, automatically identifies outliers in combination with covariance analysis, the spatiotemporal dimension compression module uses a DBA average sequence algorithm DTW dynamic time warping technology to spatiotemporally align multi-site data, reduce data redundancy, the database module is used to provide sufficient training samples for hydrological forecasting and discriminant data support for risk zoning warning, the database module includes a historical hydrological monitoring database module, a current real-time water conservancy engineering information database module, and a future hypothetical scenario database module, the hydrological forecasting module selects input parameters based on the database module, establishes regional hydrological forecasting through LSTM deep learning, the warning application module compares the water level value of the control node predicted by the hydrological forecasting module with the corresponding warning water level to form a hierarchical warning label for decision-making reference, the hierarchical warning decision module assesses the reliability of the model in real time based on the Nash coefficient, and triggers model self-checking when NSE<0.95, the three-dimensional space warning mapping module is used to combine DEM digital elevation model to map the water level prediction value to the hill-to-plain transition zone to generate a submergence depth heat map, and the crowdsourcing data feedback channel is used to access social media flood pictures and verify the accuracy of the warning area through CV technology; The historical hydrological monitoring database module includes historical rainfall and evaporation information of each rainfall station in the region, historical water level and flow information of the control station, and tidal level information along the river at the regional outlet, the current real-time water conservancy engineering information database module includes length, direction, bottom elevation, bottom width, and slope ratio information of the main flood drainage river in the region, and reservoir, flood detention area, polder area, and sluice station water conservancy engineering operation rule information in the region, the future hypothetical scenario database module is based on historical rainfall of each rainfall station in the region, obtains rainfall and rainfall distribution under different frequencies through frequency calculation, obtains tide types under different frequencies through frequency calculation based on historical Yangtze River tidal level data, and obtains water level, flow, and flow velocity information of the main control station under different future flood tide combination scenarios through hydrological and hydrodynamic model simulation calculation, thereby forming a future hypothetical scenario database.

Images