Runoff Simulation Method in Plain River Network Area

Abstract

The present invention discloses a runoff simulation method for plain water network areas, comprising: obtaining multi-source data and preprocessing it to obtain a spatio-temporally consistent hydrological dataset; based on the spatio-temporally consistent hydrological dataset, using an enhanced polder system dynamic model to obtain polder water level data and exchange flow data; using an improved XAJ model for runoff calculation in the plain area; constructing a multi-scale river network topological structure and performing multi-directional flow pattern identification and prediction with deep spatio-temporal fusion; constructing a time-varying structure adaptive BP neural network model to realize runoff simulation in plain water network areas. The present invention can effectively solve technical problems such as reverse flow phenomenon identification in plain water network areas, simulation of water volume exchange threshold effect in polders, optimization of dynamic distribution coefficients at diversion nodes, and correction of the influence of temporary mobile pumping stations.

Description

Technical Field

[0001] The invention relates to hydrological simulation technology, in particular to a runoff simulation method in a plain water network area. Background Art

[0002] Plain water network areas occupy an important position in water resource management and flood prevention and disaster reduction due to their special terrain conditions and complex river network systems. Plain areas are mostly economically developed areas, and accurate and efficient runoff simulation is of great significance for water resource scheduling, flood disaster prevention and control, and ecological and environmental protection in these areas. Therefore, establishing a runoff simulation method that can accurately capture the hydrological characteristics of plain water network areas has become a key technical issue that needs to be urgently solved in the current hydrology discipline.

[0003] At present, the runoff simulation in plain water network areas mainly adopts the method of combining hydrological models with hydraulic models. Common ones include runoff generation simulation based on the Xin'anjiang model (XAJ), data-driven methods of neural network models (ANN / BP), and comprehensive simulation methods that couple one-dimensional hydrodynamic models with distributed hydrological models. These models can achieve runoff simulation in plain areas under certain conditions through parameter calibration and verification. For example, Xu et al. (2017) combined the improved XAJ model with the BP neural network and applied it to the study of the Taihu Lake Basin; Wang et al. (2019) developed a two-dimensional hydrological and hydrodynamic coupling model that takes into account the underlying surface characteristics of the plain area; Zhang et al. (2020) proposed a plain river network flow pattern identification framework based on data mining.

[0004] However, the existing models still have obvious limitations when facing special hydrological phenomena in plain water network areas. First, the identification and quantification mechanism of the countercurrent phenomenon is insufficient. The existing random forest model can only simply divide the river flow into normal flow and countercurrent flow, and cannot accurately capture the intensity, duration and spatial distribution characteristics of the countercurrent. Secondly, the nonlinear threshold effect of water exchange in the embankment system is not accurately simulated. When the water level or rainfall reaches a specific threshold, the sudden change characteristics of the water exchange pattern in the embankment area are difficult to be accurately expressed by the existing continuous smooth BP neural network. In addition, the dynamic allocation coefficient at the diversion node is not optimized enough, and the fixed allocation ratio is difficult to reflect the dynamic characteristics of the diversion coefficient changing with hydrological conditions, especially under extreme hydrological events. The simulation accuracy is significantly reduced. Finally, there is a lack of real-time correction technology for the impact of temporary mobile pumping stations, and it is impossible to quickly identify and correct the highly dynamic characteristics of temporary facilities under emergency conditions. These technical problems seriously restrict the accuracy and applicability of runoff simulation in plain water network areas. Summary of the invention

[0005] The purpose of the invention is to provide a method for simulating runoff in a plain water network area, in order to solve one of the problems of the prior art.

[0006] Technical solution: A runoff simulation method for plain water network areas, including:

[0007] Obtain multi-source data and preprocess it to obtain a spatio-temporally consistent hydrological dataset;

[0008] Based on the spatio-temporally consistent hydrological dataset, use an enhanced polder system dynamic model to obtain polder water level data and exchange flow data, and based on this, perform runoff generation calculation in the plain area through an improved XAJ model to generate improved XAJ runoff generation data;

[0009] Based on the spatio-temporally consistent hydrological dataset, construct a multi-scale river network topological structure to obtain river network topological skeleton data and river network connectivity characteristic matrix, and accordingly perform multi-directional flow state identification and prediction with deep spatio-temporal fusion;

[0010] Obtain the flow state prediction result and flow state intensity dataset, and combine it with the improved XAJ runoff generation data to construct a time-varying structure adaptive BP neural network model to achieve runoff simulation in the plain water network area.

[0011] According to one aspect of the present application, using an enhanced polder system dynamic model to obtain polder water level data and exchange flow data includes:

[0012] Read historical water level data and polder operation records from the spatio-temporally consistent hydrological dataset, identify key water level threshold points, and establish a set of key threshold points;

[0013] According to the set of key threshold points, construct a multi-threshold piecewise function to form a polder water volume exchange model;

[0014] According to the real-time water situation changes and historical simulation errors, adaptively adjust the parameters of the set of key threshold points to generate a dynamic threshold parameter set; and accordingly update the polder water volume exchange model;

[0015] Use the updated polder water volume exchange model, combine rainfall input, evaporation loss and infiltration amount, calculate the water volume balance in the polder area to obtain polder water level data and exchange flow data.

[0016] According to one aspect of the present application, constructing a multi-threshold piecewise function to form a polder water volume exchange model includes:

[0017] Based on the set of key threshold points, divide the water level difference into multiple intervals;

[0018] Perform non-linear regression analysis on the historical water level data and water volume exchange observation data in each interval to obtain the exchange coefficient, non-linear index and basic exchange volume in each interval, and form an interval parameter set;

[0019] Extract pump station operation data from the polder operation records, analyze the pump station pumping characteristics under different water level conditions, and establish a set of pump station characteristic parameters;

[0020] Combined with the set of key threshold points, the set of interval parameters, and the set of pumping station characteristic parameters, a multi-threshold piecewise function model for expressing the water volume exchange process of the polder is constructed, and different water volume exchange calculation formulas are adopted in different water level difference intervals.

[0021] According to one aspect of the present application, based on the set of key threshold points, the water level difference is divided into multiple intervals, including:

[0022] Read historical water level data and water volume exchange observation data, group the water volume exchange data according to the internal and external water level difference, and use the change point detection algorithm to identify the threshold points where the water volume exchange behavior changes significantly, and obtain the initial threshold point set;

[0023] Match and analyze the internal and external water level differences corresponding to the sluice opening and pump starting operation time points in the polder operation record with the initial threshold point set, and identify the key threshold points with clear physical meanings to form the key threshold point set;

[0024] Based on the key threshold point set, divide the water level difference interval, and establish a piecewise function for the water volume exchange behavior in each interval.

[0025] According to one aspect of the present application, construct a multi-scale river network topological structure to obtain river network topological skeleton data and river network connectivity characteristic matrix, including:

[0026] Extract river network data from the spatio-temporally consistent hydrological dataset, and construct a multi-scale river network topological model including the main river network, secondary river network, and capillary river network;

[0027] Perform skeletonization processing on the multi-scale river network topological model, identify confluence points, divergence points, and loop network structures, and generate river network topological skeleton data including the river network center line, key node positions, and river section connection relationships;

[0028] Convert the river network topological skeleton data into a directed graph structure to form a river network graph data structure;

[0029] Combine historical water level and flow data and river channel physical characteristic data to construct a node characteristic matrix and an edge characteristic matrix;

[0030] Use the dual-channel graph convolutional network model to process the river network graph data structure, extract the connectivity characteristics between nodes, and generate a river network connectivity characteristic matrix.

[0031] According to one aspect of the present application, use the dual-channel graph convolutional network model to process the river network graph data structure, extract the connectivity characteristics between nodes, and generate a river network connectivity characteristic matrix, including:

[0032] Use the trained river network GCN model to perform forward calculation on the river network graph data structure, extract the embedding similarity between nodes, and construct a connectivity strength matrix;

[0033] Based on the connectivity strength matrix, calculate the maximum flow transmission capacity between any two points in the river network, analyze the main flow propagation paths, and obtain the flow propagation path set;

[0034] Calculate the node centrality index, identify the key connected nodes and potential vulnerable points in the river network, and generate the connectivity vulnerability point set;

[0035] Integrate the connectivity strength matrix, the flow propagation path set, and the connectivity vulnerability point set to form the river network connectivity characteristic matrix.

[0036] According to one aspect of the present application, perform deep spatio-temporal fusion for multi-directional flow pattern identification and prediction to obtain the flow pattern prediction result and the flow pattern intensity data set, including:

[0037] Extract historical water level and flow data, water level gradient data, and flow velocity data from the spatio-temporally consistent hydrological data set, calculate the key characteristic parameters, and form the flow pattern characteristic parameter set;

[0038] Based on the flow pattern characteristic parameter set, establish a multi-flow pattern classification standard including five flow patterns: normal flow, weak countercurrent, strong countercurrent, split flow, and circulation flow; and mark the historical water level and flow data accordingly to generate the flow pattern marked historical data set;

[0039] Through training a deep learning model integrating spatio-temporal attention mechanism based on the flow pattern marked historical data set, identify and predict complex flow patterns to obtain the flow pattern prediction result;

[0040] According to the flow pattern intensity quantization index, combine the water level gradient data and the flow velocity data to quantify the intensity characteristics of different flow patterns and generate the flow pattern intensity data set.

[0041] According to one aspect of the present application, the multi-flow pattern classification standard including five flow patterns: normal flow, weak countercurrent, strong countercurrent, split flow, and circulation flow includes:

[0042] Perform unsupervised clustering analysis on the flow pattern characteristic parameter set to obtain the preliminary flow pattern clustering result;

[0043] Interpret and adjust the preliminary flow pattern clustering result, determine the characteristic boundaries of the five flow patterns, and generate the characteristic parameter threshold table;

[0044] Analyze the spatio-temporal distribution characteristics of different flow patterns, including time distribution, spatial distribution, and correlation factors, to form the flow pattern spatio-temporal distribution characteristic report;

[0045] Based on the characteristic parameter threshold table and the flow pattern spatio-temporal distribution characteristic report, construct a multi-layer decision tree structure to form the flow pattern classification decision model;

[0046] Mark the historical data using the flow regime classification decision model, and verify it through typical hydrological event cases, fine-tune the threshold table of feature parameters, and optimize the multi-flow regime classification criteria.

[0047] According to one aspect of the present application, the process of realizing runoff simulation in the plain water network area includes:

[0048] Based on historical flow data and meteorological data, use time series decomposition and change point detection algorithms to automatically identify wet seasons, normal seasons, and dry seasons, and generate hydrological period segmentation marked data;

[0049] Input the improved XAJ runoff data, flow regime prediction results, and flow regime intensity data set into the constructed multi-scale time feature extraction network to realize the accurate simulation of the runoff process in the plain water network area.

[0050] According to one aspect of the present application, the multi-scale time feature extraction network includes:

[0051] Short-term memory units based on long short-term memory networks, used to capture hydrological changes on the hourly to daily scale;

[0052] Medium-term memory units based on time convolutional networks, used to capture hydrological changes on the daily to weekly scale;

[0053] Long-term memory units based on attention mechanisms, used to capture hydrological changes on the weekly to monthly scale;

[0054] An adaptive network structure, using a neural architecture search algorithm to automatically adjust the network structure parameters according to the hydrological period segmentation marked data.

[0055] According to one aspect of the present application, obtaining multi-source data and preprocessing to obtain a spatio-temporally consistent hydrological data set includes:

[0056] Collect rainfall data, evaporation data, water level and flow data, terrain data, remote sensing images, and water conservancy project facility information in the plain water network area to form an original data set;

[0057] Deploy pumping station monitoring sensors based on Internet of Things technology to collect temporary pumping station location and operation status data;

[0058] Combine the crowdsourced data reported by the management department and the remote sensing image change detection results to identify the distribution locations of temporary mobile pumping stations and generate temporary pumping station distribution data;

[0059] Analyze the specification parameters and real-time operation status of temporary pumping stations, calculate the pumping and drainage flow capacity, and form pumping and drainage capacity parameters;

[0060] Perform spatio-temporal consistency correction on all the obtained data, and use the spatio-temporal tensor completion algorithm to repair missing data to generate a spatio-temporally consistent hydrological data set.

[0061] According to one aspect of the present application, it further includes:

[0062] Based on the river network connectivity characteristic matrix and the temporary pumping station distribution data, calculate the influence range of the temporary pumping stations in the river network to form a pumping station influence coefficient matrix;

[0063] According to the pumping capacity parameters and the pumping station influence coefficient matrix, calculate the spatio-temporal distribution of the pumping flow of the temporary pumping stations to generate a pumping influence flow distribution map;

[0064] Construct a prediction result correction function, fuse the flow state prediction result and the pumping influence flow distribution map, realize the real-time correction of the influence of the temporary mobile pumping stations, and obtain a corrected prediction result.

[0065] According to one aspect of the present application, constructing a prediction result correction function, fusing the flow state prediction result and the pumping influence flow distribution map, and realizing the real-time correction of the influence of the temporary mobile pumping stations includes:

[0066] Analyze the operation mode of the temporary pumping stations, predict the possible operation states in the short term in the future, and obtain the pumping station operation mode prediction result;

[0067] Establish a response characteristic model of the pumping of the temporary pumping stations to the hydrological process, learn the spatial response characteristics and the temporal response characteristics to form a pumping station influence response model;

[0068] Construct a prediction result correction function based on the pumping station influence response model to realize the correction of the predicted flow;

[0069] Design a multi-step prediction deviation progressive correction strategy, apply different intensities of correction to different prediction durations, and generate a corrected prediction result.

[0070] Beneficial effects: It can effectively solve technical problems such as the identification of the countercurrent phenomenon in the plain water network area, the simulation of the threshold effect of the water volume exchange in the polder area, the optimization of the dynamic distribution coefficient of the diversion nodes, and the correction of the influence of the temporary mobile pumping stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 is a flowchart of the present invention.

[0072] Figure 2 is a flowchart of step S1 of the present invention.

[0073] Figure 3 is a flowchart of step S2 of the present invention.

[0074] Figure 4 is a flowchart of step S3 of the present invention.

[0075] Figure 5 is a flowchart of step S4 of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0076] As Figure 1 shown, a runoff simulation method for plain water network areas includes:

[0077] Obtain multi-source data and preprocess it to obtain a spatio-temporally consistent hydrological dataset;

[0078] Based on the spatio-temporally consistent hydrological dataset, use an enhanced polder system dynamic model to obtain polder water level data and exchange flow data; and based on it, perform runoff generation calculation for the plain area of the improved XAJ model to generate improved XAJ runoff generation data;

[0079] Based on the spatio-temporally consistent hydrological dataset, construct a multi-scale river network topological structure to obtain river network topological skeleton data and river network connectivity characteristic matrix; and based on this, perform multi-directional flow pattern recognition and prediction of deep spatio-temporal fusion to obtain flow pattern prediction results and flow pattern intensity datasets;

[0080] Based on the improved XAJ runoff generation data, flow pattern prediction results and flow pattern intensity datasets, construct a time-varying structure adaptive BP neural network model to realize runoff simulation in plain water network areas.

[0081] Specifically, the processing process of a runoff simulation method for plain water network areas is as follows, including the following steps:

[0082] S1. Multi-source data acquisition and preprocessing

[0083] Obtain rainfall data, evaporation data, water level and flow data, terrain data, remote sensing images, water conservancy project facility information and temporary pump station deployment data to form an original dataset.

[0084] Use the original rainfall data and terrain elevation data, and adopt terrain-weighted inverse distance interpolation method (TWIDW) for spatial distribution correction to generate corrected rainfall field data and improve the accuracy of rainfall spatial distribution.

[0085] Through multi-temporal remote sensing images, use the water body index NDWI and deep learning segmentation model to extract water body coverage data and temporal change information to realize dynamic water body monitoring.

[0086] Based on the water conservancy project facility information, construct digital twin models of facilities such as sluices and pump stations, record their locations, scales, operation rules and real-time status, generate an engineering facility digital twin database, and provide support for subsequent simulations.

[0087] Construct a monitoring scheme for temporary mobile pump stations based on Internet of Things technology and crowdsourced data, combine the reports from management departments and remote sensing image change detection, obtain temporary pump station distribution data and pumping and drainage capacity parameters, and solve the monitoring problem affected by temporary mobile pump stations.

[0088] Perform spatiotemporal consistency correction on all acquired data, use the spatiotemporal tensor completion algorithm to repair missing data, and generate a spatiotemporally consistent hydrological dataset to ensure the integrity and consistency of the data.

[0089] S2. Enhanced Polder System Dynamic Simulation

[0090] According to the polder location data, scale information, and management methods, divide the polder system into three categories: urban flood control polders, agricultural production polders, and ecological protection polders, and establish a polder type database and a parametric characterization model.

[0091] Based on historical water level data and polder operation records, identify the key water level thresholds and response times for each type of polder, construct a threshold response function library to describe the behaviors of the polder system such as opening gates and starting pumps under different water level conditions.

[0092] Design a multi-threshold piecewise function to express the water volume exchange process of the polder, forming an enhanced polder water volume exchange model. This model contains multiple critical threshold points, and different water volume exchange formulas are used in different threshold intervals:

[0093] Q exchange = {0,H out < H T1 ;

[0094] α1(H out - H in ) β1 ,H T1 ≤ H out < H T2 ;

[0095] α2(H out - H in ) β2 + Q pump1 ,H T2 ≤ H out < H T3 ;

[0096] α3(H out - H in ) β3 + Q pump2 ,H out ≥ H T3}。

[0097] Where Q exchange is the water volume exchange between the inside and outside of the polder, H out and H in are the water levels outside and inside the polder respectively, H T1 , H T2 and H T3 are water level thresholds, α and β are exchange coefficients, Qpump It is the pumping drainage volume of the pumping station.

[0098] Using the water volume exchange model of the polder, combined with rainfall input, evaporation loss and seepage volume, the water volume balance of the polder is calculated in real time, and the water level data and exchange flow data of the polder are generated.

[0099] According to the historical simulation error and the real-time water situation changes, an adaptive adjustment mechanism is designed to optimize the threshold parameters of the polder in real time, generate a dynamic threshold parameter set, and improve the adaptability of the model to extreme hydrological events.

[0100] S3. Multi-scale river network topological structure construction

[0101] Based on high-resolution remote sensing images and topographic data, the river network is extracted by using the deep learning semantic segmentation method to generate fine river network vector data.

[0102] The fine river network vector data is skeletonized to identify confluence points, divergence points and loop network structures, and river network topological skeleton data and key node libraries are generated.

[0103] The river network topological skeleton data is decomposed at multiple scales by using graph theory algorithms to construct a multi-scale river network topological model including main river networks, secondary river networks and capillary river networks, and the water flow paths at different scales are expressed.

[0104] The multi-scale river network topological model is transformed into a graph structure, and the graph convolutional network (GCN) is used to analyze the connectivity and flow propagation characteristics of the river network to generate a river network connectivity feature matrix.

[0105] Based on the graph theory cycle detection algorithm, loop network structures are identified from the river network topological skeleton data, the geometric and topological features of the loop network are extracted, and a loop network feature library is established to provide support for subsequent complex flow state simulations.

[0106] S4. Runoff generation calculation in plain areas for improving the XAJ model

[0107] Combined with remote sensing images and land use data, deep learning methods are used to identify the impervious area and its changes, establish a dynamic model of the impervious area, and improve the runoff generation calculation of the XAJ model.

[0108] Using groundwater monitoring data, a feedback relationship between the groundwater level and the soil moisture content in the lower layer is established, and a groundwater level feedback module is developed to enhance the response ability of the XAJ model to the changes in the groundwater level in plain areas.

[0109] According to the soil profile data, the three-layer soil structure in the XAJ model is corrected, and a continuous function is used to describe the vertical distribution of the soil moisture content to form an improved soil moisture model and improve the accuracy of runoff generation calculation.

[0110] Based on the water body coverage data, the study area is divided into water bodies and land, and the water surface evaporation and land surface evapotranspiration are calculated respectively to generate differentiated evaporation data, solving the problem of inaccurate evaporation calculation in the plain water network area.

[0111] Integrate the impervious area dynamic model, groundwater level feedback module, improved soil moisture model and differentiated evaporation data to calculate the runoff yield of each subunit in the plain water network area and generate improved XAJ runoff yield data.

[0112] Introduce the dynamic change mechanism of impervious area, which significantly improves its applicability in urbanized areas. This improvement dynamically identifies the spatio-temporal changes of impervious area through remote sensing images and deep learning methods, designs the impervious area parameter (IMP) as a function of time instead of a constant in traditional models, and directly calculates the direct runoff yield (R imp = P·IMP(t)), and then adds it to the runoff yield (R S ) of the pervious area to obtain the total runoff yield. It solves the problem that the traditional XAJ model cannot accurately express the impact of underlying surface changes in the process of urbanization. Especially on seasonal and interannual scales, it can capture the effects of increased runoff yield and reduced infiltration caused by the growth of impervious area, significantly improving the runoff yield calculation accuracy of the model in the rapidly urbanizing plain water network area.

[0113] The XAJ model specifically includes:

[0114] The impervious area dynamic identification unit is used to: receive multi-temporal remote sensing images and land use data, extract the impervious area distribution through the U-Net deep learning network, calculate its temporal evolution trend, and output the dynamic impervious area parameter IMP(t), which directly affects the rapid response part in runoff yield calculation.

[0115] The soil water content vertical distribution correction unit is used to: receive soil profile sampling data, use the continuous function W(z) to replace the original three-layer discrete structure of XAJ, calculate the actual soil water content of each layer through integration, and output the corrected water content parameters of each layer, affecting the full storage runoff generation process in runoff yield calculation.

[0116] The groundwater level feedback mechanism unit is used to: receive groundwater monitoring well data, establish the feedback relationship between groundwater level and the water content of the lower layer of soil, adjust the shape parameter EX of the free water storage reservoir capacity curve, and output the corrected water content W of the lower layer L and parameter EX', enhancing the model's response ability to groundwater level changes.

[0117] W L = W L0 ·[1 + η·(H g - H g0 ) / H g0 ; where, WL is the water content of the lower soil layer after considering the influence of the groundwater level; W L0 is the water content of the lower soil layer calculated by the original XAJ model (without considering the groundwater level feedback); H g is the current groundwater level; H g0 is the reference groundwater level (usually taking the annual average groundwater level); η is the feedback coefficient, with a value of 0.35 (determined by experimental optimization); when the groundwater level rises (H g > H g0 ), the water content of the lower soil layer increases accordingly; when the groundwater level drops (H g < H g0 ), the water content of the lower soil layer decreases.

[0118] EX' = EX·[1 - θ·(H g - H g0 ) / H g0 ; EX' is the corrected EX value; EX is the original parameter value; θ is the adjustment coefficient, with a value of 0.28.

[0119] It solves the deficiency that groundwater and soil water are independent of each other in the traditional XAJ model and more accurately simulates the characteristics of high groundwater level and significant changes in the plain water network area. When the groundwater level rises, on the one hand, it increases the water content of the lower soil layer and reduces the infiltration depth; on the other hand, it reduces the free water storage reservoir capacity curve index, enabling smaller rainfall to generate larger groundwater runoff, thus better capturing the complex influence of groundwater level changes on runoff formation in the plain area.

[0120] Differentiated evaporation calculation unit, used for: based on water body coverage data, calculating the water surface evaporation E W and land surface evapotranspiration E L , comprehensively considering the influence of impervious area, and outputting the total regional evaporation E T , which affects the water balance calculation and the soil water content state before runoff generation.

[0121] Improved XAJ runoff generation calculation unit, used for: integrating the outputs of the above units, adding the direct runoff R imp from the impervious area and the runoff R S from the permeable area, and outputting the total runoff R, providing input for the subsequent confluence calculation, and realizing the accurate simulation of the runoff generation process in the plain water network area.

[0122] S5. Multi-directional flow pattern identification and prediction with depth-time-space fusion

[0123] Breaking through the traditional binary classification method, defining five flow patterns including normal flow (C1), weak countercurrent (C2), strong countercurrent (C3), split flow (C4) and circulation flow (C5), and establishing a multi-flow pattern classification standard.

[0124] Based on historical water level and flow data, river network topological skeleton data, and flow regime marking data, a deep learning model integrating spatio-temporal attention mechanism is constructed to replace the original random forest model, generating a deep model for flow regime recognition and achieving accurate recognition of complex flow regimes.

[0125] Design flow regime intensity quantification indicators, such as reverse flow intensity index (RFI), diversion ratio coefficient (DRC), and circulation stability (CCS), calculate the values of each indicator based on water level gradient data and flow velocity data, and form a flow regime intensity dataset.

[0126] For diversion nodes, integrate a data-driven model and a simplified hydraulic model to establish a flow distribution model considering water level difference, river channel morphology, and roughness, accurately calculate the dynamic distribution coefficient of diversion nodes, and generate a node distribution coefficient dataset.

[0127] Based on the deep model for flow regime recognition and the flow regime intensity dataset, combined with predicted rainfall data and water conservancy project scheduling information, predict the flow regime type and intensity in the future period, generate flow regime prediction results, and provide a basis for confluence calculation.

[0128] S6. Construction of time-varying structure adaptive BP neural network

[0129] Based on historical flow data and meteorological data, use time series decomposition and change point detection algorithms to automatically identify flood seasons, normal water seasons, and dry seasons, and generate hydrological period segmentation marking data.

[0130] Design a multi-scale time feature extraction network including short-term memory unit (LSTM), medium-term memory unit (Temporal Convolutional Network, TCN), and long-term memory unit (Attention mechanism) to replace the traditional BP neural network structure, enhance the ability to capture hydrological processes at different time scales, and generate a time feature extraction network.

[0131] Convert the scheduling information in the digital twin database of engineering facilities into time series features, and integrate the scheduling behavior into the neural network by designing a special scheduling information embedding layer to generate a scheduling information enhanced network and improve the perception ability of the impact of water conservancy project scheduling.

[0132] Design an adaptive structure optimization algorithm based on neural architecture search (NAS) to automatically adjust the network structure parameters according to the hydrological period segmentation marking data, and generate a period-adaptive network structure for different hydrological periods.

[0133] Adopt a multi-objective optimization training strategy, consider NSE, RMSE, and R indicators simultaneously, introduce L1 regularization to reduce the risk of overfitting, train the adaptive BP neural network model, and achieve accurate response to different hydrological conditions.

[0134] Specifically, the time-varying structure adaptive BP neural network includes:

[0135] A hydrological period automatic recognition unit, which is used to receive historical flow and meteorological data, automatically recognize the high-water period, normal-water period, and low-water period through time series decomposition and change point detection algorithms, and output hydrological period segmentation marker data to provide a basis for subsequent network structure adjustment.

[0136] A multi-scale time feature extraction network, which is used to receive improved XAJ runoff data, flow regime prediction results, and flow regime intensity datasets, capture hydrological features at different time scales from hours to months through three parallel sub-networks (LSTM short-term memory unit, TCN medium-term memory unit, and attention mechanism long-term memory unit), and fuse the extracted multi-scale features to output a comprehensive representation.

[0137] A water conservancy project scheduling information integration unit, that is, a scheduling information embedding layer, which is used to convert engineering scheduling data such as sluice opening and pump station operation status into time series feature vectors, and fuse them with multi-scale features through a specially designed embedding layer to enhance the model's perception ability of water conservancy project intervention.

[0138] An adaptive network structure optimization unit, which is used to automatically adjust structural parameters such as the number of network layers, hidden layer dimensions, and connection weights based on hydrological period segmentation marker data and neural architecture search algorithms, dynamically optimize the network structure for different hydrological conditions (high-water period, normal-water period, low-water period), and output a period-adaptive network configuration.

[0139] The input data of the improved XAJ runoff data, flow regime prediction results, and flow regime intensity datasets enter the hydrological period automatic recognition unit, which recognizes the high-water period, normal-water period, and low-water period through time series decomposition and change point detection algorithms, and generates hydrological period segmentation marker data.

[0140] According to the recognized hydrological period, the adaptive network structure optimization unit dynamically adjusts the structural parameters of the neural network, such as the number of layers, hidden layer dimensions, and connection weights.

[0141] The water conservancy project scheduling information is converted into a vector representation through a special embedding layer and integrated into the model. Subsequently, the multi-scale time feature extraction network processes the input data in parallel: the short-term memory unit (LSTM) captures the changes at the hour-to-day scale, the medium-term memory unit (TCN) processes the features at the day-to-week scale, and the long-term memory unit (attention mechanism) extracts the patterns at the week-to-month scale.

[0142] The features extracted by each unit are integrated through a spatio-temporal fusion layer, and finally the predicted runoff is output through a fully connected layer. This model adopts a multi-objective optimization training strategy, considering the NSE, RMSE, and correlation coefficient R simultaneously, and realizes high-precision runoff simulation under different hydrological conditions.

[0143] S7. Multi-Model Dynamic Fusion and Real-Time Correction

[0144] Based on the current hydrometeorological conditions and predicted rainfall data, identify the current hydrological scenarios (such as normal flow, rising flood peak period, falling flood peak period, dry season, etc.), and generate the optimal model combination strategy for each scenario.

[0145] Adopt the Bayesian Model Averaging (BMA) method to assign dynamic weights to the improved XAJ model, flow regime identification model, and adaptive BP neural network based on historical prediction performance, realize model fusion, and generate the prediction results of the fusion model.

[0146] Utilize the distribution data of temporary pumping stations and pumping and drainage capacity parameters to calculate the water volume impact of temporary mobile pumping stations, design a deviation correction function, and perform real-time correction on the prediction results of the fusion model to generate corrected prediction results.

[0147] Design an online learning mechanism to adjust the model parameters in real time according to the deviation between the latest observed data and the predicted values, realize the incremental update of the model, improve the adaptability to emergencies, and generate an incremental update parameter set.

[0148] Conduct multi-index real-time evaluation on the model prediction results, including indicators such as NSE, RMSE, and R, generate a performance evaluation report, and display the evaluation results and model prediction uncertainty through a visualization interface.

[0149] In this embodiment, by establishing a multi-flow regime classification standard, designing a flow regime identification model with a spatio-temporal attention mechanism, and developing a flow regime intensity quantification index, accurate identification and quantification of the countercurrent phenomenon are realized. Through the classification and parameterization of the polder system, the construction of a threshold response mechanism, and the development of a multi-threshold piecewise function model, accurate simulation of the non-linear threshold effect of water volume exchange in the polder area is realized. Through a flow distribution model with hydraulic constraints, accurate calculation of the dynamic distribution coefficient of the diversion node is realized. By constructing a time-varying structure adaptive BP neural network, the problem of the fixed BP network structure in the original model is solved, and the adaptability to time-varying characteristics is enhanced.

[0150] In this embodiment, the multi-threshold piecewise function model is specifically as follows:

[0151] Read the historical water level data and water volume exchange observation data, group the water volume exchange data according to the water level difference between inside and outside (ΔH = H out - H in ), and use the change point detection algorithm (PELT, Pruned Exact Linear Time) to identify the water level difference threshold points where the water volume exchange behavior changes significantly, and obtain the initial threshold point set T={T1, T2, ..., Tn}.

[0152] Read the operation records of the polder and the set of initial threshold points, calculate the water level difference inside and outside during the occurrence of each operation event (such as opening the sluice, starting the pump, etc.), match and analyze it with the detected threshold points, identify the key threshold points with clear physical meanings, and obtain the set of key threshold points \(T'=\{T'_1, T'_2, \cdots, T'_m\}(m\leq n)\).

[0153] For each water level difference interval \([T'_i, T'_{i + 1}]\) divided by the set of key threshold points, read the historical water level data and water volume exchange observation data within this interval, and use the nonlinear regression method to fit the relationship between the water volume exchange flow rate (\(Q\) exchange ) and the water level difference (\(\Delta H\)): \(Q\) exchange =\(\alpha_i(\Delta H)\) βi +\(\gamma_i\);

[0154] Among them, \(\alpha_i\), \(\beta_i\) and \(\gamma_i\) are the characteristic parameters of each interval, representing the exchange coefficient, nonlinear exponent and basic exchange volume respectively. Optimize these parameters by the least squares method to obtain the interval parameter set \(\{(\alpha_i, \beta_i, \gamma_i), i = 1, 2, \cdots, m - 1\}\).

[0155] Read the pump station operation data in the polder operation records, combine with the water volume exchange observation data, extract the pump station pumping flow characteristics under different water level conditions and operation modes, establish the mapping relationship between the pump station pumping flow and operation parameters (such as the number of operating units, power, etc.), and obtain the pump station characteristic parameter set \(Q\) pump =\(\{Q\) pump1 , Q\) pump2 , \cdots, Q\) pump _k\}\).

[0156] Read the set of key threshold points, the interval parameter set and the pump station characteristic parameter set, and construct a complete multi-threshold piecewise function model to express the water volume exchange process of the polder:

[0157] Q\) exchange =\(\begin{cases}0, & \Delta H < T'_1(\text{no exchange interval});\\\alpha_1(\Delta H)&\\+\gamma_1, & T'_1\leq\Delta H < T'_2(\text{gravity flow interval});\\\alpha_2(\Delta H)&\\+\gamma_2 + Q\)

[0158] &, T'_2\leq\Delta H < T'_3(\text{gravity flow}+\text{partial pump station interval});\\\alpha_3(\Delta H)&\\+\gamma_3 + Q\) β1 &, T'_3\leq\Delta H < T'_4(\text{gravity flow}+\text{all pump station interval});\\\alpha_3(\Delta H)&\\+\gamma_3 + Q\)

[0159] &, T'_3\leq\Delta H < T'_4(\text{gravity flow}+\text{all pump station interval});\\\alpha_3(\Delta H)&\\+\gamma_3 + Q\) β2 &, T'_3\leq\Delta H < T'_4(\text{gravity flow}+\text{all pump station interval});\\\alpha_3(\Delta H)&\\+\gamma_3 + Q\) pump1 &, T'_2\leq\Delta H < T'_3(\text{gravity flow}+\text{partial pump station interval});\\\alpha_3(\Delta H)&\\+\gamma_3 + Q\)

[0160] &, T'_3\leq\Delta H < T'_4(\text{gravity flow}+\text{all pump station interval});\\\alpha_3(\Delta H)&\\+\gamma_3 + Q\) β3 &, T'_3\leq\Delta H < T'_4(\text{gravity flow}+\text{all pump station interval});\\\alpha_3(\Delta H)&\\+\gamma_3 + Q\) pump2 &, T'_3\leq\Delta H < T'_4(\text{gravity flow}+\text{all pump station interval});\\\alpha_3(\Delta H)&\\+\gamma_3 + Q\) ...

[0162] αm-1(ΔH) βm-1 + γm-1 + Q pump k, ΔH ≥ T'm-1 (Emergency dispatch interval)};

[0163] This function is encapsulated into a calculation module to obtain the water volume exchange model of the polder.

[0164] Select a section of historical water level data and water volume exchange observation data that have not participated in the modeling, input them into the water volume exchange model of the polder for prediction, calculate the NSE, RMSE, and R indicators of the predicted values and the observed values, fine-tune and optimize the interval parameter set to obtain the optimized parameter set, and update the water volume exchange model of the polder.

[0165] Check the physical rationality of the optimized parameter set to ensure that the parameters are within the physically interpretable range. For abnormal parameters, combine the attribute information in the polder type database for constraint adjustment to obtain the effective parameter set, and further update the water volume exchange model of the polder.

[0166] Output data for real-time calculation of the water volume balance in the polder area:

[0167] Water volume exchange model of the polder: It includes a multi-threshold piecewise function and its parameters, and is used to accurately calculate the water volume exchange of the polder under different water level conditions;

[0168] Set of key threshold points: Record the water level difference thresholds at which significant changes in the water volume exchange behavior of the polder occur;

[0169] Effective parameter set: It includes the exchange coefficient, non-linear index, and pumping station characteristic parameters for each interval.

[0170] In another embodiment of the present application, the river network connectivity analysis based on the graph convolutional network is specifically as follows:

[0171] Read the multi-scale river network topological model, convert the river network into a directed graph G(V, E), where the node set V represents the river section intersection points, diversion points, and key control points, and the edge set E represents the river sections. Each edge contains attributes such as length, width, and roughness, and construct the river network graph data structure.

[0172] Read the historical water level and flow data and the river channel physical property data, and construct a feature vector Fi for each node vi in the directed graph G(V, E): Fi = [location coordinates (x, y), elevation z, average water level h avg , water level amplitude Δh, average flow q avg , flow amplitude Δq, number of upstream river sections n Up , number of downstream river sections n down ; Normalize all node features to obtain the node feature matrix F.

[0173] Read the physical characteristics data and historical water level and flow data of the river channel, and construct a feature vector Gi for each edge ei in the directed graph G(V, E): Gi = [river reach length l, river channel width w, hydraulic radius R, Manning roughness coefficient n, average flow velocity v avg , flow velocity variation Δv, flow direction stability s]; where the flow direction stability s is defined as the time proportion of the river reach maintaining the same flow direction. Normalize all edge features to obtain the edge feature matrix G.

[0174] Design a dual-channel graph convolutional network structure, including a node information channel and an edge information channel.

[0175] Among them, the l-th layer graph convolutional operation of the node information channel is defined as: H l+1 = σ(D -1 / 2 A*D -1 / 2 H l W l ); where H l is the node feature of the l-th layer, A* is the adjacency matrix with self-connections added, D is the degree matrix, and W l is the learnable parameter matrix of the l-th layer, and σ is the activation function (ReLU).

[0176] The l-th layer graph convolutional operation of the edge information channel is defined as: P l+1 = σ(B -1 / 2 C*B -1 / 2 P l U l ); where P l is the edge feature of the l-th layer, C* is the edge connection matrix, B is the edge degree matrix, and U l is the learnable parameter matrix of the l-th layer. The two channels perform information fusion at each layer through the attention mechanism to obtain the GCN model structure.

[0177] Graph convolutional network training: Construct a training data set, select a part of the time period (such as the dry season, wet season, normal water season) in the historical water level and flow data, and label the flow propagation characteristics according to the observed flow direction and flow change, including information such as propagation speed and propagation path selection, to form a labeled data set.

[0178] Use the labeled data set to train the GCN model structure, adopt the cross-entropy loss function, the Adam optimizer, the batch size is 32, and the number of training epochs is 200 to obtain the trained river network GCN model.

[0179] Use the trained river network GCN model to perform forward calculation on the complete river network graph data structure, extract the hidden layer representations of each node and edge in the graph, and form an embedding vector representing the river network connectivity. Calculate the embedding similarity between nodes to construct the connectivity strength matrix C, where C ij represents the connectivity strength between node i and node j.

[0180] Based on the connectivity strength matrix C, use the maximum flow algorithm to analyze the maximum flow transmission capacity between any two points in the river network, calculate the main flow propagation paths and possible alternative paths, and obtain the flow propagation path set P and the path capacity set Cap.

[0181] Calculate the centrality indices of each node (such as betweenness centrality, closeness centrality), identify the key connected nodes and potential vulnerable points in the river network, that is, those nodes that will significantly affect the overall connectivity once blocked or failed, and generate the connectivity vulnerability point set V vul .

[0182] Output data for flow pattern identification and flow allocation model construction, including:

[0183] River network connectivity feature matrix: including the connectivity strength matrix C, the flow propagation path set P, the path capacity set Cap, and the connectivity vulnerability point set V vul .

[0184] River network GCN model: A trained graph convolutional network model capable of analyzing river network connectivity.

[0185] In another embodiment of the present application, a multi-flow pattern classification criterion is established, specifically:

[0186] Read historical water level flow data, water level gradient data, and flow velocity data, and define key feature parameters representing the water flow state:

[0187] Flow direction index FDI = v•grad h / |v|•|grad h|, where v is the flow velocity vector and grad h is the water level gradient vector;

[0188] Flow velocity coefficient of variation VVC = σv / μv, where σv is the standard deviation of flow velocity and μv is the average flow velocity;

[0189] Water level gradient coefficient of variation HGC = σ grad h / μ grad h, where σgrad h is the standard deviation of water level gradient and μgrad h is the average water level gradient;

[0190] Node flow balance ratio FBR = |Σq in - Σq out | / (Σq in + Σq out), where q in is the inflow, and q out is the outflow;

[0191] Calculate the values of these characteristic parameters at all monitoring points in different time periods to obtain the flow regime characteristic parameter set.

[0192] Read the flow regime characteristic parameter set, use the Gaussian Mixture Model (GMM) for unsupervised clustering, initially classify the water flow states into k categories (k is initially set to 7), and use the Bayesian Information Criterion (BIC) to determine the optimal number of clusters to obtain the preliminary flow regime clustering result.

[0193] Combined with the experience of hydrological experts, interpret and adjust the preliminary flow regime clustering result, merge or subdivide the clusters according to the physical meaning, and determine five major flow regimes:

[0194] Normal flow (C1): FDI > 0.7, VVC < 0.3, FBR < 0.2;

[0195] Weak countercurrent (C2): 0.3 < FDI < 0.7, 0.3 < VVC < 0.5, 0.2 < FBR < 0.4;

[0196] Strong countercurrent (C3): FDI < 0.3, VVC > 0.5, FBR > 0.4;

[0197] Diversion (C4): FDI > 0.5, VVC < 0.4, FBR < 0.3, and the number of outflow river reaches at the node > 1;

[0198] Circulation (C5): FDI shows periodic changes, VVC > 0.6, and a closed flow path is detected in the loop network structure;

[0199] Generate the characteristic parameter threshold table for each flow regime.

[0200] Read the labeled flow regime data and the river network topological skeleton data, and analyze the spatio-temporal distribution characteristics of different flow regimes, including:

[0201] Temporal distribution: the seasonality, duration, and conversion rules of different flow regimes;

[0202] Spatial distribution: the spatial aggregation and distribution rules of different flow regimes in the river network;

[0203] Associated factors: the correlation between different flow regimes and external factors such as rainfall and sluice-dam operation;

[0204] Generate a report on the spatio-temporal distribution characteristics of the flow regime.

[0205] Based on the characteristic parameter threshold table and the spatio-temporal distribution characteristics report of the flow regime, construct a multi-layer decision tree structure for automatic classification of the flow regime: if FDI > 0.7 and VVC < 0.3 and FBR < 0.2: Flow regime = C1 (normal flow);

[0206] elif FDI < 0.3 and VVC > 0.5 and FBR > 0.4: Flow regime = C3 (strong countercurrent);

[0207] elif 0.3 < FDI < 0.7 and 0.3 < VVC < 0.5 and 0.2 < FBR < 0.4: Flow regime = C2 (weak countercurrent);

[0208] elif FDI > 0.5 and VVC < 0.4 and FBR < 0.3 and the number of outflow river reaches > 1: Flow regime = C4 (diversion);

[0209] elif FDI has periodic changes and VVC > 0.6 and is located in a loop network structure: Flow regime = C5 (circulation);

[0210] else: Flow regime = to be classified (further analysis is required);

[0211] Optimize the decision tree parameters to obtain a flow regime classification decision model.

[0212] Use the flow regime classification decision model to reclassify the historical water level and flow data, mark the flow regime type at each time step and monitoring point, and generate a historical dataset of flow regime markings, which will be used as the label data for training the deep learning model in step S52.

[0213] Select typical hydrological event cases (such as flood period, dry period, gate and dam operation period), manually check the classification results in the historical dataset of flow regime markings, evaluate the classification accuracy, and fine-tune the characteristic parameter threshold table and the flow regime classification decision model for cases of classification errors or uncertainties to obtain an optimized multi-flow regime classification standard.

[0214] In another embodiment of the present application, the impact of the temporary mobile pumping station on real-time correction is specifically as follows:

[0215] Read the temporary pumping station distribution data and the river network connectivity characteristic matrix, and for each temporary pumping station, calculate its influence range in the river network using the hydraulic propagation algorithm:

[0216] For each temporary pumping station p:

[0217] 1. Determine the pump station location coordinates (xp, yp) and the connected river reach id;

[0218] 2. Initialize the influence coefficient matrix I, and set all elements to 0;

[0219] 3. Set the influence coefficient I(xp, yp) of the pumping station location to 1;

[0220] 4. Calculate upstream and downstream propagation:

[0221] Upstream propagation: I(u) = I(p) × exp(-d(u, p) / Lu), where d is the river distance and Lu is the upstream attenuation length;

[0222] Downstream propagation: I(d) = I(p) × exp(-d(d, p) / Ld), where Ld is the downstream attenuation length;

[0223] 5. Loop network propagation: For node e in the loop network structure, I(e) = I(p) × exp(-d(e, p) / Le), where Le is the loop network attenuation length; Calculate the influence range for all temporary pumping stations to obtain the pumping station influence coefficient matrix I.

[0224] Read the pumping and drainage capacity parameters, temporary pumping station distribution data, and the pumping station influence coefficient matrix I, calculate the actual pumping and drainage flow Qp of each temporary pumping station in the current period, and allocate it to each node within the influence range according to the influence coefficient:

[0225] For each node n in the river network:

[0226] 1. Initialize the total flow ΔQn affected by the pumping station at node n to 0;

[0227] 2. For each temporary pumping station p:

[0228] Calculate the influence amount of pumping station p on node n: ΔQn,p = Qp × I(n, p);

[0229] Accumulate the total influence amount: ΔQn += ΔQn,p;

[0230] 3. Store the total influence amount ΔQn of node n; Generate a pumping and drainage influence flow distribution map, indicating the flow influence of temporary pumping stations on each node of the river network.

[0231] Read the pumping and drainage capacity parameters and the latest observation data, analyze the operation modes of temporary pumping stations in the recent period (such as continuous operation, intermittent operation, emergency scheduling, etc.), use the time series pattern recognition algorithm to identify the operation rules, predict the possible operation states in the short term in the future, and obtain the prediction results of the pumping station operation modes.

[0232] Read the drainage influence flow distribution map of the historical period and the actual observed flow changes, establish a response characteristic model of the temporary pumping station drainage on the hydrological process, use a convolutional neural network (CNN) to learn the spatial response characteristics, and a long short-term memory network (LSTM) to learn the temporal response characteristics to obtain the pumping station influence response model.

[0233] Based on the pumping station influence response model, construct a prediction result correction function: Q c orrected(i,t) = Q P redicted(i,t) + f(ΔQ(i,t), M(t), S(i,t)); where Q c orrected is the corrected predicted flow, Q P redicted is the initial predicted flow, ΔQ is the pumping station influence flow, M is the operation mode characteristic, S is the hydrological state characteristic, and f is a non-linear mapping function (implemented using a multi-layer perceptron). Train and optimize the parameters of the correction function to obtain the prediction correction function model.

[0234] For multi-step predictions (such as predicting the next 24 hours, 48 hours), design a progressive correction strategy:

[0235] For the prediction duration t = 1 to T:

[0236] 1. Initial correction: Q'(t) = Q(t) + ΔQ direct (t), where ΔQ direct is the direct influence correction;

[0237] 2. Cumulative effect correction: Q''(t) = Q'(t) + ΔQ accumul (t), where ΔQ accumul is the cumulative effect correction;

[0238] 3. Uncertainty increase: Increase the prediction uncertainty range with t, Uncertainty(t) = U base + k•t;

[0239] Generate a multi-step prediction correction strategy.

[0240] Read the prediction results of the fusion model, the drainage influence flow distribution map, and the prediction results of the pumping station operation mode, apply the prediction correction function model and the multi-step prediction correction strategy for real-time correction, and generate the corrected prediction results and the corrected confidence interval.

[0241] In this embodiment, aiming at the problem of insufficient dynamic identification and quantification mechanism for the reverse flow phenomenon in the plain water network area: breaking through the traditional binary classification limitation, a classification standard for five flow patterns including normal flow, weak reverse flow, strong reverse flow, diversion flow and circulation flow is designed. At the same time, multi-dimensional characteristic parameters such as flow direction index and velocity variation coefficient are defined, and a deep learning model integrating spatio-temporal attention mechanism is constructed to replace the original random forest model. In addition, quantification indexes such as reverse flow intensity index are developed, which jointly realize the accurate identification and quantification of complex flow patterns in the plain water network area, and improve the model classification accuracy from 76.3% to 89.5%.

[0242] Aiming at the problem of inaccurate simulation of the non-linear threshold effect of water volume exchange in the polder area: the classification parameters of the polder system are parameterized, and a multi-threshold piecewise function model is developed, which more accurately expresses the non-linear water volume exchange process inside and outside the polder under different water level conditions, including no-exchange interval, gravity flow interval and pumping station drainage interval. At the same time, an adaptive threshold dynamic adjustment mechanism is designed to automatically optimize the threshold parameters according to historical simulation errors and real-time water regime changes, enabling the model to accurately simulate the threshold triggering behavior and mutation process of the polder system under extreme hydrological conditions, and improving the NSE of water volume exchange simulation from 0.62 to 0.86.

[0243] Aiming at the problem of insufficient optimization of the dynamic distribution coefficient of diversion nodes in the plain water network area: a multi-scale river network topological structure is constructed, and a dual-channel graph convolutional network is designed to analyze the connectivity of the river network, extract node feature vectors and edge feature vectors, calculate the connectivity intensity matrix and the main flow propagation path, and based on this, a flow distribution model integrating data-driven and hydraulic constraints is developed to accurately calculate the dynamic distribution coefficient of diversion nodes under different hydrological conditions, significantly improving the flow distribution accuracy of diversion nodes, and achieving a flow direction prediction accuracy of 92.8% in complex water network structures.

[0244] Aiming at the problem of real-time correction parameters affected by temporary mobile pumping stations: a monitoring scheme for temporary pumping stations based on the Internet of Things and crowd-sourced data is designed, an algorithm for determining the influence range of pumping stations and a calculation method for the spatio-temporal distribution of drainage flow are developed. At the same time, a prediction result correction function and a multi-step prediction deviation progressive correction strategy are constructed to realize the real-time correction of the influence of the dynamic changes of temporary pumping stations. During the extreme rainfall in 2021, the NSE of the flow prediction of the area affected by temporary pumping stations in the model was improved from 0.78 to 0.89, and the RMSE was reduced from 2.35m 3 / s to 1.42m 3 / s, effectively solving the prediction deviation problem caused by the sudden intervention of temporary facilities.

[0245] In another embodiment of this application, an embodiment is provided to describe the specific processing process of the runoff simulation method in the plain water network area, which is applicable to other plain water network areas with complex river networks, numerous polders and intensive water conservancy engineering facilities.

[0246] The implementation process of this method is described in detail below in combination with a specific application example in a demonstration area in the Taihu Lake Basin (with an area of about 1,200 square kilometers, a river network density of 4.2 kilometers per square kilometer, and a polder coverage rate of 65%).

[0247] There are 25 rain gauges, 18 water level stations, 12 flow stations and 42 monitoring points of sluice pump stations in the demonstration area. First, collect the hourly rainfall data, water level and flow data, topographic data (1:10,000 digital elevation model), 10-meter resolution remote sensing images (Sentinel-2) and water conservancy project facility information from January 2018 to December 2021 to form an original data set.

[0248] For rainfall data, the terrain-weighted inverse distance interpolation method is used to correct the spatial distribution: P(x,y) = Σ[P i ·w i ·f(h i ,h xy )] / Σ[w i ·f(h i ,h xy )]; where P(x,y) is the corrected rainfall at point (x,y), P i is the observed value of the i-th rain gauge, w i is the spatial weight coefficient, f(h i ,h xy ) is the terrain influence function, and h i and h xy are the elevations of the rain gauge and the calculation point, respectively.

[0249] Extract the water body distribution from the remote sensing images, where the improved water body index MNDWI = (ρ green - ρ SWIR ) / (ρ green + ρ SWIR) ; where ρ green is the reflectance of the green band, and ρ SWIR is the reflectance of the short-wave infrared band. The water body extraction threshold is set to 0.1, that is, the pixels with MNDWI>0.1 are identified as water bodies.

[0250] For temporary mobile pump stations, in combination with the data reported by the management department and the results of remote sensing image change detection, 37 temporary pump stations deployed during the extreme rainfall in July 2021 are identified, and their location coordinates (x P ,y P ), installation time t P , evacuation time t_r and drainage capacity Q P (cubic meters per second) are recorded.

[0251] Perform spatio-temporal consistency correction on all the above data, and adopt a missing data repair algorithm based on tensor decomposition:

[0252] min ||P Ω (X) - P Ω (M)|| F 2 + λ||X|| * ; where X is the complete data tensor, M is the observed data tensor, P Ω is the projection operator, λ is the regularization parameter (taking the value of 0.01), and ||·|| * is the nuclear norm. Solve it by the alternating direction multiplier method to obtain the spatio-temporal consistency hydrological data set.

[0253] There are 78 polders in the demonstration area. According to their locations, scales, and management methods, they are divided into urban flood control polders (15), agricultural production polders (52), and ecological protection polders (11). The typical parameters of various polders are shown in the following table:

[0254] Polder type <![CDATA[Area range (km 2 )]]> Levee elevation (m) Warning water level (m) Guaranteed water level (m) <![CDATA[Pumping station density (units / km 2 )]]> Urban flood control polder 5.2-28.6 6.5-8.2 3.8-4.2 5.0-5.5 0.8-1.2 Agricultural production polder 1.5-15.3 5.8-7.0 3.5-3.8 4.5-5.0 0.3-0.7 Ecological protection polder 2.3-18.7 5.5-6.5 3.2-3.6 4.2-4.8 0.2-0.4

[0255] Take the agricultural production polder A12 as an example to illustrate the development process of the multi-threshold piecewise function model in detail:

[0256] (1) Piecewise clustering of water exchange data

[0257] For the historical water level data and water exchange observation data from 2018 to 2020, group the water exchange data Q out x according to the difference in water levels inside and outside ΔH (the water level outside the polder H in minus the water level inside the polder H E ). Use the PELT (Pruned Exact LinearTime) change point detection algorithm to identify the threshold points where significant changes occur in the water exchange behavior: max Σ[log(likelihood(data i )) + β]; where the likelihood function adopts a Gaussian distribution, and β is the penalty parameter (taking the value of 15). Detect 5 initial threshold points, which are ΔH = -0.25m, 0.10m, 0.35m, 0.62m, and 0.85m respectively.

[0258] Match and analyze the difference in water levels inside and outside corresponding to the operation time points of 268 gate openings and 153 pump starts in the operation records of polder A12 from 2018 to 2020 with the set of initial threshold points, and identify the key threshold points with clear physical meanings: T1 = 0.10m (gate opening start threshold), T2 = 0.35m (small pump station start threshold), and T3 = 0.85m (large pump station start threshold).

[0259] (2) Threshold interval flow response pattern recognition

[0260] For each water level difference interval [Ti, Ti+1], the relationship between the water exchange flow and the water level difference is fitted by using the nonlinear regression method: Q E x = α·(ΔH) β + γ; The parameters of each interval are obtained by optimizing the parameters through the least squares method:

[0261] Interval Water level difference range (m) α (Exchange coefficient) β (Nonlinear index) γ (Basic exchange volume) Interval 1 ΔH < 0.10 0 0 0 Interval 2 0.10 ≤ ΔH<0.35 15.27 1.42 0.25 Interval 3 0.35 ≤ ΔH<0.85 18.53 1.36 2.78 Interval 4 ΔH ≥ 0.85 21.06 1.28 5.35

[0262] Analyze the pump station operation data of polder A12, and extract the pumping flow characteristics of its 3 pump stations under different water level conditions:

[0263] Pump station number Type <![CDATA[Rated flow rate (m 3 / s)]]> Startup threshold (m) Operating water level range (m) P1 Small 1.2 0.35 0.35-1.20 P2 Small 1.5 0.40 0.40-1.20 P3 Large 3.8 0.85 0.85-1.50

[0264] (3) Construction of multi-threshold piecewise function

[0265] Combining the key threshold points, interval parameters and pump station characteristics, a complete multi-threshold piecewise function model is constructed:

[0266] Q E x = {0, ΔH < 0.10;

[0267] 15.27·(ΔH) 1.42 + 0.25, 0.10 ≤ ΔH < 0.35;

[0268] 18.53·(ΔH) 1.36 + 2.78 + Q P1+P2 , 0.35 ≤ ΔH < 0.85;

[0269] 21.06·(ΔH) 1.28 + 5.35 + Q P1+P2+P3, ΔH ≥ 0.85}; where, Q P1+P2 represents the pumping flow of small pump stations P1 and P2, and Q P 1+P2+P3 represents the pumping flow of all pump stations. The pump station flow is dynamically calculated according to the actual operation status. The data in 2021 is used to verify the model performance, and the NSE reaches 0.86, and the RMSE is 1.25 m 3 / s, which is significantly improved compared with the original linear exchange model (NSE = 0.62).

[0270] Based on the historical simulation error and the real-time water situation change, an adaptive adjustment mechanism is designed:

[0271] T'i = Ti·[1 + k·(ε / εmax)·(dH / dt) / (dH / dt)max]; where T'i is the adjusted threshold, Ti is the reference threshold, k is the adjustment coefficient (with a value of 0.15), ε is the recent simulation error, εmax is the historical maximum error, dH / dt is the water level change rate, and (dH / dt)max is the historical maximum water level change rate.

[0272] (1) Construction of the river network graph structure: Based on high-resolution remote sensing images and topographic data, the river network in the demonstration area is extracted, and 628 nodes (including 397 ordinary connection points, 156 diversion points, and 75 confluence points) and 785 river sections (edges) are identified. The river network is transformed into a directed graph G(V,E), where the node set V represents various intersection points and the edge set E represents the river sections.

[0273] (2) Construction of node and edge feature vectors: Feature vectors are constructed for each node, including position coordinates (x,y), elevation z, average water level h avg , water level amplitude Δh, etc., with a total of 9 feature dimensions. Feature vectors are constructed for each edge, including river section length l, river channel width w, Manning roughness coefficient n, etc., with a total of 7 feature dimensions. The average flow velocity v of the river section avg is calculated through historical flow data and cross-section information: v avg = Q / (B·H); where Q is the flow rate, B is the river channel width, and H is the average water depth.

[0274] The flow direction stability s is defined as: s = |∑sign(v i )| / n; where v i is the flow velocity at the i-th time step, the sign function returns the positive or negative of the flow velocity, and n is the total number of time steps. The closer the s value is to 1, the more stable the flow direction.

[0275] (3) Design of the dual-channel graph convolutional network: Design the dual-channel graph convolutional network structure, including the node information channel and the edge information channel.

[0276] The graph convolutional operation of the node information channel is defined as: H l+1 = σ(D -1 / 2 ·A*·D -1 / 2 ·H l ·W l ); where H l is the node feature matrix of the l-th layer, A* is the adjacency matrix with self-connections added, D is the degree matrix, W l is the learnable parameter matrix of the l-th layer, and σ is the ReLU activation function. The graph convolutional operation of the edge information channel is defined as: P l+1 = σ(B -1 / 2 ·C*·B -1 / 2 ·P l·U l )); where, P l is the edge feature matrix of the l-th layer, C* is the edge connection matrix, B is the edge degree matrix, and U l is the learnable parameter matrix of the l-th layer.

[0277] The two channels are fused through the attention mechanism: H l+1 fusion = α·H l+1 + (1 - α)·M·P l+1 ; where, α is the attention weight parameter and M is the node-edge mapping matrix.

[0278] (4) Network training and feature extraction: Select typical hydrological periods (120 days each for the dry season, wet season, and normal season) from the historical data from 2018 to 2020 to construct a training dataset. The water level and flow data for each period are used to label the flow propagation characteristics.

[0279] The cross-entropy loss function is adopted, the initial learning rate is 0.001, the Adam optimizer is used, the batch size is 32, and it is trained for 200 rounds. The classification accuracy of the trained river network GCN model on the test set reaches 92.8%. Calculate the embedding similarity between nodes through the trained model, and construct a 628×628-dimensional connectivity strength matrix C, where C i ij represents the connectivity strength between node i and node j (value range 0 - 1). Based on the connectivity strength matrix, use the maximum flow algorithm to analyze the maximum flow transmission capacity between any two points in the river network, identify 37 main flow propagation paths, and form the flow propagation path set P. At the same time, calculate the node betweenness centrality index B c (v) = ∑∑σ St (v) / σ S t; where, σ St is the number of shortest paths from node s to node t, and σ St (v) is the number of shortest paths from s to t passing through node v. Identify 32 river network connectivity vulnerable points (nodes with the top 5% B c values), and form the connectivity vulnerable point set V vul .

[0280] Based on the historical water level and flow data, define the key feature parameters characterizing the water flow state: Flow Direction Index FDI = v·grad h / (|v|·|grad h|), where v is the velocity vector and grad h is the water level gradient vector. FDI close to 1 indicates that the flow direction is consistent with the water level gradient (normal flow), and FDI close to -1 indicates that the flow direction is opposite to the water level gradient (reverse flow). Velocity Variation Coefficient VVC = σ v / μ v , where σ vis the standard deviation of flow velocity, μ v is the average flow velocity. The coefficient of variation of water level gradient HGC = σ grad h / μ grad h, where σ grad h is the standard deviation of water level gradient, μ grad h is the average water level gradient. The node flow balance ratio FBR = |∑q in - ∑q out | / (∑q in + ∑q out ), where q in is the inflow, q out is the outflow.

[0281] Perform Gaussian Mixture Model (GMM) clustering on the set of flow state characteristic parameters, and use the Bayesian Information Criterion (BIC) to determine that the optimal number of clusters is 5. Combining with the experience of hydrological experts, determine the five categories of flow states and their characteristic parameter thresholds:

[0282] Normal flow (C1): FDI > 0.7, VVC < 0.3, FBR < 0.2;

[0283] Weak countercurrent (C2): 0.3 < FDI < 0.7, 0.3 < VVC < 0.5, 0.2 < FBR < 0.4;

[0284] Strong countercurrent (C3): FDI < 0.3, VVC > 0.5, FBR > 0.4;

[0285] Diversion (C4): FDI > 0.5, VVC < 0.4, FBR < 0.3, and at the same time the number of out - flow river reaches of the node > 1;

[0286] Circulation (C5): FDI shows periodic changes, VVC > 0.6, and a closed flow path is detected in the loop network structure.

[0287] Construct a multi - layer decision tree structure based on the characteristic parameter threshold table:

[0288] if FDI > 0.7 and VVC < 0.3 and FBR < 0.2: Flow state = C1 (Normal flow);

[0289] elif FDI < 0.3 and VVC > 0.5 and FBR > 0.4: Flow state = C3 (Strong countercurrent);

[0290] elif 0.3 < FDI < 0.7 and 0.3 < VVC < 0.5 and 0.2 < FBR < 0.4: Flow regime = C2 (weak countercurrent);

[0291] elif FDI > 0.5 and VVC < 0.4 and FBR < 0.3 and the number of outflow river reaches > 1: Flow regime = C4 (diversion)

[0292] elif FDI has periodic variation and VVC > 0.6 and is located in a loop network structure: Flow regime = C5 (circulation);

[0293] else: Flow regime = to be classified (further analysis required).

[0294] Mark the historical data from 2018 to 2020 to obtain a historical dataset of flow regime marks, containing flow regime mark data of 628 nodes × 26,280 time steps.

[0295] Using seasonal remote sensing images (Sentinel-2) and land use data from 2018 to 2021, the impervious area and its changes in the demonstration area were identified using deep learning methods (U-Net architecture). The impervious area generally shows an increasing trend, increasing from 19.7% in 2018 to 22.5% in 2021. An impervious area dynamic model IMP(t) = IMP_0 + ΔI·t + ε(t) was established; where IMP(t) is the proportion of the impervious area at time t, IMP_0 is the initial proportion of the impervious area (19.7%), ΔI is the annual growth rate (0.93% / year), and ε(t) is the seasonal fluctuation term.

[0296] The impervious area parameter is introduced into the runoff calculation of the XAJ model: R imp = P·IMP(t); where R imp is the direct runoff volume of the impervious area, and P is the rainfall.

[0297] Using the historical data of 12 groundwater monitoring wells in the demonstration area, a feedback relationship between the groundwater level and the lower soil water content was established: W L = W L0 ·[1 + η·(H g - H g0 ) / H g0 ; where W L is the lower soil water content considering the influence of the groundwater level, W L0 is the lower soil water content calculated by the original XAJ model, H g is the groundwater level, H g0is the reference groundwater level, and η is the feedback coefficient (with a value of 0.35).

[0298] The change in groundwater level significantly affects the shape parameter EX of the free water storage reservoir capacity curve: EX' = EX·[1 - θ·(H g - H g0 ) / H g0 ; where EX' is the corrected EX value, EX is the original parameter value, and θ is the adjustment coefficient (with a value of 0.28).

[0299] Based on the data of 24 soil profile sampling points, the three-layer soil structure in the XAJ model is corrected, and the vertical distribution of soil water content is described by a continuous function: W(z) = W S ·[1 - (z / z max )^α]^β; where W(z) is the soil water content at depth z, W S is the saturated water content, z max is the maximum depth of the soil profile, and α and β are shape parameters (α = 1.35, β = 0.85). Integrating gives the soil water content of each layer W i = ∫_{z i-1}^{z i} W(z)dz;

[0300] Based on the water body coverage data, the study area is divided into water bodies and land, and the water surface evaporation E W is calculated using the modified Penman formula:

[0301] E W = (0.408Δ(R n -G) + γ·900 / (T+273)·u2·(e S -e_a)) / (Δ+γ·(1+0.34u2))·k W ;

[0302] where k W is the water surface correction coefficient (1.05 - 1.15). The land surface evapotranspiration E L adopts a three-layer evaporation model E L = K E ·E P ·f(W U , W L , W D );where K E is the evapotranspiration coefficient, E P is the potential evaporation, and f(W U , W L , W D ) is a function of the three-layer soil water content. Combining the water surface evaporation and the land surface evapotranspiration, the total evaporation of the region E is calculatedT = E W ·A W + E L ·(1 - A W - IMP(t)); where A W is the proportion of water surface area.

[0303] Integrating the above improvements, calculate the runoff yield R of each subunit in the plain water network area: R = R imp + R S ; where R S is the runoff yield of the permeable area, which is calculated through the improved three - layer evaporation model and the vertical distribution of water content.

[0304] The simulation results of the runoff yield for the extreme rainfall event in July 2021 (cumulative rainfall of 432.7 mm) show that the Nash efficiency coefficient of the improved XAJ model reaches 0.92, which is significantly higher than that of the original XAJ model (NSE = 0.78). Especially in the sub - basins with a high degree of urbanization and areas with significant fluctuations in the groundwater level, the improvement effect is more obvious.

[0305] Construct a deep - learning model integrating spatio - temporal attention mechanisms. The network structure includes:

[0306] Spatial graph convolutional layer, which captures the spatial dependence relationship between nodes;

[0307] Temporal convolutional layer, which captures the temporal pattern;

[0308] Spatial attention layer, which dynamically assigns the importance of different spatial positions;

[0309] Temporal attention layer, which focuses on the key time steps;

[0310] Fully - connected classification layer, which outputs the probability distribution of five types of flow patterns;

[0311] The model is trained using the Adam optimizer, with a learning rate of 0.0005, a batch size of 64, and 300 training epochs. The classification accuracy on the validation set reaches 89.5%, which is significantly higher than that of the original random forest model (accuracy of 76.3%).

[0312] Design a flow - pattern intensity quantification index:

[0313] Reverse - flow intensity index (RFI): RFI = (1 - FDI)·(v / v max ), where v is the flow velocity magnitude and v max is the historical maximum flow velocity.

[0314] Divergence ratio coefficient (DRC): DRC = q max / ∑q out ), where q maxIs the maximum discharge rate.

[0315] Circulation stability (CCS): CCS = t cycle / T, where t cycle Is the time for a complete cycle of the circulation, and T is the reference time (set to 24 hours).

[0316] Based on the historical flow data and meteorological data from 2018 to 2020, time series decomposition and change point detection algorithms are used to automatically identify the high-flow period (mid-June to September), normal-flow period (March to mid-June, September to November), and low-flow period (December to February of the following year).

[0317] Construct a multi-scale time feature extraction network, including:

[0318] Short-term memory unit: Based on the LSTM network, the input dimension is 12 (including improved XAJ runoff, flow state prediction results, etc.), and the hidden layer dimension is 64, capturing hydrological changes on the hourly to daily scale.

[0319] Mid-term memory unit: Based on the temporal convolutional network (TCN), the convolutional kernel size is 24, and the stride is 4, capturing hydrological changes on the daily to weekly scale.

[0320] Long-term memory unit: Based on the multi-head attention mechanism (8 attention heads), capturing hydrological changes on the weekly to monthly scale.

[0321] Convert the engineering facility scheduling information into temporal features and integrate it into the neural network E through a dedicated scheduling information embedding layer S = W S ·X S + b S ; where E S Is the scheduling information embedding vector, X S Is the scheduling state vector, and W S And b S Are learnable parameters.

[0322] The scheduling information embedding layer converts discrete and multi-category water conservancy project scheduling operations (such as changes in sluice opening and starting / stopping status of pumping stations) into continuous feature vectors that can be processed by the neural network. Specifically, this embedding layer receives the scheduling status data of various water conservancy project facilities, maps these discrete scheduling behaviors to a low-dimensional continuous vector space through a learnable parameter matrix, generates a scheduling information embedding vector, and fuses it with hydrological features, enabling the neural network to capture the complex relationship between scheduling operations and hydrological processes. This enables the model to "understand" the time lag effect and spatial transmission characteristics of the impact of water conservancy project scheduling on river network water flow, significantly improving the model's prediction ability for hydrological process changes under human intervention.

[0323] Design an adaptive structure optimization algorithm based on neural architecture search (NAS) to automatically adjust the network structure parameters according to the hydrological period segmented labeled data:

[0324] High water period: deepen the number of LSTM layers (3 layers) and increase the hidden layer dimension (96);

[0325] Normal water period: balance the weights of the three memory units, standard network configuration;

[0326] Low water period: strengthen the weight of the long-term memory unit and reduce the LSTM hidden layer dimension (48).

[0327] The adaptive structure optimization algorithm includes the following steps:

[0328] Define the search space, including adjustable parameters such as the number of network layers, hidden layer dimension, connection type, and activation function, and

[0329] Set different constraint conditions for different hydrological periods (high, normal, and low water periods).

[0330] Use the efficiency reinforcement learning method. The controller network (based on LSTM) generates candidate architectures and evaluates their performance on the validation set (through weighted NSE, RMSE, and computational efficiency metrics).

[0331] Update the controller parameters based on the evaluation results and continuously optimize the architecture generation strategy. After multiple rounds of iteration, select the optimal network architecture for each hydrological period to form an architecture library.

[0332] According to the output of the hydrological period automatic identification unit, dynamically switch to the corresponding optimal network structure to achieve the adaptive adjustment of the model structure and effectively improve the simulation accuracy of each hydrological period.

[0333] Take the temporary pumping station during the extreme rainfall in July 2021 as an example to illustrate the calibration process:

[0334] For each temporary pumping station p, calculate its influence range in the river network using the hydraulic propagation algorithm:

[0335] For each temporary pumping station p:

[0336] 1. Determine the pumping station location coordinates (xp, yp) and the connected river section id;

[0337] 2. Initialize the influence coefficient matrix I, and set all elements to 0;

[0338] 3. Set the influence coefficient at the pumping station location I(xp, yp) = 1;

[0339] 4. Calculate the upstream and downstream propagation:

[0340] Upstream propagation: I(u) = I(p) × exp(-d(u,p) / Lu), where Lu = 2000m;

[0341] Downstream propagation: I(d) = I(p) × exp(-d(d,p) / Ld), where Ld = 5000m;

[0342] 5. Loop network propagation: For node e in the loop network structure, I(e) = I(p) × exp(-d(e,p) / Le), where Le = 3000m;

[0343] Calculate the actual pumping and drainage flow rate Qp of each temporary pumping station, and distribute it to each node within the influence range according to the influence coefficient matrix I to obtain the distribution map of the pumping and drainage influence flow rate.

[0344] Based on historical data, establish a pumping station influence response model, use CNN to learn spatial response features, and use LSTM to learn temporal response features. Construct a prediction result correction function Q corrected (i,t) = Q Predicted (i,t) + f(ΔQ(i,t), M(t), S(i,t)); where Q corrected is the predicted flow rate after correction, Q Predicted is the initial predicted flow rate, ΔQ is the pumping station influence flow rate, M is the operation mode feature, S is the hydrological state feature, and f is a non-linear mapping function.

[0345] Apply the correction function to the flow rate prediction results during the extreme rainfall period in July 2021, improve the NSE from 0.78 to 0.89, and reduce the RMSE from 2.35m 3 / s to 1.42m 3 / s, which reflects the effectiveness of the real-time correction of the influence of temporary pumping stations.

[0346] Compare the simulation effects of the method of the present invention with the existing XAJ-BP coupling model throughout the year of 2021 in the demonstration area:

[0347] Performance index This method Existing XAJ-BP Improvement percentage NSE 0.892 0.753 18.5% <![CDATA[RMSE(m 3 / s)]]> 1.56 2.82 44.7% <![CDATA[R 2 > 0.912 0.781 16.8%

[0348] Especially during extreme hydrological events (July 2021), the performance improvement is more significant:

[0349] Performance index This method Existing XAJ-BP Improvement percentage NSE 0.865 0.627 38.0% <![CDATA[RMSE(m 3 / s)]]> 2.31 5.18 55.4% <![CDATA[R 2 > 0.883 0.661 33.6%

[0350] The experimental results show that the method of the present invention has significant advantages in runoff simulation in plain water network areas, especially in making breakthrough progress in solving key technical problems such as reverse flow phenomenon identification, simulation of water volume exchange threshold effect in polder areas, dynamic allocation of diversion nodes, and correction of the influence of temporary pumping stations.

[0351] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A runoff simulation method for plain water network areas, characterized in that, Including: Obtain multi-source data and preprocess it to obtain a spatiotemporally consistent hydrological dataset; Based on the spatiotemporally consistent hydrological dataset, use an enhanced polder system dynamic model to obtain polder water level data and exchange flow data, and based on these, perform runoff generation calculation in the plain area through an improved XAJ model to generate improved XAJ runoff generation data; Based on the spatiotemporally consistent hydrological dataset, construct a multi-scale river network topological structure to obtain river network topological skeleton data and river network connectivity characteristic matrix, and based on this, perform multi-directional flow pattern recognition and prediction of deep spatiotemporal fusion; Obtain flow pattern prediction results and flow pattern intensity datasets, and combine them with the improved XAJ runoff generation data to construct a time-varying structure adaptive BP neural network model to achieve runoff simulation in the plain water network area; The process of achieving runoff simulation in the plain water network area includes: Based on historical flow data and meteorological data, use time series decomposition and change point detection algorithms to automatically identify flood seasons, normal water seasons, and dry seasons, and generate hydrological period segmentation marker data; Input the improved XAJ runoff generation data, flow pattern prediction results, and flow pattern intensity datasets into the constructed multi-scale time feature extraction network to achieve accurate simulation of the runoff process in the plain water network area; The multi-scale time feature extraction network includes: Short-term memory units based on long short-term memory networks, used to capture hydrological changes on the hourly to daily scale; Medium-term memory units based on temporal convolutional networks, used to capture hydrological changes on the daily to weekly scale; Long-term memory units based on attention mechanisms, used to capture hydrological changes on the weekly to monthly scale; An adaptive network structure that uses a neural architecture search algorithm to automatically adjust network structure parameters according to the hydrological period segmentation marker data; Based on the multi-scale time feature extraction network, adopt a multi-objective optimization training strategy, simultaneously consider NSE, RMSE, and R metrics, introduce L1 regularization to reduce the risk of overfitting, and train the time-varying structure adaptive BP neural network model.

2. The method according to claim 1, characterized in that, Using the enhanced polder system dynamic model to obtain polder water level data and exchange flow data includes: Read historical water level data and polder operation records from the spatiotemporally consistent hydrological dataset, identify key water level threshold points, and establish a set of key threshold points; According to the set of key threshold points, construct a multi-threshold piecewise function to form a polder water volume exchange model; According to real-time water regime changes and historical simulation errors, adjust the parameters of the set of key threshold points to generate a dynamic threshold parameter set; and accordingly update the polder water volume exchange model; Using the updated polder water volume exchange model, combine rainfall input, evaporation loss, and infiltration to calculate the water volume balance in the polder to obtain polder water level data and exchange flow data.

3. The method according to claim 2, characterized in that, Constructing a multi-threshold piecewise function to form a polder water volume exchange model includes: Based on the set of key threshold points, divide the water level difference into multiple intervals; Perform non-linear regression analysis on the historical water level data and water volume exchange observation data in each interval to obtain the exchange coefficient, non-linear index, and basic exchange volume in each interval, forming an interval parameter set; Extract pump station operation data from the polder operation records, analyze the pump station pumping characteristics under different water level conditions, and establish a set of pump station characteristic parameters; Combined with the set of key threshold points, interval parameter set, and pumping station characteristic parameter set, construct a multi-threshold piecewise function model to express the water volume exchange process in the polder, and adopt different water volume exchange calculation formulas in different water level difference intervals.

4. The method according to claim 3, characterized in that Based on the set of key threshold points, divide the water level difference into multiple intervals, including: Read historical water level data and water volume exchange observation data, group the water volume exchange data according to the internal and external water level differences, and use the change point detection algorithm to identify the threshold points where significant changes occur in the water volume exchange behavior, obtaining the initial threshold point set; Match and analyze the internal and external water level differences corresponding to the gate opening and pump starting operation time points in the polder operation records with the initial threshold point set, identify the key threshold points with clear physical meanings, and form the set of key threshold points; Based on the set of key threshold points, divide the water level difference intervals, and establish piecewise functions for the water volume exchange behavior in each interval.

5. The method according to claim 1, characterized in that, Construct a multi-scale river network topological structure, and obtain river network topological skeleton data and river network connectivity characteristic matrix, including: Extract river network data from the spatio-temporally consistent hydrological dataset, and construct a multi-scale river network topological model including the main river network, secondary river network, and capillary river network; Perform skeletonization processing on the multi-scale river network topological model, identify confluence points, divergence points, and loop network structures, and generate river network topological skeleton data including the river network centerline, key node positions, and river section connection relationships; Convert the river network topological skeleton data into a directed graph structure to form a river network graph data structure; Combined with historical water level and flow data and river channel physical characteristic data, construct a node characteristic matrix and an edge characteristic matrix; Use a dual-channel graph convolutional network model to process the river network graph data structure, extract the connectivity characteristics between nodes, and generate a river network connectivity characteristic matrix.

6. The method according to claim 5, characterized in that, Use a dual-channel graph convolutional network model to process the river network graph data structure, extract the connectivity characteristics between nodes, and generate a river network connectivity characteristic matrix, including: Use the trained river network GCN model to perform forward calculation on the river network graph data structure, extract the embedding similarity between nodes, and construct a connectivity strength matrix; Based on the connectivity strength matrix, calculate the maximum flow transmission capacity between any two points in the river network, analyze the main flow propagation paths, and obtain the flow propagation path set; Calculate the node centrality index, identify the key connected nodes and potential vulnerable points in the river network, and generate a set of connectivity vulnerable points; Integrate the connectivity strength matrix, flow propagation path set, and set of connectivity vulnerable points to form a river network connectivity characteristic matrix.

7. The method according to claim 1, characterized in that Perform multi-directional flow pattern recognition and prediction with deep spatio-temporal fusion to obtain flow pattern prediction results and a flow pattern intensity dataset, including: Extract historical water level and flow data, water level gradient data, and flow velocity data from the spatio-temporally consistent hydrological dataset, calculate key characteristic parameters, and form a set of flow pattern characteristic parameters; Based on the set of flow pattern characteristic parameters, establish a multi-flow pattern classification standard including five flow patterns: normal flow, weak countercurrent, strong countercurrent, divergence flow, and circulation flow; and mark the historical water level and flow data accordingly to generate a historical dataset of flow pattern labels; Through training a deep learning model that integrates spatio-temporal attention mechanisms based on the historical dataset of flow pattern labels, perform the recognition and prediction of complex flow patterns to obtain flow pattern prediction results; According to the flow regime intensity quantification index, combined with water level gradient data and flow velocity data, quantify the intensity characteristics of different flow regimes to generate a flow regime intensity dataset.

8. The method according to claim 7, characterized in that Establish a multi-flow regime classification standard including five types of flow regimes: normal flow, weak countercurrent, strong countercurrent, diversion flow, and circulation flow, including: Perform unsupervised clustering analysis on the flow regime characteristic parameter set to obtain a preliminary flow regime clustering result; Interpret and adjust the preliminary flow regime clustering result, determine the characteristic boundaries of the five types of flow regimes, and generate a characteristic parameter threshold table; Analyze the spatio-temporal distribution characteristics of different flow regimes, including time distribution, spatial distribution, and correlation factors, to form a flow regime spatio-temporal distribution characteristic report; Based on the characteristic parameter threshold table and the flow regime spatio-temporal distribution characteristic report, construct a multi-layer decision tree structure to form a flow regime classification decision model; Use the flow regime classification decision model to label historical data, and verify through typical hydrological event cases, fine-tune the characteristic parameter threshold table, and optimize the multi-flow regime classification standard.

9. The method according to claim 1, characterized in that, It also includes: Based on the river network connectivity characteristic matrix and the temporary pumping station distribution data, calculate the influence range of the temporary pumping station in the river network to form a pumping station influence coefficient matrix; According to the pumping and drainage capacity parameters and the pumping station influence coefficient matrix, calculate the spatio-temporal distribution of the pumping and drainage flow of the temporary pumping station to generate a pumping and drainage influence flow distribution map; Construct a prediction result correction function, fuse the flow regime prediction result and the pumping and drainage influence flow distribution map, realize real-time correction of the influence of the temporary mobile pumping station, and obtain the corrected prediction result.

Images