Adaptive correction method for river-pipe network coupling model based on multi-source data driving

Abstract

The application discloses a river-pipe network coupling model self-adaptive correction method based on multi-source data driving, which comprises the following steps: acquiring multi-source monitoring data and rainfall driving data of a river-pipe network system, and performing spatio-temporal alignment processing on the multi-source data to obtain processed multi-source monitoring data and rainfall driving data; based on a pre-stored current model parameter set, using a river-pipe network coupling model to calculate the processed rainfall driving data to generate a simulation sequence corresponding to the processed multi-source monitoring data; calculating deviation characteristics between the processed multi-source monitoring data and the simulation sequence; based on the deviation characteristics, determining a parameter correction strategy; the parameter correction strategy limits a target parameter subset and a correction space range of this correction. Through a deviation attribution driving mechanism, the application solves the problems of different parameters having the same effect and physical distortion in traditional full-amount correction, and improves the self-adaptive ability and parameter interpretability of the coupling model.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of hydrological and hydrodynamic simulation, and in particular to a river-pipe network coupling model adaptive correction method based on multi-source data driving. BACKGROUND

[0002] The urban water system is an important infrastructure for ensuring urban safety and promoting sustainable development, and its core includes the underground drainage pipe network system and the overground river system. Building a high-precision river-pipe network coupling hydrodynamic model is crucial for a deep understanding of urban hydrological processes, accurate prediction of waterlogging, and scientific flood control and drainage scheduling.

[0003] The existing technology usually adopts automatic calibration methods based on global optimization algorithms (such as genetic algorithm, particle swarm algorithm) or sequential data assimilation algorithms (such as ensemble Kalman filter) when dealing with river-pipe network coupling model correction problems. These methods mainly construct an objective function that minimizes the statistical error (such as root mean square error) between the simulation value and the observed value, and iteratively optimize the model parameters within the preset parameter space. At the same time, some schemes use Internet of Things collected liquid level and flow data to try to update the model parameters online, in order to make the model output approach the observed true value.

[0004] The existing technology currently faces the following key problems: first, the model parameter calibration (correction) process is highly dependent on human experience, and a set of available parameters is found by repeatedly adjusting parameters, running models, and comparing simulation results with limited monitoring data, which is low in automation and intelligence; second, the value of multi-source heterogeneous data has not been fully tapped and utilized, and with the development of Internet of Things and remote sensing technology, the data sources available for model verification are increasingly rich, including liquid level meters, flow meters, rain gauges deployed in pipe networks and rivers, surface rain from weather radar inversion, satellite / unmanned aerial vehicle remote sensing monitored surface water, and video monitoring images. However, existing methods can usually only use one or two types of data (such as a small number of liquid level meter data), and lack effective technical means for automatic cleaning, spatio-temporal alignment, and fusion analysis of multi-source data; third, the adaptability of the model is poor, and the underlying conditions (such as land use, impervious area) and the internal state of the pipe network (such as siltation, damage) will change over time, meaning that the effectiveness of the calibrated model parameters will gradually decrease; finally, the business support capability of the model is lagging, and traditional correction is carried out after the occurrence of historical events, which cannot correct the model in real time during the occurrence of rainstorm events. SUMMARY

[0005] The present application provides a river-pipe network coupling model adaptive correction method based on multi-source data driving to solve the above problems existing in the prior art.

[0006] The technical scheme is a river-pipe network coupling model adaptive correction method based on multi-source data driving, comprising the following steps:

[0007] Multi-source monitoring data and rainfall driving data of the river-pipe network system are acquired, and the multi-source data are processed in time and space to obtain processed multi-source monitoring data and rainfall driving data.

[0008] Based on a pre-stored current model parameter set, the river-pipe network coupling model is used to calculate the processed rainfall driving data to generate a simulation sequence corresponding to the processed multi-source monitoring data.

[0009] The deviation between the processed multi-source monitoring data and the simulation sequence is calculated to obtain a deviation feature.

[0010] Based on the deviation feature, a parameter correction strategy is determined; the parameter correction strategy limits a target parameter subset and a correction space range of this correction.

[0011] The beneficial effects are that, by using the deviation attribution driving mechanism, the problems of different parameters having the same effect and physical distortion in traditional full-amount correction are solved, and the adaptive ability and parameter interpretability of the coupling model are improved. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 A step flowchart of a river-pipe network coupling model adaptive correction method based on multi-source data driving is provided for the embodiments of the present application.

[0013] Figure 2 A step flowchart of acquiring multi-source monitoring data of a river-pipe network system is provided for the embodiments of the present application.

[0014] Figure 3 A step flowchart of determining a parameter correction strategy is provided for the embodiments of the present application.

[0015] Figure 4 A step flowchart of identifying a current system deviation type is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0016] In order to enable personnel in the art to better understand the present application scheme, the technical scheme in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0017] It is to be understood that the terms including and having and their conjugations are intended to cover an inclusive process, method, system, product, or apparatus that not only "comprises" or "has" the recited steps or units but also "consists essentially of" or "consists of" the recited steps or units, unless otherwise indicated.

[0018] In the research, it is found that the key defect of the prior art lies in the lack of physical causal diagnosis of the bias source, resulting in serious hetero-parameter homoeffect problem in the correction process, so that the physical meaning of the updated parameters is distorted. Specifically, the output bias of the coupled model may be caused by multiple mutually different sources such as rainfall input error, pipe network water delivery capacity change, or coupled interface physical state change. The existing method adopts a blind fitting strategy, and performs global optimization on all parameters without distinction, which is easy to mistakenly compensate the systematic bias of rainfall input as the dramatic change of hydraulic parameters (such as Manning coefficient). Although it can temporarily reduce the fitting error, it often destroys the physical consistency of the system such as mass conservation, resulting in the lack of robustness of the parameters between different rainfall events, and the inability to accurately identify the complex flow state of the river-pipe network coupling interface.

[0019] As shown in Figure 1 A multi-source data-driven adaptive correction method for river-pipe network coupled model is proposed, including the following steps:

[0020] Obtain multi-source monitoring data and rainfall driving data of the river-pipe network system, and perform spatio-temporal alignment processing on the multi-source data to obtain processed multi-source monitoring data and rainfall driving data.

[0021] That is, the multi-source monitoring data and rainfall driving data of the river-pipe network system are obtained, and the spatio-temporal alignment processing of the multi-source data is completed.

[0022] In this embodiment, obtaining multi-source monitoring data and rainfall driving data is the basis for driving subsequent model operation and correction. The river-pipe network system refers to a complex hydraulic system composed of underground urban drainage pipe network and natural or artificial river water system on the ground. Multi-source monitoring data is a physical quantity reflecting the system state collected in real time through various technical means, which can specifically include but is not limited to radar inversion area rainfall, point rainfall measured by ground rain gauge, water level of pipe network key node, water level and flow velocity of river section, and surface water depth identified by remote sensing or video, etc. These data often have the characteristics of multi-source heterogeneity, for example, radar data is spatial grid data, while water level meter is time series point data. Therefore, spatio-temporal alignment processing is needed to map the data of different sources, different frequencies and different spatial resolutions to a unified spatio-temporal reference. Specifically, time alignment can use linear interpolation or nearest neighbor matching method to uniformly resample all data to model calculation step length, such as 5 minutes or 1 minute; spatial alignment is to map radar grid rainfall or remote sensing water accumulation information to corresponding model sub-catchment or calculation unit. The consistency of the data input into the model in the spatio-temporal dimension lays a foundation for subsequent accurate simulation.

[0023] In some optional embodiments, the spatio-temporal alignment processing of the data can also include preprocessing of abnormal values, such as removing outliers caused by sensor failure, or filling data missing due to communication packet loss by spatial interpolation, to ensure the continuity and reliability of the input data.

[0024] Based on the pre-stored current model parameter set, the river-pipe network coupling model is used to calculate the processed rainfall driving data to generate a simulation sequence corresponding to the processed multi-source monitoring data.

[0025] Specifically, the current model parameter set refers to a set of parameters describing the physical properties of the system before correction, such as pipe Manning coefficient, river roughness, and sub-catchment impermeability, etc. The river-pipe network coupling model is a hydrodynamic model that can simulate pipe flow and river open channel flow as well as water exchange between the two, such as a mathematical model based on one-dimensional Saint-Venant equation set. By inputting the aligned rainfall driving data into the model, the change process of system state variables over time is calculated, i.e. the simulation sequence. The simulation sequence corresponds to the processed multi-source monitoring data in physical meaning, time step and spatial position, for example, if the monitoring data contains the water level process of a certain inspection well A from 08:00 to 09:00, the simulation sequence should also contain the simulated water level process at the same position and time period. The mapping from the model parameter space to the physical state space is realized, providing a comparison benchmark for subsequent deviation calculation.

[0026] The deviation between the processed multi-source monitoring data and the simulation sequence is calculated to obtain the deviation feature.

[0027] In this embodiment, the deviation feature is not only a simple numerical error between the simulation value and the monitoring value, such as the root mean square error RMSE, but also a deep feature that can reflect the error physical properties and sources. Specifically, the deviation feature can include but is not limited to the error time distribution characteristics, such as whether the error occurs at the beginning of the rainfall or at the peak moment, the spatial distribution characteristics, such as whether the error is global or limited to a certain area, and the physical conservation characteristics, such as the water balance of the whole system. By extracting these features, the system no longer just sees how big the error is, but can perceive what the error looks like. For example, calculating the mass conservation non-closed amount can reveal the systematic deviation of the input rainfall, and calculating the spatial clustering degree of the deviation can distinguish between local pipe network blockage and global rainfall error. Based on the feature extraction of physical meaning, the traditional black box optimization correction is distinguished, which provides a basis for subsequent causal diagnosis.

[0028] Based on the deviation feature, a parameter correction strategy is determined; the parameter correction strategy limits the target parameter subset and the correction space range of this correction.

[0029] Specifically, the parameter correction strategy is a top-level decision instruction generated based on the deviation feature, which solves the problems of where to change and who to change. The system identifies the most likely cause of the current deviation, i.e., the deviation attribution, through a pre-set logic or discrimination model based on the deviation feature, and formulates a strategy accordingly. The target parameter subset refers to the parameter set allowed to be adjusted in this correction, while other non-target parameters are locked and remain unchanged. For example, if the deviation feature shows rainfall input error, the strategy will limit the target parameter subset to only the rainfall correction coefficient, and strictly prohibit the modification of pipe roughness and other hydraulic parameters. The correction space range refers to the geographical or topological area where the parameter adjustment takes effect, such as adjusting the entire watershed or only adjusting the upstream area of a certain pipe section. Through directional strategy generation, the high-dimensional global optimization problem is effectively reduced to a low-dimensional local or specific type parameter optimization problem, avoiding the loss of physical meaning of parameters caused by the same effect of different parameters, so that the corrected model parameters not only have high fitting degree, but also conform to the physical reality.

[0030] This embodiment not only focuses on the numerical fitting of model parameters, but also emphasizes the diagnosis of the source of deviation through the extraction of physical features to achieve directional and accurate parameter correction. A technical architecture including data perception layer, model simulation layer, deviation diagnosis layer and decision execution layer is constructed.

[0031] In one possible implementation, the multi-source monitoring data includes data from IoT sensing devices deployed in drainage networks and waterways. The IoT sensing devices include at least: radar flow meters for monitoring network flow, pressure level gauges for monitoring node water levels, ultrasonic flow meters for monitoring river flow velocities, and optical rain gauges for monitoring localized rainfall.

[0032] In this embodiment, a hardware sensing layer is defined to support the operation of an adaptive correction method for a river-pipeline coupling model driven by multi-source data. IoT sensing devices constitute the nerve endings of the urban water system. Specifically, radar flow meters are installed non-contactly in pipe network inspection wells, capable of simultaneously measuring surface velocity and water level to calculate flow rate, suitable for complex conditions such as high flow velocities or non-full pipe flow. Pressure level gauges convert hydrostatic pressure into water level, featuring fast response and low cost, suitable for large-scale deployment at pipe network nodes to obtain extensive water level distribution information. Ultrasonic current meters are installed at river cross-sections, using the Doppler effect to measure water flow velocity, and combined with water level data, can accurately calculate river cross-section flow. Optical rain gauges are deployed on the ground in key catchment areas, using infrared sensing principles to measure instantaneous rainfall intensity, providing a high-precision ground truth reference for correcting radar surface rainfall. These devices transmit data back to the data center in real time via IoT communication protocols such as 4G / 5G or NB-IoT, forming a multi-dimensional physical monitoring network.

[0033] like Figure 2 As shown, according to one aspect of this application, acquiring multi-source monitoring data of a river-pipeline system includes:

[0034] Acquire satellite remote sensing data, UAV aerial survey data, ground sensor data, and social media data;

[0035] Feature-level fusion of satellite remote sensing data, UAV aerial survey data, ground sensor data, and social media data is used to establish a spatiotemporal distribution model of rainfall and flood elements.

[0036] Spatiotemporal distribution data of precipitation intensity, river water level and flow velocity, and ground water depth are extracted from the spatiotemporal distribution model of rainfall and flood elements and used as multi-source monitoring data.

[0037] In this embodiment, the data sources are further expanded. In addition to the ground fixed sensors, the system also accesses satellite remote sensing data, unmanned aerial vehicle photogrammetry data for high-precision terrain or key area waterlogging monitoring, and social media data. Among them, the satellite remote sensing data can be multispectral satellite images for large-scale waterlogging identification; the social media data can be waterlogging text or picture information with location tags. Preferably, a deep learning fusion algorithm is used for feature-level fusion. The deep learning fusion algorithm uses a convolutional neural network (CNN) or a long short-term memory network (LSTM) model to extract and fuse the features of the above-mentioned multi-source heterogeneous data. For example, the spatial features of radar echoes can be extracted using CNN, combined with the time series features of ground rain gauges, to train a rainfall estimation model and generate high-precision gridded precipitation data. Or use target detection algorithm to identify water depth from social media pictures, and cross-verify with model simulation results. The finally established rain and flood element spatio-temporal distribution model is a digital field containing global precipitation, water level, flow rate and waterlogging information. The standardized spatio-temporal distribution data extracted from it is used as high-quality input for subsequent correction processes, solving the problem of incomplete or biased information from a single data source.

[0038] In a specific embodiment, the deep learning fusion algorithm uses a multi-task learning framework. Specifically, the algorithm includes: using a convolutional neural network (CNN) to extract spatial features from satellite remote sensing images and unmanned aerial vehicle photogrammetry images, and using a long short-term memory network (LSTM) to extract time series features from ground sensor time series; After splicing the feature vectors extracted from each channel, the features are fused through a fully connected layer to output a unified rain and flood element spatio-temporal distribution tensor; Using the high-precision ground observation data in historical rainfall events as labels, using the mean square error loss function, using the Adam optimizer for end-to-end training, the learning rate is set to 0.001, and the training rounds are 100 rounds.

[0039] Optionally, before the spatio-temporal alignment processing of the multi-source data, it also includes cleaning the multi-source monitoring data: using the interquartile range method to identify and eliminate statistical outliers in the multi-source monitoring data to obtain monitoring data after eliminating outliers; using the moving average method to smooth the monitoring data after eliminating outliers, filling in the missing time nodes, to obtain cleaned multi-source monitoring data.

[0040] Specifically, for a time series data of a certain sensor, the first quartile Q1 and the third quartile Q3 are calculated, and the interquartile range IQR = Q3-Q1 is obtained. The threshold for judging outliers is constructed as [Q1-1.5·IQR, Q3+1.5·IQR], and any data point outside this range is marked as a statistical outlier and removed. It can effectively filter out the sharp pulses caused by equipment failure. For the holes left after removing outliers or missing points in the original data, a sliding average method is used for filling and smoothing. For example, the arithmetic mean of the M valid data points before and after the missing point can be taken as the filling value. Not only the data continuity is repaired, but also the random measurement noise is smoothed, ensuring the data quality of the input model and eliminating the interference of observation noise on the correction.

[0041] In further implementations, before generating the simulation sequence corresponding to the processed multi-source monitoring data, a parameter sensitivity screening step is further included:

[0042] Based on the model mechanism, an initial uncertainty parameter set is determined from the full parameters of the river-pipe network coupled model;

[0043] The basic effect value of each parameter in the initial uncertainty parameter set is calculated by using the Morris screening method, and non-sensitive parameters with a sensitivity lower than a preset threshold are removed according to the basic effect value;

[0044] The variance decomposition of the remaining parameters in the candidate parameter set is performed by using the Sobol global sensitivity analysis method, and the first-order sensitivity index and the total effect index are calculated;

[0045] Parameters with a first-order sensitivity index or a total effect index higher than a preset standard are determined as a key parameter set; wherein the target parameter subset in the parameter correction strategy is selected from the key parameter set.

[0046] In this embodiment, the offline preparation work before online correction is described, that is, the massive model parameters are reduced to core key parameters through two-stage sensitivity analysis. Specifically, according to the principles and experience of hydraulics, an initial uncertainty parameter set that may have errors is determined, such as the Manning coefficient n pipe with a value range of [0.010, 0.020]. A low-cost Morris screening method is used for preliminary screening. By constructing a random trajectory in the parameter space, the basic effect of each parameter change on the model output is calculated:

[0047] EE i = [Y(X1,..., X i + Δ,..., X k ) - Y(X)] / Δ;

[0048] Wherein EE iis the basic effect value of the ith parameter; X is a random sampling point in the parameter space; Δ is the step size of parameter i; Y(X) is the output value of the model under the parameter combination X; k is the total number of parameters. The mean μ and standard deviation σ of the basic effect are calculated. If the μ of a parameter is significantly less than the preset threshold, it means that the parameter has little effect on the model result, and is determined as a non-sensitive parameter and is removed. The Sobol global sensitivity analysis method with higher precision is used for the remaining parameters after preliminary screening. Based on the variance decomposition principle, the contribution proportion of each parameter to the total variance of the model output is quantified. The specific calculation formula involves the first-order sensitivity index S i =V i / V, wherein V i is the variance caused only by parameter i, and V is the total variance; and the total effect index S Ti , which includes the main effect of parameter i and the sum of the interaction with other parameters. Exemplarily, S Ti =1 - V ~i / V; wherein S Ti is the total effect index of the ith parameter; V ~i is the sum of the variances caused by all parameters except parameter i. The parameters with S Ti greater than a preset standard (for example, 0.05) are selected into the key parameter set. The parameter funnel is constructed in this embodiment, so that the algorithm only selects in the key parameter set which has the greatest effect on the result during subsequent online correction, reducing the optimization dimension and improving the correction efficiency.

[0049] In some optional embodiments, the system further comprises a sensor health monitoring and fault tolerance processing mechanism: the update time stamp of each sensor data is monitored in real time, and if a sensor does not update data for more than a preset time threshold, it is determined that the communication is interrupted or the device fails; for the faulty sensor, a spatial interpolation method is used to estimate and fill in the data of the adjacent normal sensor; if the number of faulty sensors exceeds a preset proportion, the system automatically switches to a degraded operation mode, only uses trusted data sources for coarse and fine correction, and sends an alarm to the operation and maintenance personnel.

[0050] In a preferred embodiment, a correction trigger system is further included, which comprises:

[0051] Periodic trigger monitoring: generating a trigger signal according to a preset time period;

[0052] Event-triggered monitoring: accessing the rainfall data stream in the processed multi-source monitoring data in real time, and generating a trigger signal when the rainfall intensity is identified to exceed a preset threshold;

[0053] Deviation trigger monitoring: comparing the processed multi-source monitoring data with the simulation sequence in a preset sliding time window, and generating a trigger signal when the error index between the two continuously exceeds a preset multiple of the historical background value.

[0054] When any trigger signal is detected, go to the step of calculating deviation features.

[0055] Periodic trigger monitoring: generate trigger signal according to preset time period;

[0056] Event trigger monitoring: access rainfall data stream in processed multi-source monitoring data in real time, and generate trigger signal when identifying that rainfall intensity exceeds preset threshold;

[0057] Deviation trigger monitoring: compare processed multi-source monitoring data with simulation sequence in a sliding time window, calculate root mean square error between them as error indicator, and generate trigger signal when error indicator continuously exceeds preset multiple of pre-stored historical background value;

[0058] When any trigger signal is detected, execute the step of calculating deviation features and subsequent steps.

[0059] In the embodiment, three complementary trigger modes are defined, covering different business scenarios. Periodic trigger monitoring is a passive and regular maintenance mechanism. The system has built-in timing tasks, such as setting to perform correction once every 24 hours or every Monday morning, so that even in the case of no rain or little data fluctuation, the model can periodically absorb the latest system state information (such as seasonal base flow changes) to maintain basic accuracy. Event trigger monitoring is a feedforward response mechanism. The system monitors rainfall data in real time, and when it detects that the rainfall intensity I> preset threshold I threshold (e.g. 0.5 mm / hour), it means that a rainfall event is about to impact the drainage system, and a trigger signal is generated immediately. This allows the model to quickly adjust parameters at the initial stage of a key hydrological event, preparing for the upcoming flood peak simulation. Deviation trigger monitoring is a feedback error correction mechanism. During operation, the system continuously calculates the error indicator of the simulated water level and the observed water level in the sliding window, such as the root mean square error RMSE in the current sliding window. Maintain the historical background value sequence of the error indicator, such as the RMSE statistical distribution of the past week, and calculate the 95% quantile as the baseline value RMSE base . When real-time RMSE> K*RMSE base and this condition lasts for N calculation steps, it is determined that the model has performance degradation or abnormal drift, and a trigger signal is generated, where K is the amplification factor, for example, K=1.5. Deviation trigger monitoring can capture deviations caused by non-rainfall factors, such as sudden gate scheduling or sensor drift.

[0060] In some optional implementations, to avoid conflicts or frequent system restarts caused by multiple trigger signals, priority arbitration and cooldown period logic can be introduced. Specifically, the priority order is set as: event trigger > deviation trigger > periodic trigger. That is, when a rainfall event occurs, the system responds to the event trigger signal first, ignoring concurrent periodic checks. Furthermore, a cooldown period T is set. cool T after a correction action is completed cool Within a given timeframe, such as 2 hours, the system blocks all deviation and periodic trigger signals unless a new heavy rainfall event is detected (event triggering is not limited by a cooldown period). This avoids misinterpreting fluctuations in the transition process after parameter adjustments as new deviations, ensuring the stability of the online calibration system.

[0061] The triggering mechanism in this embodiment is a preliminary step in the calibration process. In actual operation, the system continuously performs trigger monitoring; once a trigger signal is detected, a complete adaptive calibration process begins. This embodiment, by establishing a triple triggering system including periodicity, event, and deviation, can balance the regularity of daily maintenance with the response speed to emergencies. Furthermore, through priority arbitration and cooldown period design, it ensures system stability and resource utilization efficiency.

[0062] Within a preset sliding time window, a residual signature containing multi-dimensional physical attributes is constructed based on the processed multi-source monitoring data and simulation sequence as a deviation feature. The residual signature includes at least: a mass conservation non-closed quantity calculated based on the integral of the total inflow and total outflow of the system extracted from the processed multi-source monitoring data; a spatial clustering feature calculated based on the distribution of deviation measurement points in the pipeline topology; and a deviation time series feature calculated based on the delay of the deviation occurrence time relative to the rainfall start time.

[0063] In this embodiment, feature vectors characterizing the ill-conditioned physical mechanisms of the system are extracted, i.e., residual signatures. The sliding time window is a computational interval W that shifts forward over time, with a length T. W Based on the system response time setting, for example, 1 hour. Within this window, the system calculates not only point-to-point errors, but also overall macroscopic indicators of the system.

[0064] Specifically, the mass conservation non-closure quantity is a key indicator for determining the deviation of external input (rainfall), representing the difference in water quantity inflow and outflow within a time window. The specific formula for calculating the mass conservation non-closure quantity can be expressed as:

[0065] ΔV(W) =Σ t=tstart tend (Q in (t) - Q out (t)) *Δt - (S end - Sstart );

[0066] where ΔV(W) is the mass conservation non-closure in window W; tstart is the start time of the window; tend is the end time of the window; Q in (t) is the total inflow of the system at time t, including rainfall runoff and external inflow; Q out (t) is the total outflow of the system at time t, including pump outflow and overflow; Δt is the sampling interval, S end and S start are the total storage of the system at the end and start of the window, respectively, including pipe network storage, river storage and surface water storage. On this basis, in order to eliminate the influence of rainfall level, the normalized mass non-closure index is further calculated:

[0067] nMB = ΔV(W) / max(V in , V min );

[0068] where nMB is the normalized mass non-closure index; V in is the total inflow volume of the window, V min is the minimum volume threshold to prevent the denominator from being zero. If nMB presents a negative value, it indicates that the input rainfall is too small.

[0069] The spatial clustering feature is used to distinguish between local faults and global deviations. Its calculation is based on the concept of connected components in graph theory. Specifically, the set of abnormal measurement points whose absolute values of residuals exceed the threshold is identified, and it is determined whether these abnormal points are connected based on the topological relationship of the pipe network. The spatial clustering index is defined as:

[0070] γ= 1 - (N cc / N total );

[0071] where N cc is the number of connected components of abnormal points, and N total is the total number of abnormal points. If γ tends to 1, it indicates that the abnormal points are highly clustered, such as a blocked main pipe; if γ tends to 0, it indicates that the abnormal points are dispersedly distributed in space, which may be a global rainfall deviation or a random fault of the sensor.

[0072] The deviation timing feature is used to distinguish between input deviation and underlying surface parameter deviation. Specifically, the rainfall start time is defined as t rain , and the time when the deviation is significantly present is defined as t onset , i.e., the time when the residual first exceeds the threshold continuously. The normalized timing index is calculated as:

[0073] τ= (t onset - t rain ) / T response ;

[0074] wherein τ is the normalized bias timing feature; T response is the characteristic response time of the catchment. If τ is small, it indicates that the bias appears at the beginning of the rainfall, which tends to be the rainfall measurement error; if τ is large, it indicates that the bias is gradually accumulated in the confluence process, which tends to be the impermeability or roughness parameter error of the underlying surface.

[0075] As Figure 3 shown in the further embodiment, the parameter correction strategy is determined, comprising:

[0076] The residual signature is input into a pre-constructed bias attribution discriminator to identify the current system bias type; the bias type at least includes rainfall input bias, underlying surface parameter bias, pipe network water transport capacity bias, river water transport capacity bias, and coupling interface bias.

[0077] In the present embodiment, the bias attribution discriminator can map the high-dimensional residual signature to a specific physical cause. The bias type is a physical classification of the error source. Specifically, the rainfall input bias refers to the systematic overestimation or underestimation of radar or rain gauge data; the underlying surface parameter bias refers to the inaccuracy of the impermeability, depression storage, and other parameters of the sub-catchment; the pipe network water transport capacity bias refers to the error of the pipe Manning coefficient or the existence of unknown siltation and collapse; the river water transport capacity bias refers to the error of the river section geometry or roughness parameter; the coupling interface bias refers to the inconsistency between the state of the gate, overflow weir, or orifice at the connection between the pipe network and the river and the model setting.

[0078] According to the identified bias type, a parameter correction strategy is generated; wherein the parameter correction strategy limits the target parameter subset to a mutually exclusive parameter group corresponding to the bias type, and limits the correction space range to a physical region associated with the bias type.

[0079] As Figure 4 shown in the further embodiment, the parameter correction strategy is determined, comprising:

[0080] The mass conservation non-closure quantity and the bias timing feature in the residual signature are extracted; when the normalized value of the mass conservation non-closure quantity exceeds a preset significance threshold, and the bias timing feature indicates that the delay between the bias occurrence time and the rainfall start time is less than a preset time threshold, the bias type is determined to be rainfall input bias.

[0081] In other words, the mass conservation non-closure quantity and the bias timing feature in the residual signature are extracted; the mass conservation non-closure quantity is normalized to obtain a normalized value; when the normalized value exceeds a preset significance threshold, and the bias timing feature indicates that the delay between the bias occurrence time and the rainfall start time is less than a preset time threshold, the bias type is determined to be rainfall input bias.

[0082] Specifically, if the model not only mismatches the water level but also exhibits an imbalance in water volume (mass non-conservation), and this imbalance occurs at the very beginning of rainfall, the only explanation is that the input water volume is incorrect. This is because pipe or channel parameters only alter water distribution and flow velocity, not create or eliminate water volume out of thin air (on a short timescale ignoring infiltration errors). The significance threshold can be set according to the actual system characteristics, typically ranging from 0.10 to 0.20. For example, a preset significance threshold can be set to 0.15, representing a 15% water volume error, and a time threshold can be set to 0.2, meaning the error occurs within the first 20% of the response time. When these conditions are met, the discriminator outputs the rainfall input deviation as a conclusion, guiding subsequent adjustments to only correct the rainfall data, thus avoiding erroneous adjustments to pipe parameters to manipulate the water level.

[0083] In some possible embodiments, the discriminator also includes rules for other types of deviations. For example, for coupling interface deviations, the system calculates the signed product of the interface network-side residual and the river-side residual, i.e.:

[0084] φ flip =sign(r pipe )*sign(r river );

[0085] Where φ flip The sign-to-flip index is used, and sign() is the sign function; r pipe The residual water level on the pipeline side of the coupling interface; r river This represents the residual water level on the riverside side of the coupling interface. If φ flip A negative value indicates an overestimation on one side and an underestimation on the other, and a high spatial clustering characteristic γ near the interface indicates an error in the physical parameters of the coupling interface (such as connectivity or weir height). Regarding deviations in the water conveyance capacity of the pipeline network, if the residual water level in the network is monitored to increase monotonically with the rise of the observed water level (i.e., correlation coefficient ρ > 0.8), it suggests that the pipeline's flow capacity is limited, indicating excessive roughness or siltation.

[0086] This embodiment constructs physically meaningful residual signatures and inputs these features into an attribution discriminator, thereby achieving accurate diagnosis of the source of deviation. By locking specific parameter subsets through the correction action graph for targeted correction, it avoids the distortion of the physical meaning of parameters caused by different parameters having the same effect.

[0087] In a specific numerical case, a city's drainage system experienced a short-duration heavy rainfall event from 08:00 to 09:00 on July 15, 2025. The system has 12 water level monitoring points. Rainfall began at 08:00, peaked at 08:30 (rainfall intensity 15 mm / h), and ended at 09:00. Based on multi-source monitoring data, the total inflow volume V within a one-hour window was calculated. in= radar inversion rainfall x catchment area = 12mm x 5km 2 = 60000 m 3 ; total outflow volume V out = pump station discharge + spillage = 45000 + 8000 = 53000 m 3 ; system storage change AS = S end - S start = 2500 m 3 . AV(W) = V in - V out - AS = 60000 - 53000 - 2500 = 4500 m 3 ; nMB = AV(W) / V in = 4500 / 60000 = 0.075 = 7.5%. Rainfall start time t rain = 08:00; deviation significant occurrence time t onset = 08:10, at which time the water level residuals of multiple measurement points begin to continuously exceed the threshold value; watershed characteristic response time T response = 40 minutes; tau = (08:10 - 08:00) / 40 min = 10 / 40 = 0.25. Total number of abnormal measurement points N total = 5 (of the total 12 measurement points), number of connected components formed by abnormal measurement points N cc = 2 (based on the topology of the pipe network, the 5 points form 2 independent clusters); gamma = 1 - N cc / N total = 1 - 2 / 5 = 0.6. Input the residual signature Phi = (nMB = 0.075, tau = 0.25, gamma = 0.6) obtained above into the deviation attribution discriminator: check condition 1: nMB = 0.075 < 0.15 (significance threshold), does not satisfy the quality closure significant condition; check condition 2: tau = 0.25 > 0.2 (time threshold), the deviation occurrence time is relatively late; check condition 3: gamma = 0.6 < 0.7 (clustering threshold), the spatial distribution is relatively dispersed. Since the quality conservation closure amount does not exceed the significance threshold, and the deviation occurrence time is relatively delayed from the rainfall start, it is determined that the current deviation type is the underlying surface parameter deviation, that is, the impervious rate or the initial depression storage parameter of the sub-catchment area may have errors. Query the correction action atlas to map the underlying surface parameter deviation to the target parameter subset: the impervious rate and the depression storage of the sub-catchment area upstream of the abnormal area. Lock the rainfall correction coefficient and the pipe Manning coefficient unchanged.

[0088] In an embodiment of the present application, according to the identified deviation type, a parameter correction strategy is generated, comprising:

[0089] querying the preset correction action atlas, mapping the deviation type to a corresponding parameter group to be corrected;

[0090] determining the parameter group to be corrected obtained through the mapping as a target parameter subset, and keeping the values of other parameter groups in the river-pipe network coupling model unchanged except the target parameter subset;

[0091] The mapping relationship defined by the correction action atlas at least includes: mapping the rainfall input deviation to a rainfall correction parameter group, mapping the pipe network water delivery capacity deviation to a pipe network roughness and pipe segment deposition parameter group, and mapping the coupling interface deviation to a coupling interface connectivity coefficient and loss coefficient parameter group.

[0092] In the embodiment, the correction action atlas is a bridge connecting diagnosis and treatment, and embodies the decoupling idea. Through the correction action atlas, the system enforces exclusive update of parameters, and mathematically eliminates the multiple solution. Specifically, the correction action atlas can be embodied as a lookup table. When the input is the rainfall input deviation, the mapped parameter group only includes the radar correction factors α and β, at this time all the Manning coefficients and impermeability rates of the entire network are regarded as constants and are prohibited from being modified. When the input is the pipe network water delivery capacity deviation, the mapped parameter group only includes the Manning coefficients n pipe and the pipe diameter correction coefficient (representing deposition) of the pipes within the R radius upstream of the abnormal area, at this time the rainfall data is regarded as true value and is prohibited from being modified. When the input is the coupling interface deviation, the mapped parameter group only includes the physical property parameters of the specific interface. Through the strong constraint of keeping the values of other parameter groups unchanged, the correction result has physical interpretability, and parameter drift of the model for fitting data is prevented.

[0093] In a possible embodiment, the river-pipe network coupling model adopts parameterized modeling for the coupling interface between the pipe network and the river; the parameterized modeling introduces connectivity coefficients and energy loss coefficients as parameters to be corrected; wherein the connectivity coefficients represent the proportion of the effective water area of the coupling interface, and the energy loss coefficients represent the local head loss when water flows through the coupling interface; when the parameter correction strategy involves the coupling interface deviation, the connectivity coefficients and the energy loss coefficients are included in the target parameter subset.

[0094] Specifically, in a traditional stormwater management model (SWMM) or InfoWorks model, the outlet is usually simplified as free outflow or fixed water level boundary. The embodiment establishes an interface physical model. Taking orifice outflow as an example, its flow formula is modified as:

[0095] Q e =κ* A design * sqrt(2g * (ΔH -ζ* v 2 / 2g))

[0096] wherein Q e is the exchange flow through the coupling interface; A design is the design overwater area of the interface; g is the gravity acceleration; ΔH is the pressure difference between the pipe network water level and the river water level; κ is the to-be-corrected connection coefficient, the value range of which is [0, 1], which reflects the opening degree of the flap valve or the blockage degree of the rainwater inlet; ζ is the to-be-corrected energy loss coefficient, which reflects the local water head loss when the water flow enters and exits the interface; v is the average flow velocity through the interface. When the system is diagnosed as interface deviation, the optimization algorithm will be specifically optimized for the two parameters (κ, ζ) to accurately reproduce complex hydraulic phenomena such as jacking, backflow or local water blockage.

[0097] According to one aspect of the present application, the method further comprises:

[0098] According to the correction space range defined in the parameter correction strategy, a target function containing physical constraints is constructed.

[0099] Preferably, the construction of the target function containing physical constraints comprises:

[0100] A data fitting error term is constructed to represent the statistical error between the simulation sequence and the processed multi-source monitoring data;

[0101] A physical consistency penalty term is constructed, which is positively correlated with the mass conservation non-closure amount of the system within the correction time window;

[0102] The target function is generated by weighted sum of the data fitting error term and the physical consistency penalty term.

[0103] That is, based on the difference between the simulation sequence and the processed multi-source monitoring data, the root mean square error is calculated to construct a data fitting error term for representing the statistical error between the simulation sequence and the processed multi-source monitoring data; based on the simulation sequence, the water volume balance difference of the system within the correction time window is calculated to construct a physical consistency penalty term, which is positively correlated with the mass conservation non-closure amount represented by the water volume balance difference; the target function is generated by weighted sum of the data fitting error term and the physical consistency penalty term.

[0104] In this embodiment, a mathematical target for driving automatic optimization is constructed. The target function J(θ) not only pursues data fitting, but also pursues physical self-consistency. The specific formula can be expressed as:

[0105] J(θ) = w1 * RMSE H (θ) + w2 *RMSE Q (θ)+λ MB *|ΔV(W, θ)|+λ reg|| θ - θ0||2 2 ;

[0106] where J(θ) is the total objective function value; θ is the target parameter subset; RMSE H and RMSE Q are the root mean square errors of water level and flow, representing the data fitting error term; w1 and w2 are the weights of water level and flow terms, respectively; ΔV(W, θ) is the mass conservation non-closure calculated by the model under the parameter θ, representing the physical consistency penalty term; λ MB is the physical consistency penalty coefficient; λ reg is the regularization coefficient; θ0 is the initial prior value of the parameter; λ reg || θ - θ0||2 2 is the prior regularization term, used to prevent the parameter from deviating too much from the initial value. The size of the coefficient λ MB determines the importance of mass conservation. By minimizing the objective function, the algorithm is forced to meet the water balance constraint while searching for the optimal parameter, eliminating those pseudo-solutions that fit well but are not water-consistent.

[0107] The optimization algorithm is used to optimize the objective function to update the parameter values in the target parameter subset.

[0108] Specifically, the system can choose different optimization algorithms according to different trigger types. In the preferred embodiment, for deviation trigger scenarios that require high real-time performance, usually requiring completion within minutes, sequential data assimilation algorithms such as Ensemble Kalman Filter (EnKF) can be used to update the state and parameters in real time as observation data arrives. For event-triggered or periodic-triggered deep correction scenarios that allow longer computation time (e.g., hours), global heuristic optimization algorithms such as Differential Evolution (DE) or Particle Swarm Optimization (PSO) can be used. Iterative search is performed in the parameter space until the objective function J(θ) converges or the maximum number of iterations is reached, and the optimal parameter θ* is finally output.

[0109] In further embodiments, after updating the parameter values in the target parameter subset, the method further comprises:

[0110] starting a sandbox environment independent of the online business model, loading the updated target parameter subset in the sandbox environment and running the model for smoke testing; when the smoke testing is passed, switching the business traffic to the model instance loaded with the updated parameters through dynamic link library reloading or API gateway switching, completing the dynamic synchronization of the model.

[0111] It can also be said that the updated target parameter subset is loaded in the pre-configured sandbox environment independent of the online service model, and the model is run for smoke testing to obtain test results; when the test results meet the preset accuracy threshold, it is determined that the smoke test is passed, and the business traffic is switched to the model instance loaded with the updated parameters through dynamic link library reloading or application program interface gateway switching, completing the dynamic synchronization of the model.

[0112] In this embodiment, the hot update mechanism of an industrial-grade software system is described. Modifying parameters directly in online services can cause service interruption or computing crashes, so a sandbox environment is introduced. The sandbox is a computing container isolated from the production environment. The system loads new parameters into a copy of the model in the sandbox and runs a quick simulation, such as 1 hour of past data, for smoke testing to verify whether the model converges normally, whether the water level appears non-physical shock, and whether the accuracy index meets the standard. Smoke testing can be used to verify whether the updated model can run normally and produce output results that meet the basic accuracy requirements. Preferably, the passing standard of smoke testing can be set as: Nash efficiency coefficient NSE > 0.7 and peak relative error < 20%. Only when the test passes, the system performs the switching action. Dynamic link library reloading refers to reloading the computing core DLL without restarting the main process; API gateway switching refers to directing the routing of external query requests to the new model instance container. This makes the business users unaware of the model update process, and realizes 7x24 hours uninterrupted service.

[0113] In an embodiment of the present application, the method further comprises a knowledge base closed-loop updating step:

[0114] The deviation characteristics obtained by this correction calculation, the generated parameter correction strategy, and the updated target parameter subset are packaged as a correction case record; the correction case record is stored in the model parameter knowledge base, and the correction case record in the model parameter knowledge base is used for subsequent training and updating of the deviation attribution discriminator to improve the attribution accuracy of the deviation attribution discriminator.

[0115] Specifically, each correction is not only a parameter update, but also a valuable experience accumulation. The system encapsulates this process as a standardized case record, whose data structure can be represented as Tuple = {Time, Feature_Vector (Φ), Bias_Type (b*), Optimal_Params (θ*)}, where Tuple is the correction case record, Time is the timestamp of this correction, Feature_Vector (Φ) is the bias feature vector, Bias_Type (b*) is the bias type label, and Optimal_Params (θ*) is the optimal parameter set obtained by this correction. These records are stored in a relational database or a time series database, forming a model parameter knowledge base. As the running time passes and the case library becomes richer, the system can retrain the bias attribution discriminator using machine learning algorithms such as random forests or gradient boosting trees. For example, Feature_Vector is used as input and Bias_Type is used as label for supervised learning, so that the discriminator can achieve higher and higher accuracy in attributing complex nonlinear features, realizing the evolution from rule-based diagnosis to data-driven intelligent diagnosis.

[0116] According to another aspect of the present application, a cross-validation strategy is used to evaluate the correction effect before the updated parameters are put into use; the cross-validation strategy includes time cross-validation, space cross-validation, and event cross-validation.

[0117] In this embodiment, three verification methods to prevent model overfitting are described. Time cross-validation refers to dividing a long-duration rainfall event into a correction segment (the first 2 / 3 period) and a verification segment (the last 1 / 3 period). The parameter is inverted using the correction segment data, and the water level process of the verification segment is simulated using the parameter. If the Nash efficiency coefficient (NSE) of the verification segment is significantly lower than that of the correction segment, it indicates that the parameter has a time adaptability problem. Spatial cross-validation refers to dividing the monitoring sensors into a correction set (e.g., 80% of the stations) and a verification set (the remaining 20% of the stations, and the positions are independent). Only the data of the correction set is used to construct the objective function for parameter optimization, and the fitting error of the model on the verification set stations is evaluated. It can effectively detect whether the model has a spatial overfitting phenomenon. Event cross-validation refers to using the parameters obtained by correcting a typical rainfall event A to simulate another rainfall event B with completely different hydrological characteristics (e.g., from short-duration heavy rainfall to long-duration continuous rain). Only when the parameters perform stably in multiple events, the parameters are determined to have update eligibility.

[0118] Optionally, it also includes strict version control of model files and parameter sets, generating a new version number each time an update is made, and retaining historical versions for rollback.

[0119] Specifically, each parameter update operation triggers a version commit. The system records a complete version log, including: version number, update timestamp, trigger reason, bias attribution conclusion, updated parameter subset snapshot, and corresponding validation indicators (NSE, RMSE). The system retains all historical version model configuration files and parameter databases in storage media. When the online model experiences unexpected performance degradation or anomalies, operations personnel or automatic monitoring scripts can roll back the system state to the last stable version within seconds by tracing back the version number, thereby minimizing business risks.

[0120] In an embodiment of the present application, the quality conservation non-closure quantity is calculated, including: reading the water level simulation sequence, the flow simulation sequence and the time axis data, performing resampling of a uniform sampling interval and timestamp alignment on each sequence to obtain the aligned water level simulation sequence and the aligned flow simulation sequence. In the alignment process, if there are missing points in the simulation output, linear interpolation is used to fill in the missing points, and a simulation missing marker is generated for subsequent quality control. Read the rainfall driving data (rain gauge rainfall or radar rainfall field) and the runoff inflow sequence output by the model internal runoff generation module; if there is external inflow or upstream inflow, read the external inflow sequence. Superimpose the above inflows on the same time axis to obtain the total inflow sequence of the system. Read the outflow simulation sequence of the outfall, the outflow simulation sequence of the pump station, the overflow outflow simulation sequence and other possible outflow items, and perform deduplication processing according to the model topology relationship to avoid the same outflow being counted repeatedly at multiple interfaces. Superimpose the deduplicated outflow items to obtain the total outflow sequence of the system. Read the pipe network state quantity, the river state quantity and the optional surface water storage state quantity, calculate the system water storage quantity at each time according to the partial water storage method to obtain the system water storage quantity sequence. When the model cannot output sufficient state quantities, read the preset equivalent water storage parameters and representative water level simulation sequence, and use the equivalent water storage method to obtain the system water storage quantity sequence. Numerically integrate the total inflow sequence of the system and the total outflow sequence of the system in a sliding window, and combine the system water storage quantity sequence at the beginning and end of the window to calculate the mass conservation non-closure quantity of the window, and obtain the mass conservation non-closure quantity. At the same time, calculate the inflow volume of the window, and normalize the non-closure quantity to obtain the normalized mass non-closure indicator. Read the simulation missing marker and the availability information of various inflow and outflow items, and generate a mass conservation reliability marker according to the missing proportion, the absence of key outflow items and other conditions.

[0121] In another embodiment of the present application, the residual signature is constructed, including: reading the water level observation sequence, optional flow observation sequence and time axis data, performing resampling of uniform sampling interval and timestamp alignment, obtaining the aligned water level observation sequence and optional aligned flow observation sequence. Based on rules such as physical range, mutation amplitude, long-time constant, communication packet loss, the observation is marked for quality, and the observation quality mark is obtained. The aligned water level simulation sequence and the aligned water level observation sequence are subtracted point by point to obtain the water level residual sequence; if there is flow observation, the aligned flow simulation sequence and the aligned flow observation sequence are subtracted point by point to obtain the flow residual sequence. For the period with poor observation quality, the residual shielding or weight marking is performed according to the observation quality mark to generate the residual validity mark for the subsequent credibility control of feature statistics. In the sliding window, the water level residual sequence is statistically summarized to obtain the water level residual mean feature and the water level residual root mean square feature; the available flow residual sequence is also calculated to obtain the corresponding statistical quantity, and the flow residual root mean square feature is obtained. The error intensity ratio of pipe network and river is calculated at the same time, and the pipe-river error comparison feature is obtained to distinguish the pipe network capacity deviation and the river capacity deviation. The rainfall driving data is read and the rainfall starting time is located, and then the residual significant occurrence time is calculated based on the water level residual sequence in the window, so as to calculate the deviation occurrence relative timing index and obtain the deviation timing feature.

[0122] In the sliding window, the aligned water level observation sequence and the water level residual sequence are read, the correlation between the observed water level and the residual is calculated to represent the capacity-limited monotonic amplification effect, and the pipe network monotonicity feature and the optional river channel monotonicity feature are obtained. In the sliding window, the abnormal measurement point set is identified according to the water level residual sequence and the threshold rule, and the pre-defined measurement point adjacency relationship data (which can be based on spatial proximity or pipe network topological adjacency) is read, the connectivity clustering degree of the abnormal measurement point is calculated, and the spatial clustering feature is obtained. Read the pre-defined interface association data (used to divide the measurement points into the upstream side and the downstream side of the interface), and calculate the mean value sign of the water level residual sequence on the upstream side and the downstream side in the sliding window respectively, and judge whether the signs are opposite; if they are opposite and the residual amplitude on both sides exceeds the threshold, record the candidate interface. Output the interface sign flip feature and the candidate interface set, which is used for pure rule discrimination of coupled interface deviation and directly drives the subsequent correction spatial range limitation. When there are both rain gauges and radar rainfall, read the rainfall driving data and calculate the normalized index of cumulative rainfall difference in the sliding window, and obtain the rainfall consistency feature. When any type of rainfall data is missing, generate a rainfall consistency availability marker, and the subsequent attribution rule automatically skips or de-weights this item to avoid introducing isolated unreliable features. The water level residual mean feature, the water level residual root mean square feature, the flow residual root mean square feature, the pipe river error comparison feature, the deviation time sequence feature, the pipe network monotonicity feature, the river channel monotonicity feature, the spatial clustering feature, the interface sign flip feature, the rainfall consistency feature, the normalized quality non-closed index, and various availability and reliability markers are structured and packaged to obtain the residual signature feature vector.

[0123] According to one aspect of the present application, the river-pipe network coupling model adaptive correction method based on multi-source data driving can also be:

[0124] Step S1, construct a rain flood element model of multi-source data, use a deep learning fusion algorithm to process satellite remote sensing, unmanned aerial vehicle surveying, ground sensors, social media data and other multi-source data, and establish a multi-dimensional rain flood element spatio-temporal distribution model including precipitation intensity, river water level and flow rate, ground water area, water depth, and water duration.

[0125] Specifically, by deploying Internet of Things sensing terminals in the drainage pipe network system and the river system, first-type in-situ monitoring data is collected at a preset frequency. The sensing terminals include but are not limited to pressure water level gauges, radar flowmeters, ultrasonic flowmeters, photoelectric rain gauges, and water quality multi-parameter sensors (pH, COD, ammonia nitrogen, etc.). The data is transmitted to the data center through an Internet of Things communication protocol (such as 4G / 5G, NB-IoT, LoRa). Using remote sensing technology and geographic information system (GIS) data, a digital elevation model (DEM) of the basin or region is constructed to accurately describe the terrain characteristics, including the elevation, slope, and slope direction of each landform unit in the basin; a statistical model (such as the interquartile range (IQR) method) or a machine learning algorithm (such as Isolation Forest) is used to automatically identify and mark outliers in the data. For outliers, a moving average method can be used to smooth the data before and after the time point, or the outliers can be directly marked as missing and subsequently interpolated. The time of all data is converted to the same standard time base, and resampling is performed according to the minimum time step calculated by the model. The areal rainfall data obtained by meteorological radar inversion and the waterlogging data extracted by remote sensing are matched to the center points of the coupled model subcatchment (SWMM model) or calculation grid (river model) through spatial interpolation methods. The data of fixed-point sensors, such as the liquid level meter under a certain manhole cover, are matched to the model unit to which they belong through spatial correlation by the geographic information system (GIS).

[0126] A strategy combining data-level fusion and feature-level fusion is adopted to generate the final driving data set. For hydrological data: the spatial data obtained by interpolation is locally corrected based on the high-precision in-situ sensor data. For example, the true data of the rain gauge is used to correct the deviation of the radar estimated rainfall product. For waterlogging data: the large-scale waterlogging information identified by remote sensing and the fixed-point waterlogging information obtained by video analysis, as well as the waterlogging depth simulated by the pipe network model, are cross-verified and fused to generate a more reliable urban waterlogging field data set, which can be used as key verification data. A standard multi-dimensional data table with spatio-temporal index is generated, including but not limited to: time, model unit, monitoring rainfall, radar rainfall, water level monitoring value, flow monitoring value, remote sensing waterlogging depth, etc., providing strong data support for subsequent model correction.

[0127] Step S2, river-pipe network coupled hydrodynamic model establishment and uncertainty parameter identification: based on river data and pipe network data, a bidirectional tightly coupled river-pipe network coupled hydrodynamic model is constructed; based on the local correction Morris screening method and the global Sobol method, key parameters with high sensitivity and large uncertainty to the model output results are screened out as the parameter set to be corrected; the parameters to be corrected include pipe network roughness coefficient, river Manning coefficient, impermeability, and characteristic width, etc.

[0128] Specifically, based on model mechanism, field survey report, pipe material, river section and other prior knowledge, a set of possible uncertainty parameters is initially selected, and a reasonable physical variation range is determined for each parameter, such as the Manning coefficient range of the pipe, which can be set as [0.01, 0.025]. The parameters include but are not limited to: pipe network roughness coefficient (Manning coefficient), river Manning coefficient, surface storage capacity, impermeability, and sub-catchment characteristic width. The initial parameter set is subjected to preliminary sensitivity analysis by the Morris screening method. Within the value space of the parameters, multiple trajectories are sampled and the model is run. The average absolute effect μ and the standard deviation σ of each parameter on different output variables are calculated. Parameters with low μ and σ values are determined as insensitive parameters, which are removed from the calibration parameter set to prepare for more accurate but higher computational cost analysis. The Sobol method is used to analyze the parameter subset screened by Morris. The first-order sensitivity index (i.e. the main effect) of each single parameter and the total-order sensitivity index of the interaction between parameters are accurately quantified. A large number of Quasi-Monte Carlo sampling is used to generate a parameter sample set, the model is run and the total variance of the model output is calculated. Through variance decomposition, the variance proportion contributed by each parameter is calculated.

[0129] Step S3, set an adaptive trigger correction mechanism based on multiple strategies, and establish a correction system including three correction trigger conditions of periodic trigger, event trigger and deviation trigger.

[0130] Specifically, multiple trigger condition monitors are run in parallel: periodic trigger monitor: built-in timing task scheduler, which can be configured to automatically generate trigger signals at fixed calendar periods or model running periods; event trigger monitor: real-time access to rainfall data stream, generate trigger signal when the system identifies a rainfall event intensity exceeding the preset threshold; deviation trigger monitor: the system continuously compares the simulated value (such as node water level) of the online model with the real-time monitoring value. Statistical process control method based on sliding time window is adopted: calculate the root mean square error (RMSE) or mean absolute error (MAE) of the residual error between the simulated value and the monitoring value in the window. When the above error indicators exceed K times of their historical background values for N consecutive calculation steps, it is determined that the system has a significant deviation, and a trigger signal is generated. When multiple conditions are met at the same time, the system arbitrates according to the preset priority: event trigger > deviation trigger > periodic trigger. At the same time, in order to avoid frequent triggering within a single event, the system will set a cooling period, for example, within 6 hours after a trigger and completion of correction, new periodic or deviation triggers are not responded to, but strong rainfall event triggers are excluded.

[0131] Step S4, using multi-source fusion data as the observation benchmark, the parameter set is automatically optimized and corrected by using deep learning optimization algorithm.

[0132] Specifically, for high-frequency, real-time correction (such as bias triggering), the sequential data assimilation algorithm such as ensemble Kalman filter is preferred. For global depth correction after the event or periodically, the intelligent optimization algorithm such as particle swarm optimization or differential evolution algorithm is preferred. The objective function is constructed, and the objective function J(θ) that can comprehensively reflect the fitting degree of multi-source data is constructed for optimization, wherein θ is the parameter set to be corrected:

[0133] J(θ)= w1 * RMSE(Q) + w2 *RMSE(H)+ w3 *RMSE(V);

[0134] Wherein RMSE(Q), RMSE(H), RMSE(V) are the root mean square errors between the simulated values and the monitored values of different observation variables such as flow, water level and flow rate. ω1, ω2, ω3 are weight coefficients of each variable, which can be dynamically adjusted according to the data reliability (such as sensor accuracy) and the importance of optimization target. For remote sensing and other surface data, spatial correlation coefficient or critical success index (CSI) and other indicators can be used to join the objective function. Optimization execution and constraint processing: when solving min J(θ), the optimization of all parameters must be carried out within the determined physical reasonable range; at the same time, the system monitors the optimization process, sets the maximum iteration number or the target function improvement threshold as the stopping criterion; finally, the optimal parameter vector θ* and its corresponding posterior uncertainty estimate are output.

[0135] Step S5, verifying the precision of the corrected model, and storing the correction result to the model parameter knowledge base.

[0136] Specifically, cross-validation is performed. It includes: temporal cross-validation: for the correction of a rainfall event, use the data of the first 2 / 3 period of the event for correction, and reserve the data of the last 1 / 3 period for verification. Spatial cross-validation: use 80% of the sensor data for correction, and the other 20% of the location-independent sensor data for verification. Event cross-validation: use a historical rainfall event for correction, and another completely different rainfall event for verification. Calculate multiple performance indicators on the verification data set: Nash efficiency coefficient (NSE), Kline-Gupta efficiency coefficient (KGE), peak flow relative error (PEQ), peak time error. Only when all major indicators (such as NSE and KGE) are higher than the preset acceptance standard (such as NSE>0.7, KGE>0.6), this correction is considered successful. The successful correction case is stored as new knowledge in the model parameter knowledge base composed of a relational database. Each record contains: meta information: trigger condition, rainfall event ID, timestamp, calculation time consumption. Input data: used monitoring data hash value or pointer. Output results: optimal parameter set θ, verification indicator value, posterior distribution of parameters (such as mean, variance). This knowledge base can be used for providing high-quality initial values for new corrections in the future, or training parameter recommendation models based on machine learning, and directly recommending the most likely optimal parameter set according to real-time rainfall prediction features.

[0137] Step S6, update the corrected parameters that pass the verification to the business running model, complete the dynamic synchronization of the model.

[0138] Specifically, to ensure uninterrupted business, model hot updating technology is adopted: in addition to online model service, a sandbox environment is started to load the corrected new parameters θ* and model structure. Quick smoke test is performed to ensure that the model can start and run normally. After the test passes, through dynamic link library reloading or API gateway switching, seamless switching of the business model is realized. Strict version control is implemented for the model and parameter set. A new model version number is generated each time the model is updated. All historical versions of the model and parameters are preserved to allow quick rollback in case of unexpected situations. Complete version update logs are recorded to ensure traceability of the model evolution process.

[0139] In summary, the multi-source data driven river-pipe network coupled model adaptive correction method includes: obtaining multi-source monitoring data and running the coupled model to generate simulation sequences; constructing residual signatures containing mass conservation non-closed quantities, spatial clustering and deviation time series features; inputting the residual signature into the deviation attribution discriminator to identify whether the current deviation belongs to a specific type such as rainfall input, pipe network capacity or coupling interface; generating parameter correction strategies based on the pre-set correction action atlas, locking mutually exclusive target parameter subsets corresponding to the deviation type; constructing a target function containing physical consistency constraints to perform directional update on the target parameter subset.

[0140] The present application introduces residual signature and deviation attribution discriminator, which can accurately identify whether the deviation is caused by external rainfall input, internal pipe network water delivery capacity or coupled interface state by extracting quality conservation non-closed quantity, deviation time sequence and spatial clustering and other features with clear physical meaning, realizing the qualitative change from numerical fitting to causal diagnosis. The correction action atlas and mutual exclusion parameter grouping strategy are adopted, according to the attribution result, the non-related parameters are forcibly locked by table lookup method, and only the specific target parameter subset is updated, for example, when the rainfall deviation is determined, the pipe roughness is strictly prohibited to be modified. The path of using hydraulic parameters to compensate for input deviation is blocked. At the same time, by explicitly adding a mass conservation penalty term in the objective function, and parameterizing the modeling of the coupled interface (introducing connectivity and loss coefficients), it further ensures that the corrected parameters meet the physical conservation law, and improves the generalization ability and robustness of the model.

[0141] The above describes the preferred embodiments of the present application, but the present application is not limited to the specific details in the above embodiments, and various equivalent transformations can be made to the technical solutions of the present application within the technical concept of the present application, which all belong to the protection scope of the present application.

Claims

1. An adaptive correction method for a river-pipeline coupling model driven by multi-source data, characterized in that, include: Acquire multi-source monitoring data and rainfall-driven data of the river-pipeline system, and perform spatiotemporal alignment processing on the multi-source data to obtain the processed multi-source monitoring data and rainfall-driven data; Based on the pre-stored current model parameter set, the processed rainfall-driven data is calculated using the river-pipeline coupling model to generate a simulation sequence corresponding to the processed multi-source monitoring data. The deviation between the processed multi-source monitoring data and the simulation sequence is calculated to obtain the deviation characteristics; Based on the deviation characteristics, a parameter correction strategy is determined; the parameter correction strategy limits the subset of target parameters and the range of correction space for this correction. This also includes establishing a calibration triggering system, which includes: Periodic trigger monitoring: Generates trigger signals according to a preset time period; Event-triggered monitoring: Real-time access to the rainfall data stream from the processed multi-source monitoring data, and generation of a trigger signal when the rainfall intensity exceeds a preset threshold; Deviation Trigger Monitoring: Within a preset sliding time window, the processed multi-source monitoring data and the simulation sequence are compared. When the error index between the two continuously exceeds a preset multiple of the historical background value, a trigger signal is generated. When any trigger signal is detected, proceed to the deviation characteristic calculation step; The deviation characteristics are obtained, including: Within a preset sliding time window, based on the processed multi-source monitoring data and simulation sequence, a residual signature containing multi-dimensional physical properties is constructed as a deviation feature. The residual signature includes at least: a mass conservation non-closed quantity calculated based on the integral of the total system inflow and total outflow extracted from the processed multi-source monitoring data; a spatial clustering feature calculated based on the distribution of deviation measurement points in the pipeline topology; and a deviation time series feature calculated based on the delay of the deviation occurrence time relative to the rainfall start time. Determine the parameter correction strategy, including: The residual signature is input into a pre-built deviation attribution discriminator to identify the current system deviation type; the deviation type includes at least rainfall input deviation, underlying surface parameter deviation, pipeline water conveyance capacity deviation, river water conveyance capacity deviation, and coupling interface deviation. Based on the identified deviation type, a parameter correction strategy is generated; wherein, the parameter correction strategy limits the target parameter subset to a mutually exclusive parameter set corresponding to the deviation type, and limits the correction space range to the physical region associated with the deviation type; Identify the current system deviation type, including: Extract mass conservation non-closure quantities and deviation temporal features from residual signatures; When the normalized value of the non-closed quantity of mass conservation exceeds the preset significance threshold, and the time sequence characteristic of the deviation indicates that the delay between the time of deviation occurrence and the time of rainfall start is less than the preset time threshold, the deviation type is determined to be rainfall input deviation.

2. The method according to claim 1, characterized in that, Also includes: Based on the correction space range defined in the parameter correction strategy, an objective function containing physical constraints is constructed; The objective function is optimized using an optimization algorithm, and the parameter values ​​in the subset of objective parameters are updated.

3. The method according to claim 1, characterized in that, Acquire multi-source monitoring data of the river-pipeline system, including: Acquire satellite remote sensing data, UAV aerial survey data, ground sensor data, and social media data; Feature-level fusion of satellite remote sensing data, UAV aerial survey data, ground sensor data, and social media data is used to establish a spatiotemporal distribution model of rainfall and flood elements. Spatiotemporal distribution data of precipitation intensity, river water level and flow velocity, and ground water depth are extracted from the spatiotemporal distribution model of rainfall and flood elements and used as multi-source monitoring data.

4. The method according to claim 1, characterized in that, Based on the identified deviation type, a parameter correction strategy is generated, including: Query the preset correction action map and map the deviation type to the corresponding parameter group to be corrected; The parameter set to be corrected obtained by mapping is determined as the target parameter subset, while keeping the values ​​of other parameter sets in the river-pipeline coupling model unchanged except for the target parameter subset; The mapping relationships defined in the correction action map include at least the following: mapping the rainfall input deviation to a rainfall correction parameter set, mapping the pipeline water conveyance capacity deviation to a pipeline roughness and pipe section siltation parameter set, and mapping the coupling interface deviation to a coupling interface connectivity coefficient and loss coefficient parameter set.

5. The method according to claim 2, characterized in that, Construct an objective function that includes physical constraints, including: A data fitting error term is constructed to characterize the statistical error between the simulation sequence and the processed multi-source monitoring data; A physical consistency penalty term is constructed, which is positively correlated with the mass conservation non-closure quantity of the system within the correction time window; The objective function is generated by weighted summation of the data fitting error term and the physical consistency penalty term.

6. The method according to claim 1, characterized in that, The river-pipeline coupling model uses parametric modeling for the coupling interface between the pipeline network and the river. The parametric modeling introduces connectivity coefficient and energy loss coefficient as parameters to be corrected. Among them, the connectivity coefficient represents the effective water passage area ratio of the coupling interface, and the energy loss coefficient represents the local head loss when the water flows through the coupling interface. When the parameter correction strategy involves coupling interface deviation, the connectivity coefficient and energy loss coefficient are included in the target parameter subset.

Images