A multi-period water replenishment evaluation method based on big data simulation

Abstract

The application discloses a multi-period water supplement evaluation method based on big data simulation, and comprises the following steps: taking the Yangtze River estuary and the downstream basin as an object, establishing a hydrodynamic salinity diffusion numerical model, and training a physically constrained Fourier neural operator proxy model; dividing spatial grid nodes to construct a spatial topology connection graph; establishing data interaction relationship between nodes and updating parameters in real time through an improved spatial perception Mamba network; calculating uncertainty based on the salinity simulation results output by the proxy model, and dynamically feeding back and optimizing the Mamba network; reversely transmitting the optimized prediction results to dynamically update the constraint conditions of the proxy model; and determining an ecological water supplement position, time and water quantity scheme. The application realizes high-precision and high-frequency ecological water supplement decision support.

Description

Technical Field

[0001] This invention relates to the field of hydrology and environmental engineering technology, and in particular to a multi-period water replenishment assessment method based on big data simulation. Background Technology

[0002] With increasingly stringent requirements for ecological environmental protection in estuaries and nearshore watersheds, effectively controlling salt tide intrusion and accurately ensuring ecological water demand have become crucial aspects of comprehensive estuary and watershed management. Currently, conventional ecological water replenishment assessment methods largely employ traditional numerical simulation techniques. These methods construct estuarine hydrodynamic salinity diffusion models and utilize observational data to simulate and predict salt tide diffusion trends and ecological water demand. However, traditional numerical simulations suffer from high computational complexity and long simulation times, making it difficult to support high-frequency, dynamic ecological water replenishment decisions, especially during droughts and periods of low water levels.

[0003] In recent years, with the development of artificial intelligence, big data, and simulation technologies, researchers have attempted to apply deep learning algorithms to the assessment of ecological water replenishment in estuaries. For example, some existing technical solutions introduce deep neural networks to train models using historical monitoring data to predict salt tide propagation or ecological water demand. While this improves computational efficiency, the models lack generalization ability, are easily affected by changes in data distribution, and their accuracy is difficult to guarantee. Especially in the context of complex environments and dynamic changes in hydrological conditions in estuaries, prediction methods that rely solely on data-driven models often lack physical mechanism constraints, resulting in insufficient reliability and stability in predictions.

[0004] To improve the accuracy of ecological water replenishment assessment methods, researchers have attempted a combination of numerical and data-driven models, initially demonstrating the advantage of balancing simulation accuracy and computational efficiency. However, existing technologies lack a robust collaborative mechanism between models and effective means to address the complex environmental characteristics of estuaries, spatiotemporal dynamics, sudden salt tide events, and multi-period differences in ecological water demand. Furthermore, there are currently no successful applications that effectively integrate quantum computing principles into deep learning frameworks to significantly improve model sensitivity and prediction stability.

[0005] Therefore, how to provide a multi-period water replenishment assessment method based on big data simulation is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a multi-period water replenishment assessment method based on big data simulation. Addressing the problem that existing technologies struggle to accurately and frequently support dynamic simulation of salt tide intrusion and ecological water replenishment decisions, this invention proposes a deep proxy model combining physically constrained Fourier neural operators and a quantum-enhanced spatial perception Mamba network. Through deep collaborative fusion and a dynamic closed-loop feedback mechanism, it achieves accurate simulation and prediction of multi-period salt tide propagation and ecological water demand. This invention possesses the advantages of high simulation efficiency, high prediction accuracy, and high decision response frequency.

[0007] A multi-period water replenishment assessment method based on big data simulation according to an embodiment of the present invention includes:

[0008] Establish a basic numerical model of hydrodynamics and salinity diffusion, and obtain historical observation data of the Yangtze River Estuary and its downstream basin;

[0009] The simulation results of the basic numerical model were used as physical constraints, and a deep surrogate model of Fourier neural operators was trained based on historical observation data to obtain the dynamic simulation results of salt tide in the estuary area.

[0010] Based on the spatial distribution characteristics of the Yangtze River Estuary and its downstream basin, the area is divided into multiple spatial grid nodes, and a spatial topology connection map is constructed.

[0011] By using an improved spatially perceptive Mamba neural network, data interaction relationships between multiple spatial nodes are established based on a spatial topology connection graph. The network parameters are continuously updated based on real-time data collected in the estuary area, and the prediction results of salt tide intrusion and ecological water demand are output.

[0012] Based on the dynamic simulation results of salt tide in the estuary area, the prediction results of salt tide intrusion and ecological water demand are optimized.

[0013] The optimized salt tide intrusion prediction results and ecological water demand prediction results are back-transmitted to the Fourier neural operator deep proxy model to obtain the updated simulation results of estuarine salt tide propagation.

[0014] Based on the updated simulation results of estuarine salt tide propagation and the optimized ecological water demand prediction results, a multi-period ecological water replenishment plan was determined.

[0015] Optionally, the establishment of a basic numerical model for hydrodynamics and salinity diffusion, and the acquisition of historical observation data of the Yangtze River Estuary and its downstream basin, specifically involves:

[0016] For the Yangtze River Estuary and downstream basin, the boundary of the study area is dynamically delineated based on real-time tidal level monitoring data, and a spatial digital terrain model is constructed based on the lowest and highest ranges of river water levels during the dry season and drought season.

[0017] Salinity monitoring stations were selected within the study area, and salinity values ​​were continuously monitored at various stages of the dry and arid seasons.

[0018] Simultaneously, continuous runoff observation sections were set up along the upstream direction to the river mouth within the study area, and the upstream inflow was continuously recorded at each stage of the dry season and drought period.

[0019] Real-time acquisition of tide level data from estuary tide observation stations in the study area;

[0020] Based on monitoring and observation data, initial tidal level boundaries, inflow boundaries, and initial salinity distribution boundaries were set for different stages of the dry season and drought season, and independent solutions and simulation verifications were performed to form a basic numerical model of hydrodynamics and salinity diffusion for different stages.

[0021] Historical data on changes in estuary tide level, upstream runoff, and estuary salinity for no less than three consecutive years were obtained at each monitoring station and observation section to form historical observation data.

[0022] Optionally, the process of training a deep surrogate model based on historical observation data using Fourier neural operators to obtain dynamic simulation results of salt tides in the estuary area specifically involves:

[0023] Using the simulation results of the basic numerical model as initial constraints, simulation data are extracted to form an initial simulation dataset;

[0024] Based on historical observation data, the measured values ​​corresponding to the spatial grid nodes of the initial simulation dataset are extracted to form the historical observation dataset.

[0025] Fourier neural operator networks are used as deep proxy models to be trained;

[0026] The position coordinates of each spatial grid node and the simulation time corresponding to the initial simulation dataset are used as the model input data;

[0027] The initial simulation dataset and the historical observation dataset were used together as the training target dataset for the model;

[0028] The set of governing equations, consisting of the mass conservation equation, the salt diffusion equation, and the estuary hydrodynamic equation, serves as the physical constraints during the training process.

[0029] The deviation between the tide level, current velocity, and salinity values ​​predicted by the deep proxy model and the corresponding historical measured values ​​in the training target dataset is used as the training loss function value.

[0030] The parameters of the deep proxy model are updated step by step through the backpropagation algorithm until the training loss function value meets the preset threshold, at which point training stops and the Fourier neural operator deep proxy model is obtained.

[0031] Based on the Fourier neural operator deep surrogate model, the tidal level, current velocity and salinity of each spatial grid node in the Yangtze River Estuary and downstream basin are simulated in real time to obtain the dynamic simulation results of salt tide in the estuary area.

[0032] Optionally, based on the spatial distribution characteristics of the Yangtze River Estuary and its downstream basin, the area is divided into multiple spatial grid nodes, and a spatial topology connection map is constructed, specifically as follows:

[0033] Based on the spatial digital terrain model, the study area is divided into spatial grid nodes;

[0034] For each spatial grid node, the water flow connectivity intensity and salt diffusion and transport intensity between nodes are calculated based on historical observation data.

[0035] Each spatial grid node is used as a node in the topology graph, and the water flow connectivity strength and salt diffusion and transport strength between nodes are used as the weight values ​​of the edges in the topology graph to establish an initial spatial topology connection graph.

[0036] Based on the salt diffusion and transport intensity weight value of each node in the initial spatial topology connection graph, a simplified spatial topology connection graph is obtained by using the weight threshold screening method.

[0037] A connectivity analysis is performed on the simplified spatial topology graph to ensure that each node in the simplified spatial topology graph has at least one valid connection edge. Connection edges are added to nodes that do not meet the connectivity requirements to obtain the spatial topology graph.

[0038] Optionally, the improved spatially perceptive Mamba neural network outputs salt tide intrusion prediction results and ecological water demand prediction results, specifically as follows:

[0039] Each node and its corresponding edge in the spatial topology connection graph are mapped to a data processing node and a data interaction path in the improved spatially aware Mamba neural network.

[0040] Meteorological data, upstream runoff data, and salinity observation data from the corresponding spatial grid node are input into each data processing node to form a real-time observation data sequence for the node.

[0041] The spatial awareness attention mechanism is used to determine the spatial interaction weight value of each node's real-time observation data sequence as it propagates between neighboring nodes;

[0042] By using a time sliding window mechanism, dynamic convolution calculations are performed on the real-time observation data sequence of nodes to obtain short-term predicted state sequences.

[0043] After each sliding window period, quantum state mapping is performed on the short-term predicted state sequence between adjacent spatial nodes, and a collaborative optimization signal for node states is generated based on the quantum state mapping result to update the parameters in the data processing nodes of the improved spatial perception Mamba neural network.

[0044] Using the updated and improved spatially aware Mamba neural network, the prediction results of salt tide intrusion and ecological water demand for each spatial node in the spatial topology connection graph during the drought and dry seasons are output respectively.

[0045] Optionally, the improved spatially aware Mamba neural network includes a data processing node initialization layer, a data interaction path construction layer, a spatially aware attention calculation layer, a temporal sliding window convolutional layer, a quantum state collaborative optimization layer, and a prediction result output layer.

[0046] The data processing node initialization layer is used to map each node in the spatial topology connection graph to a data processing node, and to receive meteorological data, upstream runoff data and salinity observation data collected in real time at the corresponding location of the node. After standardization processing, a real-time observation data sequence is formed.

[0047] The data interaction path construction layer is used to map each connection edge in the spatial topology connection graph to a data interaction path between data processing nodes, and to initialize the interaction intensity parameters on the data interaction path according to the water flow connectivity intensity and salt diffusion and transmission intensity between nodes.

[0048] The spatial awareness attention calculation layer uses the real-time observation data sequence of each data processing node as the query vector and the real-time observation data of adjacent nodes as the key vector and value vector. It calculates the cosine similarity between the query vector and the key vector, and normalizes the cosine similarity to determine the spatial interaction weight value on the data interaction path between data processing nodes.

[0049] The time-sliding window convolutional layer performs dynamic convolution calculations on the real-time observation data sequence of the data processing node based on the spatial interaction weight values ​​to obtain the short-term predicted state sequence of the data processing node.

[0050] The quantum state collaborative optimization layer uses a short-term predicted state sequence as the quantum initial state, uses qubits for state encoding, and uses a quantum entanglement operator to entangle the qubit states between adjacent data processing nodes pairwise, so that the predicted states of adjacent data processing nodes form quantum entangled states. The entanglement strength of the quantum entangled states between adjacent data processing nodes is measured, and the measurement results are converted into collaborative optimization signals for state interaction between nodes to update the network parameters in the data processing nodes.

[0051] The prediction result output layer is used to receive the updated network parameters of the data processing nodes, perform high-dimensional feature decoding through a multi-head attention mechanism, and output the salt tide intrusion prediction results and ecological water demand prediction results for each data processing node during the drought and dry seasons, respectively.

[0052] Optionally, the optimization of the salt tide intrusion prediction results and the ecological water demand prediction results based on the dynamic simulation results of salt tide in the estuary area is specifically as follows:

[0053] Based on the values ​​of tide level, flow velocity and salinity corresponding to each node in the dynamic simulation results of salt tide in the estuary area, random perturbation sampling is performed, and the sampled data is used as the input of the improved spatial perception Mamba neural network to generate a set of prediction results for each spatial grid node.

[0054] The standard deviation of the prediction result set for each spatial grid node is used as an index to characterize the uncertainty of the node prediction results. The uncertainty of tidal level prediction, current velocity prediction and salinity prediction are calculated.

[0055] An ensemble Kalman filter method was adopted, using the salt tide intrusion prediction results and ecological water demand prediction results output by the improved spatial perception Mamba neural network as state variables to establish the error covariance matrix of the network state variables, and the node prediction uncertainty index was introduced into the update calculation process of the error covariance matrix.

[0056] By using online data assimilation, the network parameters of each data processing node in the improved spatial awareness Mamba neural network are dynamically adjusted based on the updated calculation results of the error covariance matrix.

[0057] An improved spatially perceptive Mamba neural network with dynamic adjustment was used to generate optimized salt tide intrusion prediction results and ecological water demand prediction results for each spatial grid node in the estuary area.

[0058] Optionally, the optimized salt tide intrusion prediction results and ecological water demand prediction results are back-transmitted to the Fourier neural operator deep surrogate model to obtain updated simulation results of estuarine salt tide propagation, specifically as follows:

[0059] Based on the optimized salt tide intrusion prediction results and ecological water demand prediction results, the differences between the optimized prediction values ​​and the initial simulation values ​​are calculated respectively.

[0060] The difference between the optimized predicted value and the initial simulated value is back-propagated to the Fourier neural operator deep proxy model. The difference is used as a feedback adjustment term to update the constraint parameters corresponding to each spatial grid node in the Fourier neural operator deep proxy model through the error backpropagation mechanism.

[0061] Based on the updated constraint parameters of the Fourier neural operator deep proxy model, the loss function value of the Fourier neural operator deep proxy model is recalculated, and the model parameters are further optimized and adjusted.

[0062] The process of backpropagating errors and optimizing model parameters continues until the mean square error between the predicted tide level, flow velocity, and salinity values ​​generated by the Fourier neural operator deep surrogate model and the optimized predicted tide level, flow velocity, and salinity values ​​all meet the set convergence threshold, at which point the update stops.

[0063] Based on the Fourier neural operator deep surrogate model obtained after stopping updates, dynamic simulation results of salt tide in the estuary area were obtained after dynamically updating the constraints.

[0064] Optionally, based on the updated simulation results of estuarine salt tide propagation and the optimized ecological water demand prediction results, a multi-period ecological water replenishment plan is determined, specifically as follows:

[0065] Based on the dynamic simulation results of salt tide in the estuary area after dynamic updating of constraints, the spatial distribution values ​​and trends of salinity concentration of each spatial grid node in the Yangtze River Estuary and its downstream basin are extracted.

[0066] Based on the optimized ecological water demand forecast results, the predicted values ​​of ecological water demand and the temporal distribution characteristics of ecological water demand for each spatial grid node in the Yangtze River Estuary and its downstream basin are extracted respectively.

[0067] For the dry season and drought season, spatial grid nodes with a high risk of salt tide intrusion are identified respectively;

[0068] Spatial grid nodes whose ecological water demand forecasts exceed the ecological security threshold are designated as priority ecological water replenishment nodes, and ecological water replenishment time windows are determined for each priority ecological water replenishment node.

[0069] The nodes with high risk of salt tide intrusion and priority nodes for ecological water replenishment are identified as key locations for ecological water replenishment. The water demand for ecological water replenishment is calculated based on the degree of salinity exceeding the standard and the ecological water demand.

[0070] Under the constraint of total ecological water replenishment in multiple periods, based on the initial water demand, the initial water demand of each key ecological water replenishment location is optimized and allocated as a whole, and the actual ecological water replenishment allocation scheme of each key ecological water replenishment location is determined.

[0071] The beneficial effects of this invention are:

[0072] (1) This invention achieves dynamic and accurate simulation and prediction of salt tide intrusion and ecological water demand by synergistic fusion of physical constraint Fourier neural operator deep proxy model and quantum-enhanced spatial perception Mamba network, effectively improving the response frequency and decision accuracy of estuary ecological water replenishment decision, and enhancing the pertinence and reliability of ecological water replenishment strategy.

[0073] (2) By introducing quantum entanglement mechanism and spatial perception attention mechanism, combined with the closed-loop feedback mechanism of online data assimilation, this invention achieves high-precision dynamic optimization of salt tide intrusion and ecological water demand prediction, significantly improves the generalization ability of the prediction model, and shows better adaptability and accuracy in typical extreme hydrological scenarios such as drought and low water season.

[0074] (3) In terms of precise decision-making on ecological water replenishment in estuary areas at multiple periods, this invention effectively solves the shortcomings of existing technologies such as delayed response and insufficient accuracy in ecological water replenishment decision-making by using node-level risk identification, ecological water demand prediction and linear optimization water allocation mechanism. It breaks through the bottleneck of traditional numerical simulation and simple data-driven methods that cannot balance simulation efficiency and prediction accuracy, and achieves specific and significant progress in ecological water replenishment technology, effectively improving the application effect of ecological water replenishment assessment methods in complex estuary environments. Attached Figure Description

[0075] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0076] Figure 1 This is an overall flowchart of a multi-period water replenishment assessment method based on big data simulation proposed in this invention. Detailed Implementation

[0077] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0078] refer to Figure 1 A multi-period water replenishment assessment method based on big data simulation includes:

[0079] Taking the Yangtze River Estuary and its downstream basin as the research object, a basic numerical model of hydrodynamics and salinity diffusion was established, and historical observation data describing the tidal process of the estuary, the upstream runoff process, and the salinity change process during the dry season were obtained.

[0080] The simulation results of the basic numerical model are used as physical constraints, and the Fourier neural operator deep surrogate model is trained based on historical observation data. The Fourier neural operator deep surrogate model is optimized by the constraints composed of the mass conservation equation, the salt diffusion equation and the estuary hydrodynamic equation to obtain the dynamic simulation results of salt tide in the estuary area.

[0081] Based on the spatial distribution characteristics of the Yangtze River Estuary and its downstream basin, the area is divided into multiple spatial grid nodes, and a spatial topological connection map reflecting the hydrodynamic and salinity diffusion relationship between the nodes is constructed.

[0082] By using an improved spatial perception Mamba neural network, data interaction relationships between multiple spatial nodes are established based on a spatial topology connection graph. The network parameters are continuously updated based on real-time meteorological data, upstream runoff data, and salinity observation data collected in the estuary area. The network outputs salt tide intrusion prediction results and ecological water demand prediction results for multiple periods.

[0083] Using the dynamic simulation results of salt tide in the estuary area as input, the uncertainty of the simulation results is calculated by an improved spatially perceptive Mamba neural network. Based on the uncertainty data, the parameters of the improved spatially perceptive Mamba neural network are dynamically adjusted through online data assimilation to optimize the prediction results of salt tide intrusion and ecological water demand.

[0084] The optimized salt tide intrusion prediction results and ecological water demand prediction results are back-transmitted to the Fourier neural operator deep proxy model to dynamically update the constraints of the deep proxy model and obtain the updated simulation results of estuarine salt tide propagation.

[0085] Based on the updated simulation results of estuarine salt tide propagation and the optimized prediction results of ecological water demand, the specific location, timing, and water allocation scheme for ecological water replenishment are determined, and a multi-period ecological water replenishment assessment is completed.

[0086] In this embodiment, the establishment of a basic numerical model for hydrodynamics and salinity diffusion, and the acquisition of historical observational data describing estuarine tidal processes, upstream runoff processes, and salinity changes during the dry season, specifically involves:

[0087] For the Yangtze River Estuary and downstream basin, the boundary of the study area is dynamically delineated based on real-time tidal level monitoring data, and a spatial digital terrain model is constructed based on the lowest and highest ranges of river water levels during the dry season and drought season.

[0088] No fewer than 50 spatial grid nodes were selected as salinity monitoring stations within the study area. Each monitoring station continuously monitored the salinity values ​​at each stage of the dry season and drought period in real time on an hourly basis, and recorded the spatial location, collection time, and river hydrological conditions of each salinity value.

[0089] At the same time, no fewer than 10 continuous runoff observation sections were set up along the upstream direction to the river mouth in the study area. The upstream inflow was recorded in real time at intervals of no more than 30 minutes during each stage of the dry season and drought. The spatial location, observation time and corresponding working condition information of each section were also recorded simultaneously.

[0090] Real-time acquisition of tidal data from no less than 5 tidal observation stations in the estuary of the study area, with an observation frequency of no more than 15 minutes for each tidal station, and recording the spatial location, observation time, and corresponding tidal cycle stage of each tidal data point.

[0091] Based on monitoring and observation data, independent and targeted initial tidal level boundaries, inflow boundaries, and initial salinity distribution boundaries were set for different stages of the dry and arid periods. The governing equations consist of the mass conservation equation, momentum conservation equation, and salinity diffusion equation.

[0092]

[0093] Where h is the instantaneous water depth at any location in the estuary region, u and v are the instantaneous velocity components in the horizontal direction, t is the time variable in the model simulation process, x and y are the spatial coordinate variables along the longitudinal direction of the river channel, and Q in Q out These represent the inflow and outflow rates per unit area, respectively; g is the acceleration due to gravity; η is the instantaneous water level elevation; and C... d The overall bottom resistance coefficient is given by S, where S is the instantaneous salinity concentration and D is the salinity diffusion coefficient. inflow To determine the concentration change rate of the upstream inflow salinity source term, the finite volume method was used to spatially discretize the governing equations, and the fourth-order Runge-Kutta method was used for temporal discretization. These discretizations were then independently solved and simulated to verify the results, resulting in fundamental numerical models of hydrodynamics and salinity diffusion for different stages. The fourth-order Runge-Kutta method involves using the spatial discretization results of each variable in the discretized governing equations as initial input values. Four intermediate states are calculated sequentially at a set time step, and the predicted value for the next time step is obtained by weighted averaging of the four intermediate states. Within each time step, the boundary conditions and related parameters of each spatial grid node are updated using real-time monitoring data. The temporal discretization process is implemented independently for the hydrological scenarios of the dry season and drought season.

[0094] Historical data on estuary tide level, upstream runoff, and estuary salinity changes for no less than three consecutive years were obtained at each monitoring station and observation section. Historical observation data corresponding to each stage of the dry season and drought period were formed and systematically compared and verified with the basic numerical model of hydrodynamics and salinity diffusion.

[0095] In this embodiment, the simulation results of the basic numerical model are used as physical constraints, and a deep surrogate model of Fourier neural operators is trained based on historical observation data to obtain the dynamic simulation results of salt tide in the estuary area. Specifically:

[0096] Using the simulation output data of the basic numerical models of hydrodynamics and salinity diffusion established at each stage as initial constraints, the spatial distribution numerical results of tidal level, flow velocity and salinity of each spatial grid node in the Yangtze River Estuary and downstream basin at each stage are extracted to form the initial simulation dataset for training the Fourier neural operator deep surrogate model.

[0097] Based on three consecutive years of historical observation data, the historical measured values ​​of tide level, current velocity and salinity corresponding to the spatial grid nodes of the initial simulation dataset were extracted to form a historical observation dataset for training the Fourier neural operator deep proxy model.

[0098] Fourier neural operator network is used as the deep surrogate model to be trained. The position coordinates of each spatial grid node corresponding to the initial simulation dataset and the simulation time are used as the model input data. The initial simulation dataset and the historical observation dataset are used together as the training target dataset of the model.

[0099] The governing equations, consisting of the mass conservation equation, the salt diffusion equation, and the estuary hydrodynamic equation, are used as the physical constraints during the training process of the Fourier neural operator deep surrogate model.

[0100]

[0101] Where h is the instantaneous water depth, U is the instantaneous average flow velocity along the longitudinal direction of the river channel, t is the simulation time variable, x is the spatial coordinate along the longitudinal direction of the river channel, and Q... inflow Q represents the inflow rate per unit length of water entering the study area from the upstream of the river channel. outflow The outflow rate per unit length of the downstream study area is given by: g = gravitational acceleration, r = riverbed resistance coefficient, C = instantaneous salinity concentration, and K = salinity diffusion coefficient. inflow The rate of change in salinity concentration per unit volume of water entering the study area from the upstream of the river is calculated by comparing the deviations between the tidal level, flow velocity, and salinity values ​​predicted by the Fourier neural operator deep surrogate model and the corresponding historical measured values ​​in the training target dataset, and this deviation is used as the training loss function value.

[0102]

[0103] Where Loss is the loss function value, and N is the number of spatial grid nodes in the training dataset. Let be the instantaneous water depth value predicted for the i-th spatial grid node. Let be the instantaneous water depth value actually observed at the i-th spatial grid node. Let be the instantaneous velocity value predicted for the i-th spatial grid node. Let be the instantaneous flow velocity value actually observed at the i-th spatial grid node. Let be the instantaneous salinity concentration value predicted for the i-th spatial grid node. Let be the instantaneous salinity concentration value actually observed at the i-th spatial grid node, and α, β, γ be the weighting coefficients of the predicted and measured deviations of water depth, flow velocity, and salinity in the total error, respectively.

[0104] The parameters of the Fourier neural operator deep surrogate model are gradually updated using the backpropagation algorithm until the training loss function values ​​between the output tide level, current velocity, and salinity predictions and historical measured values ​​meet a preset threshold. Training then stops, resulting in a Fourier neural operator deep surrogate model capable of real-time dynamic simulation of tide level, current velocity, and salinity changes in the estuary region. The preset thresholds are: when the root mean square error between the output tide level prediction and historical measured tide level is less than 0.05, the root mean square error between the current velocity prediction and historical measured current velocity is less than 0.02, and the root mean square error between the salinity prediction and historical measured salinity concentration is less than 0.5, the preset thresholds are met, and training stops.

[0105] Based on the obtained Fourier neural operator deep surrogate model, the tidal level, flow velocity and salinity of each spatial grid node in the Yangtze River Estuary and downstream basin are simulated in real time to obtain the dynamic simulation results of salt tide in the estuary area.

[0106] In this embodiment, the spatial distribution characteristics of the Yangtze River Estuary and its downstream basin are divided into multiple spatial grid nodes, and a spatial topology connection map is constructed, specifically as follows:

[0107] Based on the spatial digital terrain model, with the river centerline as the baseline and the characteristics of river curvature, width variation, water depth variation and flow velocity variation as the dividing criteria, the study area of ​​the Yangtze River Estuary and its downstream basin is divided into no less than 200 spatial grid nodes.

[0108] For each spatial grid node, based on historical observation data, the water flow connectivity intensity and salt diffusion transport intensity between nodes are calculated to determine the strength of the hydrodynamic and salt diffusion relationship between every two spatial grid nodes:

[0109] Water flow connectivity between node i and node j for:

[0110]

[0111] Salt diffusion transport intensity between node i and node j for:

[0112]

[0113] in, For water flow connectivity intensity, Q represents the salt diffusion transport intensity. ij h represents the flow rate of water exchanged between nodes per unit time. i ,h j These represent the average water level elevations of the nodes, S i ,S j The average salinity concentrations of the nodes are d, respectively. ij denoted as the actual spatial distance between nodes along the centerline of the river channel, and D is the salt diffusion coefficient.

[0114] Using the node and edge representation method in graph theory, each spatial grid node is taken as a node in the topology graph, and the water flow connectivity strength and salt diffusion and transport strength between nodes are taken as the weight values ​​of the edges in the topology graph to establish an initial spatial topology connection graph.

[0115] Based on the salt diffusion and transport intensity weight value of each node in the initial spatial topology connection diagram, the average value of the salt diffusion and transport intensity weight of all connection edges in the spatial topology connection diagram is calculated using the weight threshold screening method. 0.6 to 0.8 times this average value is taken as the predetermined weight threshold. Connection edges between nodes with weights exceeding the predetermined threshold are retained, and connection edges between nodes with weights below the predetermined threshold are deleted, resulting in a simplified spatial topology connection diagram that clearly reflects the main hydrodynamic and salt diffusion relationships between nodes.

[0116] A connectivity analysis is performed on the simplified spatial topology graph to ensure that each node in the simplified spatial topology graph has at least one valid connection edge. Connection edges are added to nodes that do not meet the connectivity requirements to obtain the spatial topology graph.

[0117] In this embodiment, the improved spatially perceptive Mamba neural network establishes data interaction relationships between multiple spatial nodes based on a spatial topology connection graph. It continuously updates network parameters based on real-time data collected from the estuary region and outputs salt tide intrusion prediction results and ecological water demand prediction results. Specifically:

[0118] Each node in the spatial topology connection graph is mapped to a data processing node in the improved spatially aware Mamba neural network, and each connection edge in the spatial topology connection graph is mapped to a data interaction path between nodes in the improved spatially aware Mamba neural network.

[0119] For each data processing node in the improved spatial perception Mamba neural network, meteorological data, upstream runoff data, and salinity observation data of the corresponding spatial grid node location are input, and a real-time observation data sequence of the node is formed in hourly units.

[0120] For the real-time observation data sequence of each data processing node, the spatial interaction weight value of each node's real-time observation data sequence is determined by the spatial awareness attention mechanism in the improved spatial awareness Mamba neural network, so as to realize the dynamic adjustment of the data interaction intensity between nodes. The spatial awareness attention mechanism uses the real-time observation data of each node as the query vector of the attention mechanism, and the real-time observation data of adjacent nodes as the key vector and the value vector. The cosine similarity between the query vector and the key vector is calculated, and the calculated cosine similarity is normalized and then determined as the spatial interaction weight value.

[0121] Based on the spatial interaction weight values ​​between data processing nodes, the improved spatial perception Mamba neural network uses a time sliding window mechanism with a 24-hour sliding window period. The sliding window is moved forward once every hour, and dynamic convolution calculation is performed on the real-time observation data sequence of the nodes within each sliding window period. This captures the time dependence characteristics of the propagation of salt tide intrusion and ecological water demand in the estuary area between spatial nodes, and obtains the short-term predicted state sequence of spatial nodes.

[0122] By introducing a quantum entanglement mechanism, after each sliding window period, a quantum state mapping is performed on the short-term predicted state sequence between adjacent spatial nodes. The short-term predicted state sequence of each spatial node is used as the quantum initial state. Quantum bits are used for state encoding. The quantum entanglement operator is used to entangle the quantum bit states between adjacent spatial nodes pairwise, so that the predicted states of adjacent nodes form quantum entangled states. The entanglement strength of the quantum entangled states between adjacent nodes is measured, and the measurement result is converted into a cooperative optimization signal for the state interaction between nodes. Quantum interaction states between spatial nodes are generated, and cooperative optimization signals for node states are generated based on the quantum interaction states to update the parameters in the data processing nodes of the improved spatial perception Mamba neural network.

[0123] Using the updated and improved spatially aware Mamba neural network, the prediction results of salt tide intrusion and ecological water demand for each spatial node in the spatial topology connection graph during the drought and dry seasons are output respectively.

[0124] In this embodiment, the improved spatially aware Mamba neural network includes a data processing node initialization layer, a data interaction path construction layer, a spatially aware attention calculation layer, a temporal sliding window convolutional layer, a quantum state collaborative optimization layer, and a prediction result output layer.

[0125] The data processing node initialization layer is used to map each node in the spatial topology connection graph to a data processing node, and to receive meteorological data, upstream runoff data and salinity observation data collected in real time at the corresponding location of the node. After standardization processing, it forms a real-time observation data sequence in hourly units.

[0126] The data interaction path construction layer is used to map each connection edge in the spatial topology connection graph to a data interaction path between data processing nodes, and to initialize the interaction intensity parameters on the data interaction path according to the water flow connectivity intensity and salt diffusion and transmission intensity between nodes.

[0127] The spatial awareness attention calculation layer uses the real-time observation data sequence of each data processing node as the query vector and the real-time observation data of adjacent nodes as the key vector and value vector. It calculates the cosine similarity between the query vector and the key vector, and normalizes the cosine similarity to determine the spatial interaction weight value on the data interaction path between data processing nodes, so as to realize the dynamic adjustment of the data interaction intensity between nodes.

[0128] The time-sliding window convolutional layer uses 24 hours as a sliding window period, moves the sliding window forward once every hour, and performs dynamic convolution calculation on the real-time observation data sequence of the data processing node based on the spatial interaction weight value within each sliding window period. This captures the time dependence characteristics of salt tide intrusion and ecological water demand in the estuary area propagating between spatial nodes, and obtains the short-term predicted state sequence of the data processing node.

[0129] The quantum state collaborative optimization layer uses the short-term predicted state sequence of the data processing node as the quantum initial state, uses qubits for state encoding, and uses a quantum entanglement operator to entangle the qubit states between adjacent data processing nodes pairwise, so that the predicted states of adjacent data processing nodes form a quantum entangled state. The entanglement strength of the quantum entangled state between adjacent data processing nodes is measured, and the measurement result is converted into a collaborative optimization signal for state interaction between nodes to update the network parameters in the data processing node.

[0130] The prediction result output layer is used to receive the updated network parameters of the data processing nodes, perform high-dimensional feature decoding through a multi-head attention mechanism, and output the salt tide intrusion prediction results and ecological water demand prediction results for each data processing node during the drought and dry seasons, respectively.

[0131] In this embodiment, the dynamic simulation results of salt tide in the estuary area are used as input. An improved spatially perceptive Mamba neural network is used to calculate the uncertainty of the simulation results. Based on the uncertainty data, the parameters of the improved spatially perceptive Mamba neural network are dynamically adjusted through online data assimilation to optimize the salt tide intrusion prediction results and the ecological water demand prediction results. Specifically:

[0132] The Monte Carlo sampling method was adopted, and the values ​​of tide level, flow velocity and salinity corresponding to each node in the dynamic simulation results of salt tide in the estuary area were used as the benchmark. No less than 100 random perturbation samplings were carried out, and the sampled data were used as the input of the improved spatial perception Mamba neural network to generate the prediction result set for each spatial grid node.

[0133] The standard deviation of the prediction result set for each spatial grid node is used as an index to characterize the uncertainty of the node prediction results. The uncertainty of the tidal level prediction, the uncertainty of the current velocity prediction, and the uncertainty of the salinity prediction are calculated.

[0134] The predicted uncertainty values ​​of node tide level, flow velocity and salinity are used as inputs to the online data assimilation process. The ensemble Kalman filter method is adopted, and the salt tide intrusion prediction results and ecological water demand prediction results output by the improved spatial perception Mamba neural network are used as state variables to establish the error covariance matrix of the network state variables. The node prediction uncertainty index is introduced into the update calculation process of the error covariance matrix.

[0135] By using online data assimilation, the network parameters of each data processing node in the improved spatial perception Mamba neural network are dynamically adjusted based on the updated calculation results of the error covariance matrix, so as to reduce the uncertainty of the network prediction results.

[0136] An improved spatially perceptive Mamba neural network with dynamic adjustment was used to generate optimized salt tide intrusion prediction results and ecological water demand prediction results for each spatial grid node in the estuary area.

[0137] In this embodiment, the optimized salt tide intrusion prediction results and ecological water demand prediction results are back-transmitted to the Fourier neural operator deep proxy model to dynamically update the constraints of the deep proxy model and obtain the updated dynamic simulation results of salt tide in the estuary area, specifically:

[0138] Based on the optimized prediction results of salt tide intrusion and ecological water demand, the optimized prediction values ​​of tide level, flow velocity and salinity corresponding to each spatial grid node are extracted respectively.

[0139] Using the initial simulated values ​​of tidal level, current velocity, and salinity generated by the corresponding spatial grid nodes in the Fourier neural operator deep surrogate model as a benchmark, the differences between the optimized predicted values ​​and the initial simulated values ​​are calculated respectively to obtain the tidal level difference, current velocity difference, and salinity difference for each spatial grid node.

[0140] The tidal level difference, velocity difference, and salinity difference of each spatial grid node are transmitted in reverse to the Fourier neural operator deep proxy model. The difference is used as a feedback adjustment term, and the constraint parameters corresponding to each spatial grid node in the Fourier neural operator deep proxy model are updated through the error backpropagation mechanism.

[0141] Based on the updated constraint parameters of the Fourier neural operator deep surrogate model, the loss function value of the Fourier neural operator deep surrogate model is recalculated. The mean square error between the tide level prediction, current velocity prediction, and salinity prediction values ​​generated by the Fourier neural operator deep surrogate model and the optimized tide level prediction, current velocity prediction, and salinity prediction values ​​is used as the loss function value, and the model parameters are further optimized and adjusted.

[0142] The error backpropagation update and model parameter optimization adjustment process is continuously executed until the mean square error between the predicted tide level, current velocity and salinity generated by the Fourier neural operator deep proxy model and the optimized predicted tide level, current velocity and salinity all meet the set convergence threshold, at which point the update stops.

[0143] Based on the Fourier neural operator deep surrogate model obtained after the update stops, the dynamic simulation results of salt tide in the estuary area after dynamic updating of constraints are obtained, which serve as input data for the subsequent determination of ecological water replenishment strategies.

[0144] In this embodiment, based on the updated refined simulation results of salt tide propagation and the optimized ecological water demand prediction results, the specific location, time, and water allocation scheme for ecological water replenishment are determined, and a multi-period ecological water replenishment assessment is completed, specifically as follows:

[0145] Based on the dynamic simulation results of salt tide in the estuary area after dynamic updating of constraints, the spatial distribution values ​​of salinity concentration and their time-varying trends of each spatial grid node in the Yangtze River Estuary and its downstream basin are extracted.

[0146] Based on the optimized ecological water demand forecast results, the predicted values ​​of ecological water demand and the temporal distribution characteristics of ecological water demand for each spatial grid node in the Yangtze River Estuary and its downstream basin are extracted respectively.

[0147] For the dry season and drought season, spatial grid nodes with high risk of salt tide intrusion are identified respectively. The length of time that the predicted salinity concentration of each spatial grid node exceeds the ecological risk threshold is used as the judgment criterion to determine the location of spatial grid nodes with high risk of salt tide intrusion during the dry season and drought season, and the start time, duration and corresponding node location of the salt tide intrusion risk are clearly marked.

[0148] Spatial grid nodes whose ecological water demand prediction values ​​exceed the ecological security threshold are designated as priority nodes for ecological water replenishment. The ecological water replenishment time window corresponding to each priority node is determined. Based on the start and end times when the ecological water demand prediction values ​​of the nodes exceed the ecological security threshold, the start and end times of ecological water replenishment are determined, thus obtaining the specific time window for ecological water replenishment.

[0149] The nodes with high risk of salt tide intrusion and the nodes with priority for ecological water replenishment are identified as key locations for ecological water replenishment. For each key location, the water demand for ecological water replenishment is calculated based on the degree of salinity exceeding the standard and the ecological water demand. The amount of dilution water required for the degree of salinity exceeding the standard and the amount of water replenishment required for the predicted ecological water demand are calculated separately for each key location. The two are then summed to obtain the initial water demand for the key location.

[0150] Under the constraint of total ecological water replenishment in multiple periods, based on the initial water demand, the initial water demand of each key ecological water replenishment location is optimized and allocated in an overall manner through linear programming optimization method. The actual ecological water replenishment allocation scheme of each key ecological water replenishment location is determined, and the specific ecological water replenishment location, ecological water replenishment time window and ecological water replenishment allocation scheme are determined, forming a multi-period ecological water replenishment assessment scheme for dry season and drought period.

[0151] Example 1:

[0152] To verify the feasibility of this invention in practice, it was applied to an ecological water replenishment decision support task in a river estuary basin. Targeting the typical environmental scenario of frequent salt tide intrusions and significant abrupt changes in ecological water demand during drought and low water seasons, multi-period ecological water replenishment assessments were conducted. In this practical application scenario, traditional numerical models require substantial computational resources, with a single simulation typically exceeding 5 hours. This makes it impossible to frequently and dynamically update the predictions of salt tide intrusion and ecological water demand, leading to delayed ecological water replenishment decisions, difficulty in effectively controlling the risk of salt tide intrusion, and inaccurate assurance of the effectiveness of ecological water replenishment.

[0153] In practice, a basic numerical model of hydrodynamics and salinity diffusion in the estuary area was first established. Real-time monitoring data, including estuary tide level, upstream runoff and river salinity, were used. Based on long-term observation data, a digital terrain model covering no less than 200 spatial grid nodes in the estuary basin was constructed. Parameters such as water level, salinity and flow velocity of the nodes were continuously monitored and recorded in real time to build a high-precision basic numerical model.

[0154] Subsequently, using the simulation results of the basic numerical model as initial constraints, and based on three consecutive years of historical observation data, historical measured data of tide level, current velocity, and salinity corresponding to spatial grid nodes were extracted to train a physical constraint-based Fourier neural operator deep surrogate model. Target error thresholds were set for model training, with tide level error less than 0.05 meters, current velocity error less than 0.02 meters per second, and salinity error less than 0.5‰. After 42 rounds of training iterations, the surrogate model error reached the convergence threshold, the average simulation time was reduced to less than 5 seconds, and the accuracy met the preset standard.

[0155] Meanwhile, based on the spatial digital terrain model, no less than 200 spatial grid nodes are divided to construct an initial spatial topology connection map. The water flow connectivity intensity and salt diffusion and transmission intensity between nodes are calculated through historical data between nodes. The weight threshold screening method is used to remove and adjust the connection edges with low weights to obtain a clear and accurate spatial topology connection map.

[0156] An improved spatially aware Mamba neural network was built based on this spatial topology connection graph. This network includes a data processing node initialization layer, a data interaction path construction layer, a spatially aware attention computation layer, a temporal sliding window convolutional layer, a quantum state collaborative optimization layer, and a prediction result output layer. The temporal sliding window length for the spatially aware attention mechanism was set to 24 hours, with a step size of 1 hour. During model prediction, the spatial interaction weights of data propagation between nodes were calculated using real-time data sequences. The network parameters were optimized at the end of each sliding window period through a quantum entanglement mechanism, effectively improving the network's prediction accuracy and generalization performance.

[0157] Subsequently, the dynamic simulation results of estuarine salinity tides output from the Fourier neural operator deep surrogate model were used as input. A Monte Carlo sampling method was employed to randomly perturb the prediction results of each node 100 times, and the uncertainty indices of tidal level, flow velocity, and salinity at each node were calculated using the standard deviation of the prediction result set. The parameters of the improved spatially aware Mamba network were dynamically adjusted using ensemble Kalman filtering. After 20 online assimilation adjustments, the network prediction uncertainty indices decreased significantly, with an average reduction of 38.6% in tidal level prediction error, 41.2% in flow velocity error, and 45.9% in salinity error.

[0158] The improved salt tide intrusion prediction and ecological water demand prediction results are then transmitted back to the Fourier neural operator surrogate model for feedback parameter updates. The feedback updates are continuously executed until the mean square error between the model output and the actual observations reaches the preset convergence threshold, thus forming a high-precision refined simulation result of salt tide propagation.

[0159] Based on the refined simulation results and optimized ecological water demand prediction results, the locations of nodes with high risk of salt tide intrusion and the locations of nodes with priority for ecological water replenishment were determined. The duration of salinity exceeding the standard, the start and end times of ecological water replenishment were clearly recorded, the initial ecological water replenishment demand of each node was calculated, and then the water allocation scheme was optimized as a whole using the linear programming optimization method to clarify the specific ecological water replenishment location, time window and water allocation.

[0160] To verify the actual implementation effect, the table below shows a comparison between the prediction results and the measured results of the method of the present invention at some nodes.

[0161] Table 1 Comparison of Predicted and Measured Performance of Some Nodes in the Estuary Area

[0162]

[0163] As can be clearly seen from the data in Table 1 above, the method of this invention achieves a high level of accuracy in predicting tide level, flow velocity, salinity, and ecological water demand. For example, at node C-076, the prediction error for tide level is only 0.02 meters, the prediction error for flow velocity is only 0.02 meters per second, the prediction error for salinity is only 0.2‰, and the prediction error for ecological water demand is only 0.4 m³ / s, which is significantly better than the error range of traditional numerical models and single deep learning models. The prediction errors of other nodes are also significantly lower than the error range of traditional technologies, demonstrating the high accuracy and generalization ability of this invention in predicting salt tide intrusion and dynamically simulating ecological water demand.

[0164] Based on the results of the above embodiments, the present invention demonstrates significant advantages in ecological water replenishment decision support tasks in estuary areas, exhibiting high prediction accuracy, fast response frequency, and short simulation time. This fully verifies the technical applicability and practical value of the present invention in the field of high-frequency ecological water replenishment decision-making, effectively solving the problems of insufficient accuracy and long computation time of traditional methods, and greatly improving the accuracy and efficiency of ecological water replenishment decision-making, thus possessing broad application prospects.

[0165] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-period water replenishment assessment method based on big data simulation, characterized in that, include: Establish a basic numerical model of hydrodynamics and salinity diffusion, and obtain historical observation data of the Yangtze River Estuary and its downstream basin; The simulation results of the basic numerical model were used as physical constraints, and a deep surrogate model of Fourier neural operators was trained based on historical observation data to obtain the dynamic simulation results of salt tide in the estuary area. Based on the spatial distribution characteristics of the Yangtze River Estuary and its downstream basin, the area is divided into multiple spatial grid nodes, and a spatial topology connection map is constructed. By using an improved spatially perceptive Mamba neural network, data interaction relationships between multiple spatial nodes are established based on a spatial topology connection graph. The network parameters are continuously updated based on real-time data collected in the estuary area, and the prediction results of salt tide intrusion and ecological water demand are output. Based on the dynamic simulation results of salt tide in the estuary area, the prediction results of salt tide intrusion and ecological water demand are optimized. The optimized salt tide intrusion prediction results and ecological water demand prediction results are back-transmitted to the Fourier neural operator deep proxy model to obtain the updated simulation results of estuarine salt tide propagation. Based on the updated simulation results of estuarine salt tide propagation and the optimized ecological water demand prediction results, a multi-period ecological water replenishment plan was determined.

2. The multi-period water replenishment assessment method based on big data simulation according to claim 1, characterized in that, The establishment of a fundamental numerical model for hydrodynamics and salinity diffusion, and the acquisition of historical observation data of the Yangtze River Estuary and its downstream basin, specifically involves: For the Yangtze River Estuary and downstream basin, the boundary of the study area is dynamically delineated based on real-time tidal level monitoring data, and a spatial digital terrain model is constructed based on the lowest and highest ranges of river water levels during the dry season and drought season. Salinity monitoring stations were selected within the study area, and salinity values ​​were continuously monitored at various stages of the dry and arid seasons. Simultaneously, continuous runoff observation sections were set up along the upstream direction to the river mouth within the study area, and the upstream inflow was continuously recorded at each stage of the dry season and drought period. Real-time acquisition of tide level data from estuary tide observation stations in the study area; Based on monitoring and observation data, initial tidal level boundaries, inflow boundaries, and initial salinity distribution boundaries were set for different stages of the dry season and drought season, and independent solutions and simulation verifications were performed to form a basic numerical model of hydrodynamics and salinity diffusion for different stages. Historical data on changes in estuary tide level, upstream runoff, and estuary salinity for no less than three consecutive years were obtained at each monitoring station and observation section to form historical observation data.

3. The multi-period water replenishment assessment method based on big data simulation according to claim 1, characterized in that, The deep surrogate model based on historical observation data, trained with Fourier neural operators, yields dynamic simulation results of salt tide in the estuary region, specifically: Using the simulation results of the basic numerical model as initial constraints, simulation data are extracted to form an initial simulation dataset; Based on historical observation data, the measured values ​​corresponding to the spatial grid nodes of the initial simulation dataset are extracted to form the historical observation dataset. Fourier neural operator networks are used as deep proxy models to be trained; The position coordinates of each spatial grid node and the simulation time corresponding to the initial simulation dataset are used as the model input data; The initial simulation dataset and the historical observation dataset were used together as the training target dataset for the model; The set of governing equations, consisting of the mass conservation equation, the salt diffusion equation, and the estuary hydrodynamic equation, serves as the physical constraints during the training process. The deviation between the tide level, current velocity, and salinity values ​​predicted by the deep proxy model and the corresponding historical measured values ​​in the training target dataset is used as the training loss function value. The parameters of the deep proxy model are updated step by step through the backpropagation algorithm until the training loss function value meets the preset threshold, at which point training stops and the Fourier neural operator deep proxy model is obtained. Based on the Fourier neural operator deep surrogate model, the tidal level, current velocity and salinity of each spatial grid node in the Yangtze River Estuary and downstream basin are simulated in real time to obtain the dynamic simulation results of salt tide in the estuary area.

4. The multi-period water replenishment assessment method based on big data simulation according to claim 1, characterized in that, Based on the spatial distribution characteristics of the Yangtze River Estuary and its downstream basin, the area is divided into multiple spatial grid nodes, and a spatial topology connection diagram is constructed, specifically as follows: Based on the spatial digital terrain model, the study area is divided into spatial grid nodes; For each spatial grid node, the water flow connectivity intensity and salt diffusion and transport intensity between nodes are calculated based on historical observation data. Each spatial grid node is used as a node in the topology graph, and the water flow connectivity strength and salt diffusion and transport strength between nodes are used as the weight values ​​of the edges in the topology graph to establish an initial spatial topology connection graph. Based on the salt diffusion and transport intensity weight value of each node in the initial spatial topology connection graph, a simplified spatial topology connection graph is obtained by using the weight threshold screening method. A connectivity analysis is performed on the simplified spatial topology graph to ensure that each node in the simplified spatial topology graph has at least one valid connection edge. Connection edges are added to nodes that do not meet the connectivity requirements to obtain the spatial topology graph.

5. The multi-period water replenishment assessment method based on big data simulation according to claim 1, characterized in that, The improved spatially perceptive Mamba neural network outputs salt tide intrusion prediction results and ecological water demand prediction results, specifically as follows: Each node and its corresponding edge in the spatial topology connection graph are mapped to a data processing node and a data interaction path in the improved spatially aware Mamba neural network. Meteorological data, upstream runoff data, and salinity observation data from the corresponding spatial grid node are input into each data processing node to form a real-time observation data sequence for the node. The spatial awareness attention mechanism is used to determine the spatial interaction weight value of each node's real-time observation data sequence as it propagates between neighboring nodes; By using a time sliding window mechanism, dynamic convolution calculations are performed on the real-time observation data sequence of nodes to obtain short-term predicted state sequences. After each sliding window period, quantum state mapping is performed on the short-term predicted state sequence between adjacent spatial nodes, and a collaborative optimization signal for node states is generated based on the quantum state mapping result to update the parameters in the data processing nodes of the improved spatial perception Mamba neural network. Using the updated and improved spatially aware Mamba neural network, the prediction results of salt tide intrusion and ecological water demand for each spatial node in the spatial topology connection graph during the drought and dry seasons are output respectively.

6. The multi-period water replenishment assessment method based on big data simulation according to claim 5, characterized in that, The improved spatially aware Mamba neural network includes a data processing node initialization layer, a data interaction path construction layer, a spatially aware attention calculation layer, a temporal sliding window convolutional layer, a quantum state collaborative optimization layer, and a prediction result output layer. The data processing node initialization layer is used to map each node in the spatial topology connection graph to a data processing node, and to receive meteorological data, upstream runoff data and salinity observation data collected in real time at the corresponding location of the node. After standardization processing, a real-time observation data sequence is formed. The data interaction path construction layer is used to map each connection edge in the spatial topology connection graph to a data interaction path between data processing nodes, and to initialize the interaction intensity parameters on the data interaction path according to the water flow connectivity intensity and salt diffusion and transmission intensity between nodes. The spatial awareness attention calculation layer uses the real-time observation data sequence of each data processing node as the query vector and the real-time observation data of adjacent nodes as the key vector and value vector. It calculates the cosine similarity between the query vector and the key vector, and normalizes the cosine similarity to determine the spatial interaction weight value on the data interaction path between data processing nodes. The time-sliding window convolutional layer performs dynamic convolution calculations on the real-time observation data sequence of the data processing node based on the spatial interaction weight values ​​to obtain the short-term predicted state sequence of the data processing node. The quantum state collaborative optimization layer uses a short-term predicted state sequence as the quantum initial state, uses qubits for state encoding, and uses a quantum entanglement operator to entangle the qubit states between adjacent data processing nodes pairwise, so that the predicted states of adjacent data processing nodes form quantum entangled states. The entanglement strength of the quantum entangled states between adjacent data processing nodes is measured, and the measurement results are converted into collaborative optimization signals for state interaction between nodes to update the network parameters in the data processing nodes. The prediction result output layer is used to receive the updated network parameters of the data processing nodes, perform high-dimensional feature decoding through a multi-head attention mechanism, and output the salt tide intrusion prediction results and ecological water demand prediction results for each data processing node during the drought and dry seasons, respectively.

7. The multi-period water replenishment assessment method based on big data simulation according to claim 1, characterized in that, The optimization of salt tide intrusion prediction and ecological water demand prediction results based on the dynamic simulation results of salt tide in the estuary area is as follows: Based on the values ​​of tide level, flow velocity and salinity corresponding to each node in the dynamic simulation results of salt tide in the estuary area, random perturbation sampling is performed, and the sampled data is used as the input of the improved spatial perception Mamba neural network to generate a set of prediction results for each spatial grid node. The standard deviation of the prediction result set for each spatial grid node is used as an index to characterize the uncertainty of the node prediction results. The uncertainty of tidal level prediction, current velocity prediction and salinity prediction are calculated. An ensemble Kalman filter method was adopted, using the salt tide intrusion prediction results and ecological water demand prediction results output by the improved spatial perception Mamba neural network as state variables to establish the error covariance matrix of the network state variables, and the node prediction uncertainty index was introduced into the update calculation process of the error covariance matrix. By using online data assimilation, the network parameters of each data processing node in the improved spatial awareness Mamba neural network are dynamically adjusted based on the updated calculation results of the error covariance matrix. An improved spatially perceptive Mamba neural network with dynamic adjustment was used to generate optimized salt tide intrusion prediction results and ecological water demand prediction results for each spatial grid node in the estuary area.

8. The multi-period water replenishment assessment method based on big data simulation according to claim 1, characterized in that, The optimized salt tide intrusion prediction results and ecological water demand prediction results are then transmitted back to the Fourier neural operator deep proxy model to obtain updated simulation results of estuarine salt tide propagation. Specifically: Based on the optimized salt tide intrusion prediction results and ecological water demand prediction results, the differences between the optimized prediction values ​​and the initial simulation values ​​are calculated respectively. The difference between the optimized predicted value and the initial simulated value is back-propagated to the Fourier neural operator deep proxy model. The difference is used as a feedback adjustment term to update the constraint parameters corresponding to each spatial grid node in the Fourier neural operator deep proxy model through the error backpropagation mechanism. Based on the updated constraint parameters of the Fourier neural operator deep proxy model, the loss function value of the Fourier neural operator deep proxy model is recalculated, and the model parameters are further optimized and adjusted. The process of backpropagating errors and optimizing model parameters continues until the mean square error between the predicted tide level, flow velocity, and salinity values ​​generated by the Fourier neural operator deep surrogate model and the optimized predicted tide level, flow velocity, and salinity values ​​all meet the set convergence threshold, at which point the update stops. Based on the Fourier neural operator deep surrogate model obtained after stopping updates, dynamic simulation results of salt tide in the estuary area were obtained after dynamically updating the constraints.

9. The multi-period water replenishment assessment method based on big data simulation according to claim 1, characterized in that, Based on the updated simulation results of estuarine salt tide propagation and the optimized ecological water demand prediction results, a multi-period ecological water replenishment scheme is determined, specifically as follows: Based on the dynamic simulation results of salt tide in the estuary area after dynamic updating of constraints, the spatial distribution values ​​and trends of salinity concentration of each spatial grid node in the Yangtze River Estuary and its downstream basin are extracted. Based on the optimized ecological water demand forecast results, the predicted values ​​of ecological water demand and the temporal distribution characteristics of ecological water demand for each spatial grid node in the Yangtze River Estuary and its downstream basin are extracted respectively. For the dry season and drought season, spatial grid nodes with a high risk of salt tide intrusion are identified respectively; Spatial grid nodes whose ecological water demand forecasts exceed the ecological security threshold are designated as priority ecological water replenishment nodes, and ecological water replenishment time windows are determined for each priority ecological water replenishment node. The nodes with high risk of salt tide intrusion and priority nodes for ecological water replenishment are identified as key locations for ecological water replenishment. The water demand for ecological water replenishment is calculated based on the degree of salinity exceeding the standard and the ecological water demand. Under the constraint of total ecological water replenishment in multiple periods, based on the initial water demand, the initial water demand of each key ecological water replenishment location is optimized and allocated as a whole, and the actual ecological water replenishment allocation scheme of each key ecological water replenishment location is determined.

Images