River and lake regulation and storage system flood routing intelligent forecasting method based on dynamic storage and discharge relation

Abstract

The invention discloses a river and lake regulation and storage system flood routing intelligent forecasting method based on a dynamic storage and discharge relation, and the method comprises the steps: constructing a space-time dynamic association diagram structure which reflects the waterpower relation of a drainage basin, and mapping multi-source hydrological data to generate a model input set with a unified space-time scale; establishing an aggregation reservoir parameterized structure model which comprises a dynamic water level-reservoir capacity relation driven by inflow change characteristics and a composite water level-flow relation integrating a time lag effect and backwater jacking feedback; constructing a flood regulation calculation equation set, and according to the real-time delay state in the composite relationship, adopting an adaptive solution mechanism to switch among different algorithm channels so as to solve the equation set; and correcting the model state by using the real-time observation data and executing rolling prediction. The method effectively solves the problems that in a complex river and lake system, dynamic characteristics of the storage and discharge relation are difficult to depict, the non-in-phase evolution process is unstable to solve, and high-dimensional parameters are difficult to calibrate, and the continuity and physical consistency of flood forecasting are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of flood forecasting, and in particular to a flood evolution intelligent forecasting method for a river-lake regulation system based on dynamic storage-discharge relationship. BACKGROUND

[0002] With the deepening of large-scale water resources development, the main stream river and the lake connected to the river form a complex river-lake joint regulation system. The flood evolution process is restricted by the storage along the river, the lake diversion, and the backwater support of the downstream, showing a high degree of nonlinearity and time-varying characteristics. Accurate analysis of the water exchange and propagation law of the system under dynamic boundary, and construction of a high-fidelity flood evolution model, have important engineering and technical value for improving the scientificity and timeliness of the decision-making of the flood control regulation of the basin.

[0003] The prior art usually adopts Saint-Venant equation set to construct one-dimensional and two-dimensional hydrodynamic models, or uses hydrological methods such as Muskingum model for generalization simulation. Such methods generally regard the river and lake as static units with fixed geometric boundaries, describe the regulation capacity by pre-setting fixed water level-storage capacity curves according to topographic data, and represent the water level-flow relationship of the section based on single-valued processing or an empirical rope sleeve curve. In the calculation and solving level, the finite difference method or fixed iteration algorithm is often used to step by step deduce the discrete control equation.

[0004] However, in the face of scenes with severe flood fluctuation or complex hydraulic connection, the above conventional schemes have the problems of rigid representation of storage-discharge structure and insufficient adaptability of solving mechanism. Specifically, firstly, the traditional model ignores the real-time influence of flood fluctuation force effect on the storage capacity, and the fixed water level-storage capacity function cannot reflect the adjustment of water surface slope caused by sudden change of inflow, resulting in systematic deviation in dynamic storage calculation; secondly, the water level and flow response of the control section often have significant non-synchronous (lag) characteristics, and are affected by the coupling of backwater support and propagation lag time, so it is difficult for the traditional single solving algorithm to ensure the stable convergence of water balance equation in a large lag interval, and calculation shock or divergence is easy to occur. SUMMARY

[0005] The present application relates to the field of flood forecasting, and in particular to a flood evolution intelligent forecasting method for a river-lake regulation system based on dynamic storage-discharge relationship.

[0006] Technical scheme, according to one aspect of the present application, a flood evolution intelligent forecasting method for a river-lake regulation system based on dynamic storage-discharge relationship, comprising:

[0007] A space-time dynamic correlation graph structure reflecting the hydraulic connection of the basin is constructed, the collected multi-source hydrological data is mapped into the space-time dynamic correlation graph structure, and the model input data set of unified space-time scale is generated through node feature aggregation and edge weight updating.

[0008] extracting inflow variation features based on the model input dataset, and establishing an aggregated reservoir parameterized structure model representing the overall storage and discharge characteristics of the main stream and the lake, which includes a dynamic water level-storage relationship driven by the inflow variation features in the model input dataset, and a composite water level-flow relationship integrating the hysteresis effect and backwater feedback;

[0009] constructing a flood regulation equation set based on the aggregated reservoir parameterized structure model, solving the flood regulation equation set by using an adaptive solving mechanism according to the real-time hysteresis state identified from the composite water level-flow relationship, and obtaining the calculation state variables;

[0010] correcting the aggregated reservoir parameterized structure model based on real-time observation data, performing rolling prediction by using the corrected model and the calculation state variables, and outputting water level and flow process information in a future period.

[0011] Beneficial effects, the application effectively solves the problems of difficult description of dynamic characteristics of storage and discharge relationship, unstable solution of non-in-phase evolution process and difficult calibration of high-dimensional parameters in a complex river and lake system, and improves the continuity and physical consistency of flood prediction. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 is the overall flow process schematic diagram of the river and lake regulation system flood evolution intelligent prediction method based on dynamic storage and discharge relationship according to the embodiment of the application.

[0013] Figure 2 is the flow process schematic diagram of constructing a space-time dynamic correlation graph reflecting the hydraulic connection of a basin and generating a model input dataset of a unified space-time scale according to the embodiment of the application.

[0014] Figure 3 is the flow process schematic diagram of establishing an aggregated reservoir parameterized structure model representing the overall storage and discharge characteristics of the main stream and the lake according to the embodiment of the application.

[0015] Figure 4 is the flow process schematic diagram of constructing a hysteresis correction model according to the embodiment of the application. DETAILED DESCRIPTION

[0016] Embodiment one, the overall framework of the river and lake regulation system flood evolution intelligent prediction method based on dynamic storage and discharge relationship is described in detail, and an intelligent prediction scheme that can adapt to the nonlinear and non-in-phase storage and discharge response characteristics of a complex river and lake system is provided, as shown in Figure 1 .

[0017] Step 101, constructing a space-time dynamic correlation graph reflecting the hydraulic connection of a basin, and mapping the collected multi-source hydrological data to the space-time dynamic correlation graph structure, generating a model input dataset of a unified space-time scale through node feature aggregation and edge weight updating.

[0018] In this embodiment, the selection of data sources is important for constructing accurate hydraulic connections. The study area covers a complex river-lake system, such as the Three Gorges to Luoshan River section of the Yangtze River Basin, which involves the confluence of multiple tributaries and the regulation of large lakes (such as Dongting Lake). In order to comprehensively capture the hydrological process, the system first collects multi-source hydrological data, including water level and flow data measured by ground hydrological stations, rainfall data inverted by weather radar, water area data monitored by remote sensing satellites, and future rainfall forecast data output by numerical weather prediction models.

[0019] Further, the process of constructing a spatio-temporal dynamic correlation graph structure involves digital mapping of geographical spatial information and hydraulic connection relationships. The system generalizes hydrological stations, key control sections on rivers, and lake inlets and outlets as nodes in the graph structure. The connection edges between nodes are not based solely on Euclidean distance, but are established based on actual river physical connectivity. For example, there is a direct river connection between the Three Gorges reservoir outlet node upstream and the Zhicheng node downstream, so a directed edge is established.

[0020] On this basis, multi-source hydrological data is mapped to the corresponding nodes. Since data from different sources often differ in time resolution and spatial scale, such as hourly data from ground stations and 6-hourly data from numerical forecasts, temporal and spatial scale unification is required.

[0021] Specifically, interpolation, resampling, or aggregation methods are used to align all data to a pre-set standard spatio-temporal scale, such as a 6-hour time step. Through node feature aggregation, the information of the node itself and its neighborhood nodes is fused, and through a dynamic edge weight updating mechanism, the changes in influence strength between different nodes during flood propagation are reflected. The final generated model input dataset is a set of feature tensors of all basin nodes in a unified spatio-temporal framework, providing a standardized input basis for subsequent structure modeling.

[0022] In some optional embodiments, the computing device performing this step can be a server configured with a high-performance graphics processing unit (GPU) to support the training and inference of large-scale graph neural network models. If certain data sources (such as radar rainfall) are unavailable, the system can be configured to use only ground hydrological station data to construct a simplified static graph structure as a kind of degraded operation mode.

[0023] Step 102, based on the model input dataset, extract the inflow change feature, and establish an aggregated reservoir parameterization structure model representing the overall storage and discharge characteristics of the main stream and the lake; the aggregated reservoir parameterization structure model includes a dynamic water level-storage capacity relationship driven by the inflow change feature in the model input dataset, and a complex water level-flow relationship integrating the time delay effect and the backwater toppling feedback.

[0024] Specifically, this step solves the problem that traditional hydrological models are difficult to describe the dynamic storage characteristics of river-lake systems. In complex river-lake systems, there is frequent exchange of water between the main stream and the lake connected to the river, and it is greatly affected by water level fluctuations. In order to avoid the need for fine but time-consuming two-dimensional hydrodynamic simulation of the complex river diversion and lake flow field, the present embodiment uses the concept of aggregated reservoir. Aggregated reservoir refers to the aggregation of the main stream section and the lake area with close hydraulic connection into a unified water storage unit. For example, the main stream from the Three Gorges Dam downstream to the Luoshan section and the Dongting Lake area are regarded as a unified virtual reservoir.

[0025] The dynamic water level-storage relationship is one of the important components of the model. The traditional storage curve is usually static, that is, a water level strictly corresponds to a storage value. However, during the flood process, the actual storage under the same water level will change due to the influence of the inflow size and the fluctuation speed, showing a rope sleeve curve characteristic. The present embodiment introduces the inflow variation characteristics (such as the size trend and fluctuation rate of the inflow) as the driving variable, so that the water level-storage function can dynamically deform over time. For example, during the rapid rising period of the flood, the inflow surge drives the storage curve to shift to the high storage direction, reflecting the storage increment caused by the additional slope of the river and the backwater of the lake.

[0026] Correspondingly, the composite water level-flow relationship is used to describe the discharge capacity of the outlet section of the aggregated reservoir. In actual flood evolution, the outlet flow not only depends on the current water level, but also is significantly affected by the flood propagation lag (time lag effect) and the downstream water level backwater (backwater feedback). The composite relationship model is not just a simple water level-flow lookup table, but a parameterized equation that integrates multiple physical mechanisms. The time lag parameter can be automatically adjusted according to the current inflow gradient, and a negative feedback term can be introduced according to the change rate of the downstream water level, accurately simulating the non-synchronous water level-flow response process.

[0027] Step 103, constructing a flood regulation equation set based on the parameterized structure model of the aggregated reservoir, and solving the flood regulation equation set by using an adaptive solving mechanism according to the real-time time lag state identified from the composite water level-flow relationship, to obtain the regulation state variables.

[0028] In this step, the flood regulation equation set is usually composed of a water balance equation and a storage and discharge equation. The water balance equation describes the conservation relationship between the inflow, outflow and storage change of the aggregated reservoir in a unit time, that is, the inflow minus the outflow in a period of time is equal to the increment of the storage. The storage and discharge equation is provided by the parameterized structure model of the aggregated reservoir in step 102, which is specifically embodied as the dynamic water level-storage relationship and the composite water level-flow relationship.

[0029] The adaptive solving mechanism is to solve the convergence and stability problem of the equation set under different hydraulic conditions. During the flood routing process, the system will show different response states. For example, during the stable period, the response of water level and flow is relatively synchronous; while during the flood peak period or under the serious influence of backwater, the response of water level and flow will show significant lag or out-of-phase characteristics, i.e. the so-called real-time lag state. If a single numerical solution method is always used, it may lead to calculation divergence or oscillation when the out-of-phase is serious. Therefore, the system introduces a discrimination logic to monitor the lag state in real time and switch between different solving algorithms. When it is determined that the response is synchronous, the synchronous iteration algorithm with higher efficiency can be used; when it is determined that there is significant lag, the hybrid algorithm containing the prediction and correction link is switched to, so as to ensure that the equation set can be stably solved and the water level, flow and storage capacity and other calculation state variables at the next time are output.

[0030] Step 104, based on the real-time observation data, the state of the aggregated reservoir parameterized structure model is corrected, and the rolling prediction is performed by using the corrected model and the calculation state variables, and the water level and flow process information in the future period are output.

[0031] This step realizes the real-time update and rolling release of the forecast. With the passage of time, new real-time observation data (such as the latest water level and flow measured values) will continuously enter the system. The model compares the measured data with the simulation values at the last time to calculate the error. Based on the error feedback mechanism, the state variables (such as the current storage capacity and the current lag factor value) in the aggregated reservoir parameterized structure model are corrected, and even some sensitive parameters are fine-tuned, so that the model state can approach the real basin hydrological state as much as possible.

[0032] On this basis, taking the corrected state as the starting point and combining the future meteorological rainfall forecast data as input, the system uses the aforementioned flood regulation calculation framework to deduce to the future. For example, starting from the current time t, the water level and flow at t+1, t+2 and t+K (K is the time step) are calculated step by step. The above deduction is rolling, i.e. every time step, such as 6 hours, the prediction window slides one step back, and the prediction sequence for a period of time in the future (such as 5 days in the future) is always provided. The finally output information not only includes specific numerical values, but also can include key flood control indicators such as flood peak arrival time and highest water level, which provides support for flood control decision-making.

[0033] Embodiment two, the specific process of constructing a spatio-temporal dynamic graph structure, dynamic evolution and feature reconstruction is elaborated in detail, which focuses on solving the unevenness of multi-source hydrological data in spatio-temporal distribution and data quality problems, and realizes the enhancement and repair of data through graph embedding technology, as shown in Figure 2 .

[0034] Step 201, time synchronization and spatial registration are performed on the collected multi-source hydrological data, a multi-modal hydrological time series matrix is constructed, and an initial graph structure is established according to the physical connection relationship of the river.

[0035] In the embodiment, the multi-source hydrological data often has heterogeneity. For example, the ground hydrological station provides single-point time series data, the weather radar provides gridded spatial data, and the remote sensing image is non-continuous periodical surface data. Time synchronization refers to resampling all data to a unified time axis, for example, all data are unified to a time series with an interval of 6 hours. Spatial registration is to map the radar grid data and remote sensing pixel data to the geographical position of the hydrological station or river section through coordinate transformation, such as unified conversion to WGS84 geographical coordinate system or Lambert projection coordinate system.

[0036] The process of constructing the multi-modal hydrological time series matrix M(t) is actually splicing the above-mentioned aligned data. For each node v, the feature vector at time t includes not only the water level and flow observation value of itself, but also the rainfall and water surface area information corresponding to the position. The above-mentioned feature vector is arranged in time sequence to form a multi-modal time series matrix.

[0037] The establishment of the initial graph structure G0 is based on the physical topology of the river. The node set V includes all key hydrological control points in the study area. For any two nodes v and u, if there is a direct river connection between them, an edge is established. The edge weight w vu The initial value calculation can comprehensively consider the river distance and historical flow correlation.

[0038] Preferably, the calculation formula of the initial edge weight can be represented as:

[0039] w vu =exp(-d vu / σ d )*(1+ρ vu ) / 2;

[0040] Wherein, w vu represents the initial weight between nodes v and u, d vu represents the actual river distance between the two nodes, with the unit of kilometer (km). σ d is a distance attenuation coefficient for controlling the influence degree of distance on weight, which is preferably set to 80 km in the embodiment. With the increase of distance, the influence between nodes decreases exponentially. ρ vu is the Pearson correlation coefficient of the historical flow sequence of nodes v and u, with the value ranging from-1 to 1. The formula makes the nodes closer and the flow change more synchronous, the hydraulic connection more closely, and the initial weight larger.

[0041] Step 202, use the multi-modal hydrological time series matrix to drive the initial graph structure to evolve over time, update the edge weight by calculating the dynamic similarity between nodes, and obtain the spatio-temporal dynamic correlation graph structure.

[0042] Although the physical river channel is fixed, the influence strength of water flow changes with the flood process. For example, in the dry season, the influence of the upstream on the downstream may be weak and lag for a long time; while in the flood season, the wave speed is accelerated, and the connection between upstream and downstream is obviously enhanced. This step introduces a dynamic evolution mechanism. The system uses the node feature vector at each time, i.e. the slice in the multi-modal hydrological time series matrix, to calculate the feature similarity between nodes in real time.

[0043] Specifically, the dynamic edge weight w vu (t) can be updated by attention mechanism, and its calculation formula can be:

[0044] w vu (t)=Softmax u (α*sim(x v (t),x u (t)));

[0045] Where x v (t) and x u (t) represent the feature vectors of nodes v and u at time t, sim represents the similarity function, such as cosine similarity, which is used to measure the closeness of two feature vectors in direction, and α is the self-attention coefficient, which is a learnable parameter, used to adjust the sensitivity of the attention mechanism, and Softmax u represents the normalization processing of the weight of all neighbor nodes u connected to node v, ensuring that the sum of all outgoing edge weights is 1.

[0046] Through the above mechanism, the graph structure can automatically adjust the information transmission path and strength according to the current water regime, and generate a spatio-temporal dynamic correlation graph structure G t .

[0047] Step 203, extract the node embedding features of the spatio-temporal dynamic correlation graph structure, perform node missing completion, anomaly detection and residual feedback correction, and obtain an optimized graph structure containing reconstructed features.

[0048] This step uses the feature extraction capability of graph neural network GNN to process data quality problems. The spatio-temporal embedding features of each node are extracted through graph convolution or graph attention network. The spatio-temporal embedding features not only contain the information of the node itself, but also aggregate the information of the neighborhood nodes.

[0049] The specific steps for performing node missing completion, anomaly detection, and residual feedback correction include: aggregating the features of neighboring nodes in the spatiotemporally dynamic relational graph structure using a graph message passing mechanism to generate a node completion matrix; applying a sliding window residual threshold detection algorithm to the observation subset of the node completion matrix to identify and correct abnormal observation points; introducing a residual gated feedback unit to calculate the node prediction error and feeding the node prediction error back to the edge weight update process of the neighboring nodes to generate a feedback correction matrix; and fusing the corrected observation subset with the feedback correction matrix to obtain the optimized graph structure.

[0050] Specifically, to address data missingness, the system utilizes a graph message passing mechanism. If data for a node (such as the Qingjiang Gaobazhou Station) is missing at a certain time, its value can be inferred by aggregating the weighted features of its upstream and downstream neighboring nodes (such as Zhicheng and Yichang), thus generating a node completion matrix.

[0051] To address data anomalies, a sliding window residual detection algorithm is employed. Specifically, for the observed sequence, a time window is defined, for example, with a width of 12 hours. Within this window, the imputation value X is calculated. v_hat The residual r between (t) and the original observation value v (t)=|X v_hat (t)-x v (t)|。 The system presets a threshold, for example, 3 times the standard deviation 3σ. If the calculated residual exceeds this threshold, the observation is determined to be abnormal, possibly due to sensor failure or radar false alarm, and a replacement correction is performed using the completed value or weighted smoothed value.

[0052] Furthermore, to improve the accuracy of feature reconstruction, especially considering the non-stationarity of time-series data, this embodiment introduces a residual gated feedback mechanism. Specifically, this includes: introducing a residual gated feedback unit to calculate the node prediction error, and feeding the node prediction error back to the edge weight update process of neighboring nodes to generate a feedback correction matrix.

[0053] The specific calculation formula for this process can be expressed as follows:

[0054] e v (t)=g(t)*(X v_hat (t)-X v (t-1));

[0055] Among them, e v (t) represents the prediction error feedback amount of node v at time t. v_hat (t) is the current node completion value or prediction value, X v (t-1) is the node state value at the previous time step. g(t) is the gating coefficient, which is usually generated by an activation function (such as Sigmoid) and takes a value between 0 and 1. It is used to control the proportion of error feedback.

[0056] The calculated error e v (t) is not only used to correct the current node feature, but also back-propagated to adjust the connection weight between the node and its neighbor nodes, so that the model can focus on the area with larger prediction error and adaptively optimize the feature extraction strategy in the subsequent steps.

[0057] The system fuses the corrected observation subset and the feedback correction matrix to obtain an optimized graph structure. The node features in this structure fill in the missing values, eliminate the outliers, and are enhanced by spatiotemporal correlation, thus having higher reliability.

[0058] In step 204, the optimized graph structure is converted into a high-dimensional embedding matrix, and tensor expansion and spatiotemporal normalization processing are performed to generate a model input data set.

[0059] After the above processing, each node has an enhanced high-dimensional feature vector. The feature vectors of all nodes are stacked to form a high-dimensional embedding matrix E(t). In order to adapt to the input requirements of a time series prediction model (such as a long short-term memory network LSTM or a transformer model), the embedding matrices of consecutive multiple time points need to be expanded in the time dimension.

[0060] For example, if the prediction period is set to 5 days and the time step is 6 hours, then the time step K = 20. The system expands the embedding matrix sequence from time t to t+K into a three-dimensional tensor, which can be represented as (K, N, d), where N is the number of nodes, such as 17, and d is the feature dimension, such as 64. The Z-Score standardization method is used to perform spatiotemporal normalization processing, which subtracts the mean and divides by the standard deviation to eliminate the order of magnitude difference between different physical quantities (such as water level and flow). The final generated standardized tensor is the model input data set, which is directly delivered to the subsequent aggregated reservoir parameterization structure model.

[0061] Example Three: Detailed description of the construction process of the aggregated reservoir parameterization structure model, especially the specific implementation of the dynamic water level-storage capacity relationship, focusing on solving the problem that the traditional static hydrology method cannot describe the dynamic deformation of the river and lake storage capacity during the flood rise and fall process.

[0062] In step 301, the multi-level equivalent unit method is used to dynamically aggregate the main stream river channel and the lake storage area, determine the adaptive aggregation boundary and equivalent dam site section, and generalize the water body within the adaptive aggregation boundary as an aggregated reservoir unit.

[0063] In this embodiment, for the complex river-lake relationship of the Three Gorges to Luoshan River section in the middle reaches of the Yangtze River, although the direct global two-dimensional hydrodynamic simulation has high precision, the calculation time is difficult to meet the real-time rolling prediction demand. This step introduces the physical generalization concept of aggregated reservoir. The multi-level equivalent unit method is a technology for logically combining dispersed water bodies in space. Specifically, the main river channel (such as Zhicheng to Luoshan section) and the laterally connected Dongting Lake area and the flood storage area along the river are divided into the same hydraulic unit by using clustering algorithm or connectivity analysis based on elevation.

[0064] The adaptive aggregation boundary refers to the spatial range of the unit, which is not fixed but dynamically adjusted according to the water level magnitude in the prediction period. For example, at low water level, the aggregation boundary only includes the main river channel; and at high water level, the boundary automatically expands to include the floodplain and the flood storage area. The equivalent dam site section is determined as the control outflow section of the aggregated reservoir, and in this embodiment, the section where the Luoshan hydrological station is located is preferably selected as the equivalent dam site section.

[0065] On this basis, the system needs to calculate the total inflow of the aggregated reservoir. Combining the outflow of the upstream reservoir, the inflow of the tributary and the interflow into the aggregated reservoir inflow sequence is the key to driving the subsequent dynamic deformation. The calculation formula of the aggregated reservoir total inflow sequence Q in (t) can be specifically represented as:

[0066] Q in (t) = ∑Q out_u (t) + ∑Q trib_k (t) + Q int (t);

[0067] Where ∑Q out_u (t) represents the sum of the outflow of all upstream cascade reservoirs (such as the Three Gorges Reservoir and Geheyan Reservoir) at time t; ∑Q trib_k (t) represents the sum of the inflow of all tributaries (such as Xiangjiang River, Zishui River, Yuanjiang River and Lishui River) within the aggregation boundary at time t; Q int (t) represents the lateral interflow generated by the rainfall in the aggregated area which is not controlled by the hydrological station.

[0068] Through the above-mentioned manner, the originally complex multi-source inflow is integrated into a single driving variable.

[0069] Step 302, based on the historical observation data in the model input data set, a water level-storage capacity adaptive static structure is fitted and generated; the aggregated reservoir total inflow sequence is signal decomposed to obtain a trend item reflecting the slow change characteristics of the inflow and a disturbance item reflecting the mutation characteristics of the inflow.

[0070] The step establishes the static benchmark and dynamic driving source of the storage-discharge relationship. The historical water level and flow data are used in combination with the river terrain data to obtain the historical reservoir storage sequence by water balance back calculation. The least square method is used for regression analysis of the historical water level-storage data pairs to obtain the basic water level-storage curve, i.e., the water level-storage adaptive static structure V s (Z). The structure represents the average storage capacity under the condition of steady water flow.

[0071] Further, in order to capture the dynamic influence of flood fluctuation on the storage capacity, the system performs signal decomposition on the aggregated reservoir total inflow sequence Q in (t) obtained in step 301. Specifically, the moving average filtering, wavelet transform or empirical mode decomposition (EMD) method can be used. The decomposition process can be represented as:

[0072] Q in (t) = Q bar_in (t) + Q' in (t);

[0073] wherein Q bar_in (t) is a trend item representing the low-frequency slowly changing component in the inflow process, reflecting the overall magnitude and background flow of the flood; and Q' in (t) is a disturbance item representing the high-frequency sudden change component in the inflow process, reflecting the rapid fluctuation rate of the flood wave. For example, when the flood peak arrives quickly, Q' in (t) will show a significant positive value; and at the initial stage of the flood recession, Q' in (t) may turn to a negative value. The two components will be used to control the overall rate and local shape of the storage capacity curve.

[0074] Step 303, a time-varying driving factor controlling the change rate of the storage capacity is constructed using the trend item, and a nonlinear enhancement operator controlling the curve shape correction is constructed using the disturbance item; the parameter weight of the water level-storage adaptive static structure is adjusted in real time using the time-varying driving factor and the nonlinear enhancement operator, to generate a dynamic water level-storage relationship that evolves dynamically with the inflow process.

[0075] The traditional static curve cannot reflect the wedge-shaped storage or the increase in storage capacity caused by the additional slope when the flood rises. The dynamic parameter adjustment mechanism of the embodiment solves this problem. Specifically, the time-varying driving factor a(t) is mainly used to adjust the overall change rate of the storage capacity with the water level, and the construction formula thereof is preferably:

[0076] a(t) = Q barin (t) / Q max ;

[0077] wherein Q max is the maximum inflow value in the historical observation sequence, used for normalization.

[0078] Nonlinear enhancement operator B i Mainly used to control the nonlinear bending degree of the curve, i.e. the shape correction, and the construction formula is preferably:

[0079] B i = b i * Q in (t) ;

[0080] Wherein, b i is the enhancement operator coefficient to be rated, usually in the range of 0 to 1, used to control the influence weight of the disturbance term.

[0081] The water level-storage adaptive static structure is represented by a third-order polynomial. The specific calculation expression of the dynamic water level-storage relationship V(Z) is:

[0082]

[0083] Wherein, Z is the water level at the current time t; V(Z) is the calculated dynamic storage; a0, a1, a2, a3 are static polynomial coefficients determined by historical data fitting. B1, B2, B3, B4 are nonlinear enhancement operators corresponding to each order term of the polynomial, defined as the product of the disturbance term and the preset enhancement operator coefficient.

[0084] As can be seen from the formula, when the flood is in the steady period (Q in (t) is close to 0), the exponential term B i tends to 0, tends to 1, and at this time the dynamic relationship V(Z) degenerates into the static relationship V s (Z), which conforms to the physical law. When the flood is rapidly rising (Q in (t) is very large), B i increases, and each coefficient is nonlinearly amplified or reduced by α(t) (depending on whether α(t) is less than 1 or greater than 1 and the base characteristics), so that the calculated storage V(Z) is offset relative to the static value. For example, in the rising stage, due to the increase of the water surface slope, the same dam front water level corresponds to a larger actual water volume, and the formula can accurately simulate the physical phenomenon of the expansion of the storage by parameter adjustment.

[0085] Optionally, the aggregated reservoir parameterization structure model representing the overall storage and discharge characteristics of the main stream and the lake can also include the following processes, as shown in Figure 3 :

[0086] The multi-level equivalent unit method is used to dynamically aggregate the main stream river channel and the lake flood detention area to determine the adaptive aggregation boundary and the equivalent dam site section, and the water body within the adaptive aggregation boundary is generalized as an aggregated reservoir unit; ​

[0087] Based on the model input data set, the upstream reservoir outflow, tributary inflow and interval confluence flow are extracted, and the three are combined into an aggregated reservoir total inflow sequence;

[0088] The storage-discharge relationship of the aggregated reservoir unit is constructed with the aggregated reservoir total inflow sequence as the driving variable.

[0089] Compared with the traditional method of using fixed reservoir capacity curve, the dynamic water level-storage capacity relationship of the application can adaptively deform with the flood fluctuation rate, and accurately reflects the additional slope and wedge-shaped storage change.

[0090] A typical calculation example is given below:

[0091] Suppose the trend term Q bar_in (t)=15000m³ / s, the historical maximum inflow Q max =50000m³ / s, then the time-varying driving factor α(t)=15000 / 50000=0.3; suppose the disturbance term Q' in (t)=2000m 3 / s, the enhancement operator coefficient b1=0.5, then B1=0.5×2000=1000. Set the static coefficient a0=100×10 6 m 3 , a1=50×10 6 m 3 / m, a2=10×10 6 m 3 / m 2 , a3=1×10 6 m 3 / m 3 , when the water level Z=30m, the dynamic reservoir capacity V(Z)≈275×10 8 m 3 is calculated by substituting the formula, compared with the static reservoir capacity, which reflects the reservoir capacity expansion effect in the rising water period.

[0092] Example four, the construction process of the composite water level-flow relationship is described in detail, which focuses on solving the non-same phase response problem commonly existing in river and lake system, that is, the water level-flow relationship is influenced by multiple factors such as time lag, top support and scouring and silting, and presents complex rope sleeve curve characteristics.

[0093] Step 401, extract the water level and flow characteristics in the observation subset of the model input data set, and construct the water level-flow master control relationship based on the power law function through multi-scale nonlinear mapping.

[0094] This step is used to establish the baseline framework of the discharge capacity. The system selects the time period data from the historical data that is not significantly affected by backwater uplift and is in a stable flow state, such as dry season or end of ebb data. Regression analysis is performed using these data to construct the water level-flow master control relationship Q b (Z). The relationship is preferably in the form of a power law function simplified from the Manning formula or Saint-Venant equation set:

[0095] Q b (Z)=c*(Z-Z0) p ;

[0096] Or simplified as: Q b (Z)=c*Z' p ;

[0097] Where c is the flow coefficient, p is the flow index, usually close to 1.5 to 2.0, Z0 is the river bottom elevation or zero flow weighted water level, and Z' is the effective water depth. This master control relationship represents the single mapping ability of the control section water level to the flow under ideal and undisturbed conditions.

[0098] Step 402, on the basis of the water level-flow master control relationship, a lag correction model representing the delay of flood propagation time, a backwater feedback model representing the suppression effect of downstream water level on outflow, and a erosion and deposition compensation model representing the change of riverbed topography are constructed respectively; through a hierarchical parameterization coupling mechanism, the lag correction model, the backwater feedback model and the erosion and deposition compensation model are superimposed into the water level-flow master control relationship, to establish a composite water level-flow relationship integrating time domain, space domain and morphological domain interactive driving.

[0099] This step expands the single master control relationship into a multi-dimensional coupled dynamic relationship. The hierarchical parameterization coupling mechanism refers to parameterizing the three main factors affecting the flow, namely lag (time domain), backwater (space domain) and erosion and deposition (morphological domain), as independent correction terms, and then integrating them into the master control equation in a superimposed or nested manner.

[0100] Specifically, the general expression form of the composite water level-flow relationship Q(Z) can be constructed as:

[0101] Z real (t)=[Q(t-τ(t)) / c] (1 / p) +Z bk (t)+Δh bed ;

[0102] Where Z real (t) is the measured water level, the first term is the theoretical water level (including lag) calculated by flow back calculation, the second term Z bk (t) is the water level uplift caused by backwater uplift, and the third term Δh bed is the water level deviation caused by riverbed erosion and deposition.

[0103] Or expressed in the form of inverse flow: Q(t) = c * [Z(t + τ'(t)) - Z bk (t + τ'(t)) - Δh bed ] p ;

[0104] Where Q(t) is the actual flow at time t, Z(t + τ'(t)) is the water level observation value at future time t + τ'(t) that needs to be used to obtain the flow at time t, τ'(t) is the reverse lag time used to compensate for the time delay of flood wave propagation, Z bk (t + τ'(t)) is the correction term of the top support water level corresponding to the time, which needs to be used synchronously with the water level, and p is the flow index.

[0105] Step 403, constructing a lag correction model, as shown in Figure 4 , specifically including: calculating the time variation rate of the aggregated reservoir total inflow sequence to obtain the inflow lag gradient; performing nonlinear dynamic amplification on the preset baseline lag parameter based on the inflow lag gradient to generate a time-varying lag factor; using the time-varying lag factor to construct a time sequence residual feedback channel in the water level-flow master control relationship, and establishing a correction structure that drives the water level-flow master control relationship to dynamically time delay map with the change of inflow.

[0106] This step elaborates the implementation of the lag correction model (time domain). In a long river type reservoir or river-lake system, the flood peak needs time to propagate from the upstream to the downstream control section, and the propagation time (lag) is not a constant, but varies with the flow size and fluctuation speed. The specific calculation formula of the time-varying lag factor τ(t) is:

[0107] τ(t) = τ0 * (1 + k1 * dQ in );

[0108] Where τ(t) is the time-varying lag factor, τ0 is the baseline lag parameter to be rated, representing the propagation time under the average flow rate; k1 is the lag amplification coefficient to be rated, which is a positive real number to be rated; dQ in is the inflow lag gradient. Dynamic time delay mapping is realized by introducing the time offset term Q(t-τ(t)) in the water level-flow master control relationship.

[0109] The calculation of dQ in can use the difference method:

[0110] dQ in (t) = [Q in (t) - Q in (t-Δt)] / Q in (t-Δt);

[0111] When the flood is in the rapid rising stage (dQ in >0), the flood wave becomes steep, and although the wave speed increases, the overall lag effect caused by the flood peak flattening may be amplified due to the enhanced storage effect of the river channel, or in some river types, it may appear as an increase in propagation. By adjusting the sign and size of the parameter k1, this formula can flexibly adapt to different basin characteristics of fast rising and slow falling or slow rising and fast falling. By using the calculated τ(t), the system introduces the term Q(t-τ(t)) in the master control relationship, realizing dynamic time delay mapping.

[0112] The value of the parameter k1 has a significant impact on the time delay response: increasing k1 will make the time delay factor more sensitive to the inflow gradient, which is suitable for steep slope sections with rapid changes in water surface slope; reducing k1 will make the time delay change more gentle, which is suitable for gentle slope sections. The typical value range is 0.1 to 1.0, which can be determined according to historical flood data.

[0113] Step 404, real-time water level data of the equivalent dam site section is obtained, and the water level change rate of the real-time water level data with time is calculated; the backwater feedback function depicting the degree of backwater in the downstream river channel is constructed by taking the water level change rate as the driving variable; the backwater feedback function is added as a negative feedback term to the water level variable of the composite water level-flow relationship to compensate for the jacking effect of downstream water level rise on the outflow process.

[0114] This step elaborates the implementation of the jacking feedback model (spatial domain). In the middle reaches of the Yangtze River, the outflow of Luoshan section is not only controlled by the upstream inflow, but also severely affected by the downstream Hankou water level and the jacking of Dongting Lake. This jacking effect is manifested as: at the same flow, the water level is jacked up; or at the same water level, the discharge capacity is reduced.

[0115] To quantify this effect, the embodiment constructs the backwater feedback function Z bk (t). Preferably, the function takes the normalized acceleration of the water level change rate as the driving variable:

[0116] dZ=[Z(t)-Z(t-Δt)] / [Z(t-Δt)-Z(t-2Δt)+ε];

[0117] Where ε is a small constant to prevent the denominator from being zero, for example, 10 -6 . The backwater feedback function can be specifically represented as:

[0118] Z bk (t)=Z bk0 *(1+k2*dZ);

[0119] Where Z bk0 is the reference jacking water level, and k2 is the jacking amplification coefficient.

[0120] The physical meaning of the formula is that when the acceleration of water level rise suddenly increases (dZ increases significantly), there is a resistance or backwater downstream, for example, the downstream water level rises faster than the upstream, causing the water level of this section to be lifted quickly and passively. At this time, Z bk (t) increases, in the equation Z real = Z theory + Z bk (Z theory is the theoretical water level, Z bk is the increment of the top-up water level) occupies a larger proportion, which explains why the measured water level Z real is high but the actual flow does not increase accordingly.

[0121] Step 405, obtain the topographic evolution characteristics and historical deposition rate of the river section; construct a scouring and deposition offset correction term that can reflect the change of riverbed bottom elevation with time according to the topographic evolution characteristics; use the scouring and deposition offset correction term to offset the shape of the water level variable in the compound water level-flow relationship, so as to eliminate the long-term drift effect of topographic change on the water level-flow response relationship.

[0122] This step elaborates the implementation of the scouring and deposition compensation model (shape domain). The riverbed topography is not constant, and long-term erosion will cause the water level-flow relationship to move to the left (the water level decreases for the same flow), while deposition will cause it to move to the right. Specifically, the system regularly imports river topographic measurement data, or analyzes the water level at the same flow based on long-term hydrological data, to calculate the offset Δh bed of the riverbed bottom elevation relative to the base year. The scouring and deposition offset correction term can be constructed as a piecewise linear function or a step function of time t: ΔZ bed (t) = Δh bed_Year (y), ΔZ bed (t) is the water level offset caused by riverbed scouring and deposition, and Δh bed_Year (y) is the change of riverbed bottom elevation relative to the base year (such as the initial year of the station or a stable year). When performing real-time prediction, this term is added to the compound relationship as a slowly changing constant term. Although its short-term fluctuation is not large, it is important for maintaining the accuracy of the model across years, avoiding the systematic high error of water level prediction caused by riverbed incision.

[0123] As an improvement of the above scheme, in some embodiments, a machine learning residual compensation module can also be introduced. For the residual part that cannot be explained by the above physical parameterized model (main control + lag + top-up + scouring and deposition), a lightweight multilayer perception MLP can be used for fitting, and the output of the MLP is added to the compound relationship as the fifth correction term, forming a hybrid modeling scheme of physical mechanism + data-driven.

[0124] Embodiment five, the construction and execution process of adaptive solution mechanism is elaborated in detail, focusing on solving the calculation divergence or water imbalance problem caused by the single solution algorithm unable to adapt to the rapidly changing boundary conditions when the traditional hydrological model faces strong nonlinear hysteresis effect.

[0125] Step 501, construct a time-delay state discriminator, real-time monitor the time-varying time-delay factor in the composite water level-flow relationship, and output the real-time time-delay state reflecting the current flood response characteristics.

[0126] In this embodiment, the time-delay state discriminator is a logic control unit, similar to the control logic of an automatic transmission in a vehicle. It does not directly participate in numerical calculation, but is responsible for monitoring the operating conditions of the system. The monitoring index is the time-varying time-delay factor τ(t). This factor quantifies the degree of propagation delay of the flood wave at the current time. The discriminator reads the τ(t) value at each time, which is used to determine whether the current river and lake system is in a quasi-steady state operation mode or in a severe dynamic transition mode.

[0127] Specifically, the discriminator internally presets a set of threshold parameters that can distinguish different hydraulic response stages. The output of the discriminator, the real-time time-delay state, is an enumerated signal or flag bit, such as state 0 (static / synchronous) or state 1 (dynamic / asynchronous). This state signal will serve as the control instruction for the downstream algorithm selector, ensuring that the solver always works in the mode most suitable for the current hydraulic characteristics.

[0128] Step 502, parallelly establish a dynamic storage synchronous iteration algorithm suitable for the water level-flow synchronous response stage, and a prediction-correction hybrid algorithm suitable for the water level-flow asynchronous response stage, forming a heterogeneous complementary dual-algorithm channel.

[0129] This step constructs two parallel computing engines. Heterogeneous complementarity means that the mathematical principles and application scenarios of the two algorithms are different but their functions are complementary.

[0130] Channel A: Dynamic storage synchronous iteration algorithm. This algorithm is based on the assumption of synchronous response, which assumes that within a short time step, for example 1 hour, the changes in water level and flow are quasi-synchronous. Its logic is to solve the current water level Z(t+1) and flow Q(t+1) simultaneously by fixed point iteration method or Newton-Raphson method based on the simultaneous solution of storage and discharge equations. This method has fast convergence speed and high precision when the time delay is small and the flow is stable, and can strictly guarantee the water balance.

[0131] Channel B: Predict-correct hybrid algorithm. This algorithm is designed to deal with the case of significant phase difference. When there is large hysteresis, i.e. τ(t) is large, the outflow Q(t+1) is essentially determined by the past inflow state, and the instantaneous correlation with the current water level is weakened. If forced to use synchronous iteration, the equation set may not converge due to singular Jacobian matrix or iteration oscillation. This channel adopts a step-by-step strategy: predict the approximate water level or flow based on the hysteresis relationship, and correct it using the water balance equation. This method sacrifices a little synchronization accuracy, but gains higher numerical stability, ensuring that the program does not crash during the peak flood evolution period.

[0132] Step 503: Based on the real-time hysteresis state, automatically switch between the two algorithm channels, select the algorithm path that matches the current response characteristics to solve the flood regulation equation set.

[0133] This step realizes the dynamic routing of the solution path. At each simulation time step, for example t=k, the system first calls the discriminator in step 501 to obtain the state. If the state is static, the data stream is directed to channel A; if the state is dynamic, the data stream is directed to channel B. After calculation, regardless of which channel is passed through, the calculation state variables (water level, flow, storage) in a unified format are output, and they are written back to the system state memory as the initial conditions at time t=k+1. This mechanism avoids the rigid use of a single algorithm in the entire flood process of traditional models, and balances the calculation efficiency in the dry period and the calculation stability in the flood period.

[0134] Step 504: Use the hysteresis state discriminator to compare the time-varying hysteresis factor with the preset stable threshold and fluctuation threshold. When the time-varying hysteresis factor is lower than the stable threshold, it is determined that the real-time hysteresis state is in the static interval; when it is higher than the fluctuation threshold, it is determined that it is in the dynamic interval; when it is between the stable threshold and the fluctuation threshold, the state of the previous time is maintained.

[0135] This step details the quantitative criteria for state determination. In order to avoid frequent jumping of the algorithm at the critical point, this embodiment preferably introduces a hysteresis interval or double threshold logic.

[0136] Specifically, the stable threshold τ s and the fluctuation threshold τ d are set. The preferred empirical values are τ s =1, corresponding to 1 time step, and τ d =2, corresponding to 2 time steps. The determination logic is as follows:

[0137] If τ(t) < τ s , it is determined to be in the static interval. At this time, the flood propagation is fast or in the dry period, and the water level and flow are basically synchronous;

[0138] If τ(t) > τ d, the flood hysteresis effect is significant, and the waveform is deformed greatly;

[0139] If τ s ≤ τ(t) ≤ τ d , in order to maintain system stability, the determination state of the last time can be maintained unchanged, or defined as a transition interval (the prediction-correction method with stronger stability is adopted by default).

[0140] Step 505, when in the static interval, activate the dynamic reservoir capacity synchronous iteration algorithm, and solve the reservoir capacity, water level and flow variables simultaneously until the convergence condition is met.

[0141] In this step, the specific execution process of the dynamic reservoir capacity synchronous iteration algorithm is as follows:

[0142] Initialization: assume the initial water level at t+1 time Z (0) (t+1) is equal to the water level at the last time Z t .

[0143] Iterative calculation (kth iteration):

[0144] a) According to the current assumed water level Z (k) (t+1) and the composite water level-flow relationship Q(Z), calculate the trial flow Q (k) (t+1).

[0145] b) According to the water balance equation: V (k) (t+1)=V t +[Q in (t)+Q in (t+1)-Q t -Q (k) (t+1)]*Δt / 2, calculate the trial reservoir capacity.

[0146] c) According to the inverse function of the dynamic water level-reservoir capacity relationship V(Z), the new water level Z (k) (t+1) is obtained from V (k+1) (t+1).

[0147] Convergence determination: check whether |Z (k+1) (t+1)-Z (k) (t+1)| is less than the preset tolerance ε, for example 0.001 meters. If it is satisfied, the result is output; otherwise, let k=k+1, and return to continue iteration.

[0148] Step 506, when in the dynamic interval, enable the prediction-correction hybrid algorithm, use the historical flow sequence containing hysteresis information to predict the future water level response trend, and correct and solve according to the dynamic water level-reservoir capacity relationship and the water balance equation.

[0149] In this step, the specific implementation process of the prediction-correction hybrid algorithm is as follows: logical closed loop processing is performed on the data source of the lag time window:

[0150] Prediction step: directly calculate the predicted water level Z at t+1 time using the master control lag relationship of flow pred . The formula is:

[0151] Z pred (t+1)=g -1 (Q base );

[0152] Where g -1 (.) is the inverse function of the master water level-flow relationship, and the reference flow value Q base depends on the lag factor τ(t). Here, the flow information at t+1-τ(t) time is required. If t+1-τ(t) points to the future time, that is, the lag is very short, less than the prediction period step, then Q base The predicted flow sequence calculated by the upstream hydrological station is used; if t+1-τ(t) points to the past or current time, that is, the lag is longer, then Q base The known measured flow sequence or historical calculation value is used. This processing ensures that the logical chain is complete under all lag conditions.

[0153] Correction step: a) query the dynamic storage capacity curve V(Z) using the predicted water level Z pred to obtain the predicted storage capacity V pred .

[0154] b) Substitute V pred into the modified form of the water balance equation to inversely solve the final flow Q final that can satisfy the water balance. The formula is:

[0155] Q final =Q in (t)+Q in (t+1)-Q t -2*(V pred -V t ) / Δt.

[0156] Through the above non-iterative strategy of first determining the water level and then determining the flow, the risk of divergence of high-dimensional nonlinear equation groups in the lag interval is avoided.

[0157] Embodiment six: detailed description of the parameter calibration and correction process, focusing on solving the common problem of different parameters having the same effect in hydrological models, that is, different parameter combinations may lead to similar simulation results, reducing the physical interpretability of parameters and the generalization ability of the model.

[0158] Step 601, obtain the water level simulation value and the water level measured value at the same time, and the flow simulation value and the flow measured value.

[0159] This step is the basis for error calculation. In the model training or real-time correction phase, the system will obtain the measured water level Z obs (t) and the measured flow Q obs (t) of the control section (such as the Luoshan station) from the historical database or real-time telemetry network. At the same time, the model runs based on the current parameter set Θ, outputting the corresponding simulated water level Z sim (t) and simulated flow Q sim (t). In order to eliminate the dimension effect, normalization processing is usually performed on these data, or the weight is balanced in the subsequent objective function.

[0160] Step 602, construct a dynamic weight allocation mechanism, and give dynamic weights to the water level fitting error and the flow fitting error according to the importance difference of water level and flow at different prediction stages.

[0161] Traditional rating methods often use fixed weights, for example, water level and flow each account for 50%. However, the focus of flood control dispatch changes with water regime. This embodiment introduces a dynamic weight allocation mechanism.

[0162] Specifically, define the water level weight ω Z (t) and the flow weight ω Q (t), satisfying ω Z (t)+ω Q (t)=1. The allocation strategy can be adaptively adjusted based on the inflow size:

[0163] When the inflow is small (dry season or flat season), the flow measurement error is relatively large and the flood control attention is low, at this time the system automatically increases the water level weight (for example, ω Z =0.8), focusing on ensuring the accuracy of water level simulation, to serve the navigation or ecological dispatch.

[0164] When the inflow is large (flood season), accurately predicting the peak flow is important for dam peak shaving, at this time the system automatically increases the flow weight (for example, ω Q =0.7), focusing on capturing the peak process.

[0165] Step 603, based on the dynamic weight, the water level fitting error and the flow fitting error are weighted and summed to generate a multi-index comprehensive objective function for evaluating the fitting performance of the model, and the model parameters are corrected to minimize the multi-index comprehensive objective function.

[0166] The mathematical expression of the multi-index comprehensive objective function J can be constructed as:

[0167] J=∑t=1 T [ω Z sim obs Z 2 +ω Q sim obs Q 2 ];

[0168] where T is the length of the rating time window; σ Z and σ Q are the standard deviations of the observed stage and discharge series, respectively, to eliminate the dimensional difference.

[0169] The objective function not only considers the numerical approximation, but also implicitly focuses on the key features of flood control (such as the flood peak) through the introduction of dynamic weights. The process of model correction is to find the parameter combination that minimizes J in the parameter space.

[0170] Step 604, using the parameter decoupling strategy, map the high-dimensional parameter set of the aggregated reservoir parameterization structure model to the mutually independent storage parameter subset and discharge parameter subset; independently optimize the storage parameter subset and discharge parameter subset in the wide parameter space, and identify the physically effective interval of each parameter subset.

[0171] This step is the first phase of the mixed rating strategy - dimensionality reduction decoupling. The aggregated reservoir model contains two types of parameters with completely different properties: the storage parameter subset Θ v (such as polynomial coefficients a i , enhancement coefficients b i ) that control the storage capacity and the discharge parameter subset Θ Q (such as flow coefficient c, index p, lag coefficient k1, etc.) that control the discharge capacity. If directly rated in high-dimensional space at the same time, it is easy to fall into local optimum. This embodiment uses the physical mechanism decoupling method:

[0172] Lock Θ Q , only rate Θ v : Use the measured inflow and outflow to directly deduce the theoretical storage at each time according to the water balance equation, and then use the measured water level to fit the V-Z relationship. This step does not involve flow calculation and is a one-to-one function fitting, which can quickly determine the physically effective interval R v of Θ v .

[0173] Lock Θ v , only rate Θ Q ​​​​​​​​: Using the measured water level as input, the flow is calculated directly by the composite stage-discharge formula, and compared with the measured discharge. This step does not involve storage calculation, and can quickly determine Θ Q The physical effective range R Q of the parameters. The wide parameter space refers to the initial search range of the parameters being set very wide, such as [0, +∞] or [-10, 10], to ensure that all physically possible solutions are not missed.

[0174] Further, the high-dimensional parameter search space is reconstructed based on the physical effective range, and global collaborative optimization is performed in the space to obtain an optimal parameter combination that takes into account the storage-discharge coupling characteristics.

[0175] Step 605: The independent optimization results of the storage parameter subset and the discharge parameter subset are counted, and a fine parameter boundary covering the true physical change range and significantly narrower than the original wide space is constructed by applying a preset proportion of numerical perturbation.

[0176] This step is the second phase - space contraction. After independent optimization in step 604, anchor point values of each parameter under ideal decoupling conditions are obtained. However, in actual operation, storage and discharge are coupled, and the optimal solution may fluctuate around the anchor point. The system takes the parameter values obtained by independent optimization as the center and applies a small proportion of numerical perturbation (e.g. ±10% or ±15%) to construct a new search boundary. For example, if the independent rating obtained delay amplification coefficient k1 is 0.5, the fine parameter boundary is constructed as [0.45, 0.55]. The new boundary range is usually several orders of magnitude smaller than the original wide space, reducing the difficulty and uncertainty of subsequent search.

[0177] Step 606: The fine parameter boundary is used as a constraint condition to jointly search the storage parameter subset and the discharge parameter subset in the same solution space; a double convergence criterion including target function improvement rate and population parameter stability is set, and when the double convergence criterion is met, the optimal parameter combination is output.

[0178] This step is the third phase - global collaboration. In the contracted fine space, all parameters are released and allowed to change simultaneously to capture the weak coupling effect of the storage and discharge process. Collaborative optimization can use evolutionary algorithms such as differential evolution DE or particle swarm optimization PSO. To prevent overfitting or premature convergence, a strict double convergence criterion is set:

[0179] Target function improvement rate criterion: the relative change rate of the optimal target function value J best of the population of the last N generations (e.g. 20 generations) is less than the threshold value ε J (preferably 10 -3 ).

[0180] Population parameter stability criterion: the variance of the distribution of all individuals in the current population in the parameter space is less than a threshold value ε Θ (preferably 10 -2 ).

[0181] Only when the two conditions are met at the same time, the algorithm stops iteration, and outputs the final optimal parameter combination.

[0182] Compared with direct global optimization without using this strategy, the method of the embodiment not only improves the convergence speed, but also improves the Nash efficiency coefficient NSE of the obtained parameters in the verification period, and enhances the robustness of the prediction.

[0183] As an alternative, on embedded devices with limited computing resources, steps 604 and 605 can be omitted, and a pre-trained fixed parameter set is directly used for running, or only 2-3 parameters with the highest sensitivity (such as the flow coefficient c and the baseline lag time τ0) are fine-tuned online, while other parameters are frozen. Although this simplified scheme sacrifices some accuracy, it can meet the application requirements of low power consumption and low computing power.

[0184] Embodiment seven provides an optional scheme of a river and lake regulation system flood evolution intelligent prediction method based on a dynamic storage and discharge relationship, comprising the following steps:

[0185] Step 701, a graph embedding and dynamic structure reconstruction fusion method is constructed, multi-source heterogeneous hydrological data such as control section flow and reservoir discharge are aligned according to time sequence, and an initial spatio-temporal correlation graph structure is established based on spatial position, physical connection relationship and water power influence range. The correlation mode between multi-source information is extracted by using a node feature aggregation mechanism, the hydraulic connection strength and interaction relationship are dynamically adjusted through an edge weight self-learning mechanism, a spatio-temporal enhanced graph structure reflecting the flood evolution law is formed, and the node and edge correlation characteristics are uniformly projected to the same spatio-temporal scale. The model input data set of the unified graph structure is generated.

[0186] Step 702, the main stream river and the lake flood detention area are unified as a whole unit with a common storage and discharge mechanism, and are generalized as an aggregated reservoir. Based on the model input data set, an aggregated reservoir representation system is constructed: an adaptive polynomial framework is generated by introducing a time-varying factor, a water level-storage capacity relationship that changes dynamically with time, water level and hydraulic boundary is formed, a multi-factor interaction coupling model is constructed to represent the influence of factors such as flood fluctuation, downstream jacking and erosion and deposition, a composite water level-flow relationship with time-varying feedback characteristics is generated, and an aggregated reservoir parameterization structure model that can accurately describe the storage and discharge response characteristics of the aggregated reservoir is constructed.

[0187] Step 703, based on the aggregated reservoir parameterization structure model of the aggregated reservoir, the flood regulation calculation equation set is constructed combined with the storage and discharge balance and continuity equation, the self-adaptive solving mechanism is designed according to the time delay characteristics in the composite water level-flow relationship, so that the model can automatically match the appropriate calculation path to update the boundary conditions, and form an aggregated reservoir flood regulation calculation framework suitable for complex river and lake systems.

[0188] Step 704, a hybrid global-local intelligent optimization method is established, taking the aggregated reservoir flood regulation calculation framework as the core execution environment, automatically calibrating the key storage and discharge parameters in the aggregated reservoir parameterization structure model, and dynamically correcting the model parameters and state variables under the driving of real-time observation data, so that the model can reflect the current hydrological state change in real time, and the rolling prediction is performed using the corrected model to output the water level process, flow process and flood peak characteristic information in the future prediction period.

[0189] According to one aspect of the present application, a model input data set of a unified graph structure is generated, further comprising:

[0190] The multi-source heterogeneous hydrological data of reservoirs, rivers and lakes are mapped to generate an initial graph structure G0 through dynamic topology mapping, a multi-modal hydrological time series matrix M(t) is constructed to drive the time-varying evolution of the initial graph structure G0, and a spatio-temporal dynamic association graph structure G is obtained which can describe the change law of hydraulic connection and time sequence dependence. t ;

[0191] The node embedding feature h v (t) of the spatio-temporal dynamic association graph structure G t is extracted, the multi-stage feature reconstruction is realized through the node completion, anomaly detection, residual feedback and adaptive fusion process, and the dynamically completed optimized graph structure G t * is generated.

[0192] The optimized spatio-temporal graph structure G t is converted into a high-dimensional embedding matrix E(t), and tensor unfolding and spatio-temporal normalization processing are performed to generate a model input data set T t * of unified spatio-temporal scale.

[0193] E(t)=[h1(t);h2(t);…;h N (t)]∈R N×d ;

[0194] T t *=Norm(Unfold(E(t)));

[0195] Wherein, N is the number of nodes (spatial dimension), d is the node embedding dimension (feature dimension), Unfold is a tensor unfolding function, and Norm is a spatio-temporal normalization function.

[0196] In this embodiment, G t * The node embedding E(t) of d=64 dimensions is generated by graph embedding, and is unfolded into a three-dimensional tensor structure of K=20 (5-day forecast period), N=17, d=64, and spatiotemporal normalization is performed on each dimension to make different characteristic scales consistent. The final obtained T t * As a direct input of the storage and discharge response parameterization model.

[0197] According to one aspect of the present application, a spatiotemporal dynamic correlation graph structure G t is obtained, which can describe the change rule of the hydraulic connection and the time sequence dependence

[0198] The ground hydrological station Sg, radar rainfall R(x, t), remote sensing inversion A(t) and numerical forecast F num (t) and observation data are collected, time synchronization is performed using a unified sampling rate Δt, and spatial registration is performed through coordinate transformation and projection method to form a multi-modal hydrological time series matrix M(t);

[0199] M(t) = [Sg(t), R(x, t), A(t), F num (t)].

[0200] In this embodiment, the hourly data of 17 ground hydrological stations in the Three Gorges-Suoshan section are selected as Sg(t), the radar quantitative precipitation estimation product is interpolated to the unified grid of the basin with a resolution of 1km x 1km as R(x, t), and the rainfall forecast field of the 500m resolution remote sensing water body inversion result A(t) and the global / regional numerical weather forecast F num (t) are combined to jointly construct the multi-modal hydrological time series matrix M(t). Resampling is performed with a unified step length of Δt = 6h, and coordinate system is unified using the Lambert projection coordinate system, and finally the multi-modal hydrological time series matrix M(t) covering the entire Three Gorges-Jingjiang-Dongting Lake area is formed, solving the problem of inconsistent scales of different data sources.

[0201] According to the river connectivity relationship, an initial graph structure G0 is established, wherein the node v represents a hydrological station or a section, the node feature vector x v (t) includes water level, flow and rainfall, and the edge e vu represents the hydraulic connection and the spatial distance, and the edge weight initial value w vu is calculated according to the river distance d vu and the historical flow correlation ρ vu (the Pearson correlation coefficient of the historical flow sequence of the nodes v and u).

[0202] w vu = exp(-d vu / σ d )·(1+ρ vu ) / 2.

[0203] In this embodiment, 17 key nodes including Three Gorges, Zhicheng, Shashi, Luoshan and the entrances of each tributary are included in the graph structure, and the river distance d between any two nodes is calculated vu , the distance attenuation coefficient σ d =80km is constructed. Combined with the hourly flow data from 2013 to 2022, the Pearson correlation coefficient ρ vu is calculated, the edge weight w vu is generated by the formula, and the initial graph structure G0 is formed, in which the edges such as Three Gorges-> Zhicheng, Zhicheng-> Shashi, Dongting Lake-> Luoshan have higher weights due to close hydraulic connection, which can more truly reflect the flood propagation path in the study area.

[0204] The multi-modal hydrological time series matrix M(t) is used as a driving factor to dynamically evolve the graph structure, update the node features x v (t) and edge weights w vu (t), and obtain the spatio-temporal dynamic correlation graph structure G t .

[0205] w vu (t)=Softmax u (α·sim(x v (t),x u (t)));

[0206] Wherein, sim is a similarity function, and α is a self-attention coefficient.

[0207] In this embodiment, the attention mechanism is used to update the edge weight according to the node flow and water level change, so that the attention degree of downstream nodes to upstream nodes is automatically improved in the flood propagation strengthening stage. For example, the similarity between Three Gorges and Zhicheng nodes increases during the rising water, so that w vu (t) is automatically enhanced, and when the backwater influence is enhanced due to the high water level of Dongting Lake, the influence of the Dongting Lake node on the Luoshan node is also improved, and finally G t which is more in line with the actual hydrodynamic conditions is obtained.

[0208] According to one aspect of the present application, the dynamically completed optimized graph structure G t * is generated, which is further:

[0209] Extract the node spatio-temporal embedding feature h v (t), and use the graph message passing and neighborhood aggregation mechanism to realize the interpolation completion of the missing nodes in G t , and obtain the node completion matrix X v_hat (t);

[0210] h v (t)=Σ u∈N (v)wvu (t)·x u (t).

[0211] In this embodiment, the Qingjiang Gaobazhou Station has short-time missing data during part of the rainstorm period. The h v (t) is obtained by feature aggregation of neighboring nodes such as Zhicheng and Yichang stations, and the missing data section is smoothly completed.

[0212] The observation subset of the node completion matrix X v_hat (t) is applied to a sliding window residual threshold detection algorithm to identify abnormal points, and a weighted reconstruction method is used for error correction to obtain an abnormal correction matrix X v_tilde (t);

[0213] r v (t)=|X v_hat (t)-x v (t)|;

[0214] Wherein, r v (t) represents the residual of the node completion value and the original observation.

[0215] In this embodiment, the residual detection with a window width of 12 hours is applied to the Zhicheng station Q(t), and when the residual exceeds 3σ, it is marked as abnormal, and the false peak caused by radar rainfall false alarm is effectively eliminated by using neighborhood weighted reconstruction. The corrected sequence is smoother and more consistent with the laws of hydraulics.

[0216] The prediction subset of the node completion matrix is introduced into a residual gated feedback unit, the node prediction error e v (t) is calculated and fed back to the neighborhood to adjust the edge weight, to obtain a feedback correction matrix X f v (t).

[0217] e v (t)=g(t)*(X v_hat (t)-X v (t-1));

[0218] Wherein, g(t) is a gating coefficient.

[0219] In this embodiment, for the water level prediction deviation of the Dongting Lake area, the residual is fed back to the related nodes, so that the nodes closely related to the lake area (such as Chenglingji) are improved in weight, and the expression ability of the slow response characteristics of the lake area in the graph is improved.

[0220] The abnormal correction matrix X v_tilde (t) and the feedback correction matrix X f v (t) are fused, and the edge weight w vuThe self-regulating mechanism of (t) realizes the synergistic optimization of structure and features, and forms the optimized graph structure G after dynamic completion t *.

[0221] x v (t)=β·X v_tilde (t)+(1-β)·X f v (t);

[0222] Wherein, β is a fusion weight coefficient.

[0223] In this embodiment, the graph structure G t *On the key transmission link of Three Gorges-Suoshan, it shows more stable timing characteristics, reduces node noise, and edge weight connectivity is more consistent with hydrodynamic response.

[0224] According to one aspect of the present application, a parameterized structure model of an aggregated reservoir is constructed, which can accurately depict the response characteristics of the aggregated reservoir, and further for:

[0225] The multi-level equivalent unit method is used in combination with the basin partition clustering result to dynamically aggregate the main stream and the lake flood detention area, determine the adaptive aggregation boundary and the equivalent dam site section, and combine the reservoir outflow sequence Q out_u (t) of the upstream reservoir, the tributary inflow Q trib_k (t) and the interval confluence Q int (t) to obtain the total inflow sequence Q in (t) of the aggregated reservoir.

[0226] Q in (t)=∑Q out_u (t)+∑Q trib_k (t)+Q int (t) of the aggregated reservoir.

[0227] In this embodiment, the multi-level equivalent unit method is used to dynamically aggregate the river channel and lake area in the study area, and the hierarchical clustering method is used to determine the aggregation domain, so that the Three Gorges Reservoir outflow, Geheyan Reservoir outflow, tributary inflow and lake water are unified as a whole. The equivalent dam site section is automatically identified at the tail end of the aggregation domain, which corresponds to the Suoshan control section and is used as the unified outflow port of the aggregated reservoir. The total inflow Q in (t) is composed of the Three Gorges Reservoir outflow sequence Q out_u (t), the Xiangjiang River, Zishui River, Yuanjiang River, Lishui River and other tributary inflow sequences Q trib_k (t) and the flow Q int (t) directly inflowing from the rainfall in Zhicheng-Suoshan area, and is time-aligned and summed up with a unified time step of Δt=6h.

[0228] The water level-storage adaptive static structure of the aggregated reservoir is established based on a model input data set, and a total inflow sequence Qin(t) is taken as a driving quantity to adjust the response amplitude of the static framework, so as to form a multi-dimensional dynamic evolution water level-storage function V(Z) coupled with a water level Z(t), an inflow Qin(t), and a time t and a storage V(t);

[0229] Based on the interactive coupling mechanism, time delay effect, backwater uplift and scouring and silting evolution and other multi-dimensional influencing factors are integrated, and are coupled in a hierarchical parameterization manner to form a dynamic feedback system covering the time domain, the spatial domain and the shape domain, so as to obtain a water level-flow composite function Q(Z) with time-varying feedback characteristics.

[0230] The dynamic water level-storage function and the composite water level-flow function are coupled in the parameter space, the overall expression of the storage and discharge behavior is realized, and an aggregated reservoir parameterization structure model M representing the storage and discharge characteristics of the aggregated reservoir is formed.

[0231] M={V(Z),Q(Z)}.

[0232] In this embodiment, the dynamic water level-storage expression V(Z) and the composite water level-flow function Q(Z) are parameterized to maintain a continuous relationship in the same time step and water level range, thereby forming a complete aggregated reservoir parameterization structure model for flood regulation calculation and parameter calibration.

[0233] According to an aspect of the present application, the water level-flow composite function Q(Z) with time-varying feedback characteristics is obtained as follows:

[0234] The water level and flow characteristics are extracted in the observation subset of the model input data set, and a power-law function-based water level-flow master control relationship Q b (Z) is constructed by a multi-scale nonlinear mapping method.

[0235] Q b (Z)=c(Z') p ;

[0236] Wherein, c, p are the scale and exponential parameters of the power-law master control function to be calibrated, and Z'(t) is the water level (effective water depth) in the master control single water level-flow relationship.

[0237] In this embodiment, first, the hourly water level and flow data of Luoshan section from 2013 to 2022 are collected to form a master sample set. The double logarithmic linearization method is used to perform power-law regression on (Z, Q) to preliminarily obtain the value range of parameters c and p. The power-law function constructed on this basis can be used as a benchmark expression for subsequent time delay correction, uplift feedback and scouring compensation, and provides a consistent master control framework for the composite water level-flow relationship.

[0238] A time series residual feedback channel is constructed in the water level-flow master control relationship, and a flow lag gradient dQ is calculated in A time-varying time lag factor τ(t) is formed, a time delay mapping for dynamic updating of the driving master control function is established, and a time lag correction model is obtained.

[0239] dQ in (t)=[Q in (t)-Q in (t-Δt)] / Q in (t-Δt);

[0240] τ(t)=τ0(1+k1·dQ in );

[0241] Q(t-τ(t))=Qb(Z'(t));

[0242] Z'(t)=[Q(t-τ(t)) / c] 1 / p ;

[0243] Wherein, τ0 is a baseline time lag parameter to be rated, and k1 is a time lag amplification coefficient to be rated.

[0244] A backwater feedback function Z bk (t) is constructed by controlling the cross-section water level change rate dZ, the backwater effect of downstream river water level rise on the aggregated reservoir outflow process is described, and a backwater feedback model is obtained.

[0245] dZ=[Z(t)-Z(t-Δt)] / [Z(t-Δt)-Z(t-2Δt)+ε];

[0246] Z bk (t)=Z bk0 (1+k2·dZ);

[0247] Wherein, ε is a small constant (for example, 10 -6 ) to prevent the denominator from being zero, Z bk0 is a backstop reference strength parameter (reference backstop water level) to be rated, and k2 is a backstop amplification coefficient to be rated. For initial time t=0 and t=1, Z bk (t)=Z bk0 .

[0248] In this embodiment, the water level change rate of the screw mountain itself is used to reflect the backwater and backstop effect of the upstream river, a backwater feedback term based on the water level change rate is constructed, and the backstop strength is controlled by the parameter k2 to match the differentiated hydrodynamic force in dry and flood seasons.

[0249] According to the characteristics of topographic evolution and sedimentation change, a scouring and silting offset correction term ΔZ bed, adjust the water level-flow pattern deviation caused by terrain uplift or erosion, and generate a scouring and silting compensation model;

[0250] In this embodiment, it is considered that ΔZ bed Approximately does not change with time, and is used as a compensation term for correcting long-term pattern deviation of the master control function.

[0251] By layering parameterization, coupling time delay correction, jacking feedback and the scouring and silting compensation model, a dynamic feedback system of interactive driving of time domain, space domain and pattern domain is established, and a composite water level-flow function Q(Z) with time-varying feedback characteristics is obtained.

[0252] Z(t) = Z'(t) + Z bk (t) + Δh bed (t);

[0253] Q(Z) = c[Z(t) - Z bk (t) + Δh bed (t)] p ;

[0254] Z(t) = [Q(t - τ(t)) / c] 1 / p + Z bk (t) + Δh bed (t);

[0255] Wherein, Z(t) is the measured water level, which is composed of three parts of single master control relationship, jacking water level and scouring and silting water level.

[0256] In this embodiment, first, the master control water level-flow function is constructed, and a dynamic feedback term is superimposed, so that the outflow of the Luoshan control section can synchronously respond to the time delay effect, jacking effect and scouring and silting effect, and the complete expression of the aggregated reservoir storage and discharge response dynamic structure is realized.

[0257] According to one aspect of the present application, an aggregated reservoir flood regulation calculation framework suitable for a complex river and lake system is generated, further for:

[0258] Taking the control section as a node, the storage and discharge equation of the parameterized structure model M injected into the aggregated reservoir is combined with the water balance equation to form an aggregated reservoir flood regulation calculation equation group which can describe the storage and detention changes and hydrodynamic feedback at the same time.

[0259] V(t + 1) - V(t) = (Q in (t + 1) + Q in (t) - Q(t + 1) - Q(t)) / 2·Δt;

[0260] Q(t + 1) = Q(Z(t + 1));

[0261] V(t + 1) = V(Z(t + 1)).

[0262] In this embodiment, the control section of the spiral mountain is taken as the solution node, and the storage-outflow correlation relationship is established according to the aggregated reservoir dynamic storage function V(Z) and the composite water level-flow function Q(Z), representing the storage-discharge equation. Combined with the water balance equation, the flood regulation calculation equation set is formed to provide a unified calculation framework for subsequent adaptive solution.

[0263] The hysteresis state discriminator is constructed to dynamically monitor the hysteresis factor τ(t) in the composite water level-flow function, and a dynamic storage capacity synchronous iteration algorithm and a prediction-correction hybrid algorithm are simultaneously constructed. Based on the feedback information of the discriminator, the solution algorithm is automatically switched to form an adaptive solution mechanism.

[0264] Based on the adaptive solution mechanism, the water level, flow, storage and other information obtained at each time is written back to the aggregated reservoir parameterized structure model, the internal parameters of the water level-storage relationship and the water level-flow relationship are dynamically updated, and the boundary conditions for the next calculation are simultaneously refreshed to realize the closed-loop update of the flood regulation calculation and form the aggregated reservoir flood regulation calculation framework.

[0265] In this embodiment, the flood regulation calculation results Z(t+1), Q(t+1), V(t+1) at each time are written back to the aggregated reservoir parameterized structure model M. The model adaptively updates the parameters in the dynamic storage function V(Z, t) and the composite water level-flow function Q(Z, t) according to the new state quantity, so that the model continuously reflects the time variation characteristics of the storage and discharge capacity during the whole flood process. The updated function is used as the boundary condition and the storage and discharge structure input for the next calculation step to realize the rolling closed-loop update of the flood regulation calculation.

[0266] According to one aspect of the present application, an adaptive solution mechanism is formed, which further comprises:

[0267] The hysteresis factor τ(t) in the composite water level-flow function is calculated in real time, and the state discriminator is constructed. When τ(t) is lower than the stable threshold τ s , it is determined to be a static interval, and when τ(t) is higher than the fluctuation threshold τ d , it is determined to be a dynamic interval, and the hysteresis state signal is output.

[0268] In this embodiment, by monitoring the inflow change gradient, the hysteresis term in the composite water level-flow function is calculated, and combined with the empirical analysis of the typical flood process of the spiral mountain section, the static threshold τ s =1 and the dynamic threshold τ d =2 are selected. When τ(t)<τ s , it is considered that the water level and flow response are approximately synchronous; when τ(t)>τ d , it is considered that there is significant hysteresis; and for the case of τ s ≤τ(t)≤τ d , the hysteresis state retention rule is adopted to avoid frequent switching.

[0269] A dynamic reservoir capacity iterative algorithm is constructed to simultaneously solve for water level, flow rate, and reservoir capacity variables in the static interval, ensuring the convergence of the storage and release equation. A prediction-correction hybrid algorithm is also constructed to predict future water level trends using the lagging flow rate in the significant lag interval. Correction is performed based on the water level-storage capacity function and the water balance equation, and the real-time flow rate is solved. The two form a heterogeneous and complementary dual algorithm channel.

[0270] In this embodiment, the synchronous iterative algorithm first assumes that the initial water level at time t+1 is equal to the water level at time t: Z (k=0) (t+1)=Z(t), calculate the flow rate Q at time t+1 based on the water level-flow equation. (k) (t+1)=Q(Z (k) Substituting (t+1) into the water balance equation, we can calculate the reservoir capacity V at time t+1. (k) (t+1), the water level Z at time t+1 is calculated by inverse calculation using the dynamic reservoir capacity function. (k+1) (t+1), assuming water level Z (k) With reverse calculation of water level Z (k+1) When the accuracy requirements are met, the flood control calculation is completed at time t+1, and the calculation continues along the time scale until the end of the forecast period; otherwise, the water level Z is calculated in reverse. (k+1) Recalculate the flow rate Q (k+1) (t+1), continue iteratively to solve for the water level and flow rate at time t+1. The prediction-correction rule, based on the measured or predicted flow rate at time t+1-τ, derives the water level Z(t+1)=g at time t+1. -1 (Q(t+1-τ)) is used to derive the flow rate Q(t+1) at time t+1 based on the water balance equation, avoiding iterative solutions and exhibiting greater stability in the significant lag phase.

[0271] Based on heterogeneous complementary dual-algorithm channels, when the lag state is in a static interval, the dynamic capacity iteration algorithm channel is activated, and when it is in a dynamic interval, the prediction-correction hybrid algorithm channel is enabled. The two are dynamically switched in the time series, forming a dual-channel adaptive solution mechanism that is sensitive to lag changes.

[0272] According to one aspect of this application, generating information on water level processes, flow processes, and flood peak characteristics over a future forecast period, further includes:

[0273] Based on the principle of error minimization, a dynamic weight allocation mechanism is designed, taking into account the simulated water level value Z. sim (t) and measured value Z obs (t), simulated flow value Q sim (t) and measured value Q obs The fitting effect of (t) is used to form a multi-index comprehensive objective function J that simultaneously measures the fitting performance of water level and flow rate.

[0274] J=∑t=1 T [ω Z (t)(Z sim (t)-Z obs (t)) 2 +ω Q (t)(Q sim (t)-Q obs (t)) 2 ];

[0275] wherein, ω Z (t) is a water level fitting effect weight; ω Q (t) is a flow fitting effect weight; both satisfy ω Z (t)+ω Q (t)=1.

[0276] In the embodiment, the water level and flow fitting effect weights are adjusted through a dynamic weight distribution mechanism. When the water level storage function is calibrated alone, the flow weight ω Q (t) is set to 0, and the water level weight ω Z (t) is set to 1; when the water level and flow function is calibrated alone, the water level weight ω Z (t) is set to 0, and the flow weight ω Q (t) is set to 1.

[0277] Based on the multi-index comprehensive objective function J, a hybrid global-local intelligent optimization algorithm is used to automatically calibrate the aggregated reservoir parameterized structure model, and through the phased parameter decoupling and global collaborative convergence strategy, the optimal parameter combination Θ * is obtained.

[0278] Based on the optimal parameter combination Θ * , the latest observation data is used to assimilate and correct the internal parameters and state variables of the model, so that the aggregated reservoir parameterized structure model can immediately reflect the latest state of flood evolution, and the updated model is used as the core execution environment to implement the rolling prediction of water level process, flow process and flood peak characteristics in the future prediction period.

[0279] According to an aspect of the application, the optimal parameter combination Θ * is further:

[0280] The high-dimensional parameter set Θ of the aggregated reservoir parameterized model structure is mapped into two independent subsets of reservoir capacity curve parameters Θ V and flow curve parameters Θ Q in stages, and is optimized in a low-dimensional space wide boundary after decoupling, to obtain a stable parameter interval under the dominance of a single physical mechanism.

[0281] In this embodiment, the high-dimensional parameter space is first divided into a parameter subset Θ related to the reservoir capacity curve V The parameter subset Θ related to the water level-flow Q The parameter optimization rate is calibrated Θ by driving the water level-flow sub-model with the hourly measured water level Q The parameter optimization rate is calibrated Θ by driving the water level-flow sub-model with the hourly measured water level V This low-dimensional independent calibration avoids the high-dimensional space dimension disaster, significantly stabilizes the optimization process, and provides reliable initial values and reasonable physical parameter intervals for subsequent collaborative optimization.

[0282] Based on the multiple independent calibration results, the parameter subset value interval is calculated, and the original wide range boundary is replaced by a significantly narrowed parameter space that covers the true physical change range through a disturbance expansion strategy. In the narrow domain solution space, the high-dimensional parameter set Θ is globally optimized collaboratively.

[0283] In the global collaborative optimization process, double convergence criteria are set, and the improvement rate of the objective function and the stability of the population parameters are monitored. When the double convergence conditions are met, it indicates that the multi-parameter collaborative convergence is achieved, and the global optimal parameter combination is output.

[0284] In this embodiment, the target function convergence threshold ε J =10 -1 and the parameter change threshold ε Θ =10 -2 In actual operation, the collaborative calibration usually converges stably after 120-300 iterations. The final parameters remain consistent in independent test floods, and the comprehensive prediction error of the model is significantly improved in terms of time and accuracy compared to the initial wide range value range collaborative calibration.

[0285] The application adopts a parameterized dynamic modeling method driven by inflow characteristics. By decomposing the total inflow sequence of the aggregated reservoir into a trend item and a disturbance item, a time-varying driving factor and a nonlinear enhancement operator are constructed to dynamically modulate the parameter weight of the static water level-storage capacity polynomial in real time. This method makes the storage capacity function no longer a rigid single-value curve, but can adaptively deform with the flood fluctuation rate, accurately capturing the additional slope and wedge-shaped storage capacity changes that traditional static models cannot describe, effectively solving the problem of inaccurate dynamic storage capacity quantization. It solves the problem of systematic deviation in dynamic storage capacity calculation caused by rigid representation of storage and discharge structure (i.e. fixed water level-storage capacity curve cannot reflect the dynamic effect of flood).

[0286] The application adopts a dual strategy of physical structure parameterization and algorithm adaptive switching. At the physical level, a composite water level-flow relationship integrating time-varying hysteresis factor (based on inflow gradient) and backwater jacking feedback term (based on water level acceleration) is constructed, explicitly depicting the out-of-phase hysteresis loop characteristics of water level and flow. At the calculation level, an adaptive solving mechanism including a hysteresis state discriminator is established. The mechanism can automatically switch between dynamic reservoir capacity synchronous iteration algorithm (suitable for steady state) and prediction-correction hybrid algorithm (suitable for dynamic state) according to the real-time monitored hysteresis degree. This means effectively avoids the oscillation risk of iterative method in the strong nonlinear hysteresis interval, guarantees the convergence and water quantity balance of flood regulation equation set under the condition of dramatic boundary change, and improves the numerical stability of the prediction. The problem of unstable solution of non-synchronous evolution process under complex boundary (i.e. the traditional single algorithm is easy to diverge in the significant hysteresis interval) is solved.

[0287] The above describes the preferred embodiments of the application, but the application is not limited to the specific details of the above-described embodiments. Within the technical concept of the application, various equivalent transformations of the technical solutions of the application can be made, and these equivalent transformations all belong to the protection scope of the application.

Claims

1. A method for intelligent forecasting of flood evolution in river and lake regulation systems based on dynamic storage-discharge relationships, characterized in that, include: A spatiotemporal dynamic correlation graph structure reflecting the hydraulic connections of the watershed is constructed. The collected multi-source hydrological data is mapped to the spatiotemporal dynamic correlation graph structure. Through node feature aggregation and edge weight update, a model input dataset with a unified spatiotemporal scale is generated. Based on the inflow change features extracted from the model input dataset, a parameterized structural model of the aggregated reservoir is established to characterize the overall storage and discharge features of the main stream and the lake. It includes a dynamic water level-storage capacity relationship driven by the inflow change features in the model input dataset, as well as a composite water level-discharge relationship integrating the lag effect and backwater backwater feedback. A flood control calculation equation set is constructed based on the parameterized structural model of the aggregated reservoir. Based on the real-time lag state identified from the composite water level-discharge relationship, an adaptive solution mechanism is used to solve the flood control calculation equation set to obtain the calculation state variables. Based on real-time observation data, the parameterized structural model of the aggregated reservoir is corrected for its state. The corrected model and calculated state variables are used to perform rolling predictions and output water level and flow process information for future periods.

2. The method according to claim 1, characterized in that, A spatiotemporal dynamic correlation graph structure reflecting the hydraulic connections of the watershed is constructed, and a model input dataset with a unified spatiotemporal scale is generated, including: The collected multi-source hydrological data were synchronized in time and registered in space to construct a multimodal hydrological time series matrix and establish an initial graph structure based on the physical connectivity of the river channels. The initial graph structure is driven by a multimodal hydrological time series matrix to undergo time-varying evolution. The edge weights are updated by calculating the dynamic similarity between nodes to obtain a spatiotemporally dynamic relational graph structure. Extract the node embedding features of the spatiotemporal dynamic relational graph structure, perform node missing completion, anomaly detection and residual feedback correction to obtain an optimized graph structure containing reconstruction features; The optimized graph structure is transformed into a high-dimensional embedding matrix, and tensor expansion and spatiotemporal normalization are performed to generate the model input dataset.

3. The method according to claim 1, characterized in that, Establish a parameterized structural model of the aggregated reservoir to characterize the overall storage and discharge characteristics of the main stream and lake, including: The multi-level equivalent unit method is used to dynamically aggregate the main river channel and lake flood storage and detention areas, determine the adaptive aggregation boundary and equivalent dam site section, and generalize the water body within the adaptive aggregation boundary into aggregated reservoir units. Based on the model input dataset, the outflow from the upstream reservoir, the inflow from the tributaries, and the confluence flow between the sections are extracted and merged into a total inflow sequence of the reservoir. Using the total inflow sequence of the aggregated reservoir as the driving variable, the storage and release relationship of the aggregated reservoir unit is constructed.

4. The method according to claim 3, characterized in that, The dynamic water level-reservoir capacity relationship is constructed through the following steps: Based on historical observation data in the model input dataset, an adaptive static structure of water level and reservoir capacity is generated by fitting. Signal decomposition was performed on the total inflow sequence of the Juhe Reservoir to obtain a trend term reflecting the gradual change in inflow and a disturbance term reflecting the abrupt change in inflow. A time-varying driving factor controlling the rate of change of storage capacity is constructed using a trend term, and a nonlinear enhancement operator controlling the shape correction of the control curve is constructed using a disturbance term. By utilizing time-varying driving factors and nonlinear enhancement operators, the parameter weights of the water level-reservoir capacity adaptive static structure are adjusted in real time, generating a dynamic water level-reservoir capacity relationship that evolves dynamically with the inflow process.

5. The method according to claim 4, characterized in that, The adaptive static structure of water level-reservoir capacity is represented by a third-order polynomial; the specific calculation expression for the dynamic water level-reservoir capacity relationship is as follows: ； Where Z is the water level at the current time t, V(Z) is the dynamic reservoir capacity, and a0, a1, a2, a3 are the coefficients of the static polynomial; α(t) is a time-varying driving factor, defined as the ratio of the trend term to the historical maximum inflow; B1, B2, B3, and B4 are the nonlinear enhancement operators corresponding to each order of terms, defined as the product of the perturbation term and the preset enhancement operator coefficients.

6. The method according to claim 3, characterized in that, The composite water level-flow rate relationship is constructed through the following steps: Water level and flow characteristics are extracted from the observation subset of the model input dataset, and a water level-flow control relationship based on a power law function is constructed through multi-scale nonlinear mapping. Based on the water level-discharge control relationship, a time lag correction model representing the flood propagation time delay, a top-feedback feedback model representing the downstream water level's inhibitory effect on outflow, and a scouring and deposition compensation model representing changes in riverbed topography are constructed respectively. By using a hierarchical parameterized coupling mechanism, the time-delay correction model, the top-support feedback model, and the scouring and sedimentation compensation model are superimposed on the water level-flow master control relationship, thus establishing a composite water level-flow relationship driven by the interaction of the time domain, spatial domain, and morphological domain.

7. The method according to claim 6, characterized in that, The time-delay correction model is constructed through the following steps: Calculate the time rate of change of the total inflow sequence of the aggregated reservoir to obtain the inflow lag gradient; The preset baseline lag parameters are nonlinearly and dynamically amplified based on the inflow lag gradient to generate a time-varying lag factor. By utilizing time-varying lag factors, a time-series residual feedback channel is constructed in the water level-flow control relationship, and a correction structure is established to drive the water level-flow control relationship to dynamically map the time delay as the inflow changes. The specific formula for calculating the time-varying lag factor is: τ(t) = τ0 * (1 + k1 * dQ) in ); Where τ(t) is the time-varying lag factor, τ0 is the baseline lag parameter to be calibrated, k1 is the lag amplification factor to be calibrated, and dQ in The inflow hysteresis gradient; Dynamic time delay mapping is achieved by introducing a time offset term Q(t-τ(t)) into the water level-flow master relationship.

8. The method according to claim 6, characterized in that, The top-support feedback model is constructed through the following steps: Obtain real-time water level data of the equivalent dam site cross section and calculate its rate of change over time; Using the rate of change in water level as the driving variable, a backwater feedback function is constructed to characterize the degree of backwater in the downstream river channel; The return water feedback function is superimposed as a negative feedback term onto the water level variable in the composite water level-flow relationship to compensate for the backwater effect of downstream water level rise on the outflow process.

9. The method according to claim 7, characterized in that, The adaptive solution mechanism is constructed through the following steps: Construct a time-delay state discriminator to monitor the time-varying time-delay factor in the composite water level-discharge relationship in real time and output the real-time time-delay state that reflects the current flood response characteristics; A dynamic reservoir capacity synchronous iterative algorithm suitable for the synchronous response stage of water level and flow rate is established in parallel, and a prediction-correction hybrid algorithm suitable for the heterogeneous response stage of water level and flow rate is established, forming a heterogeneous and complementary dual algorithm channel. Based on the real-time delay state, the system automatically switches between the two algorithm channels and selects the algorithm path that matches the current response characteristics to solve the flood control calculation equations.

10. The method according to claim 9, characterized in that, The specific execution process of the adaptive solution mechanism includes: The time-varying lag factor is compared with the stability threshold and the fluctuation threshold using a lag state discriminator. When the time-varying lag factor is lower than the stability threshold, the real-time lag state is determined to be in the static range. When it is higher than the fluctuation threshold, it is determined to be in the dynamic range. When in the static interval, activate the dynamic reservoir capacity synchronous iterative algorithm, and solve the reservoir capacity, water level and flow variables simultaneously until the convergence condition is met. When in the dynamic range, the prediction-correction hybrid algorithm is activated, which uses historical flow sequences containing lag information to predict future water level response trends, and performs correction solutions based on the dynamic water level-storage capacity relationship and water balance equation.

Images