Water conservancy project intelligent design method and system based on multi-source heterogeneous data fusion

Abstract

The application relates to the technical field of hydraulic engineering, and particularly discloses a water conservancy project intelligent design method and system based on multi-source heterogeneous data fusion, which acquires meteorological radar rainfall prediction data and river network hydrodynamic monitoring data, carries out time-space alignment and cleaning to obtain a multi-source data set; adopts implicit neural representation to supervise discrete rainfall data, constructs a continuous rainfall space-time function; constructs a physically enhanced neural differential equation, embeds the continuous rainfall space-time function as an external driving term, and ensures the consistency of dynamic mapping and physical laws through residual constraint of a hydrodynamic control equation; uses the hydrodynamic monitoring data and the hydrodynamic residual at random space-time points to construct a physical constraint, and iteratively optimizes dynamic mapping parameters; uses the optimized dynamic mapping as an environment simulator, combines the probability distribution of random interference events, and solves a gate opening and closing time sequence meeting safety constraints through distributed robust optimization; and the application realizes deep fusion of multi-source data and hydrodynamic laws.

Description

Technical Field

[0001] This invention relates to the field of water conservancy engineering technology, specifically to a method and system for intelligent design of water conservancy projects based on multi-source heterogeneous data fusion. Background Technology

[0002] With the rapid development of meteorological radar observation technology, automatic hydrological monitoring and forecasting technology, and numerical simulation methods, the field of water conservancy engineering planning and design has accumulated an increasing amount of multi-source heterogeneous data. How to effectively integrate these data from different sources, in different formats, and with different sampling frequencies, and apply them to the intelligent design of pump and gate group joint scheduling schemes, has become a research hotspot in the current water conservancy engineering field.

[0003] Existing water conservancy scheduling methods based on physical information neural networks generally face a fundamental conflict between the hard constraints of physical laws and the sparsity of data when integrating multi-source data and hydraulic equations. When local geological exploration data is missing or random disturbances such as weeds in the upstream flow cannot be covered by sensing equipment, the neural network will forcibly smooth these abrupt regions not covered by data during the training process in order to strictly meet the mathematical continuity requirements of the Saint-Venant equations. This results in the generated scheduling scheme perfectly conforming to the physical conservation laws in mathematics, but in actual engineering, it happens to avoid the real weak zones of the dikes or the sections blocked by debris. The physical equations are transformed from the original "safety guardrail" into a "perfect disguise" to cover up the real engineering risks, ultimately causing the scheduling instructions to fail under extreme conditions and resulting in the consequences of dike overflow or waterlogging. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for intelligent design of water conservancy projects based on the fusion of multi-source heterogeneous data, so as to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] The intelligent design method for water conservancy projects based on multi-source heterogeneous data fusion includes the following steps:

[0007] S1: Acquire meteorological radar rainfall forecast data and river network hydrodynamic monitoring data, and perform spatiotemporal coordinate system I and outlier cleaning processing to obtain an aligned multi-source dataset;

[0008] S2: Using implicit neural representations and discrete rainfall data from multi-source datasets as supervision, a continuous spatiotemporal function for rainfall that can output rainfall intensity at any spatiotemporal location is constructed.

[0009] S3: Construct a physically enhanced neural differential equation, embedding the spatiotemporal function of continuous rainfall as an external driving term into the neural differential equation to describe the dynamic mapping of the river network hydraulic state over time. The internal structure of the neural differential equation is ensured to be consistent with physical laws by introducing residual constraints from the hydraulic control equation.

[0010] S4: Using hydrodynamic monitoring data from multiple sources as the fitting target, and constructing physical constraints based on the hydrodynamic residuals at random spatiotemporal points of the dynamic mapping, the parameters of the dynamic mapping are iteratively optimized to obtain the optimized dynamic mapping;

[0011] S5: The optimized dynamic mapping is used as an environment simulator. Combined with the probability distribution of random interference events in the upstream flow, the gate opening and closing sequence that meets the preset safety constraints is solved by a robust optimization algorithm, which serves as the design scheme for joint scheduling of pump and gate groups.

[0012] As a further aspect of the present invention: S2 specifically includes:

[0013] A learnable continuous mapping function is constructed with spatiotemporal coordinates as input and rainfall intensity as output. The mapping function uses multi-resolution hash coding to embed features of the input coordinates and processes the embedded features through multiple cascaded differentiable transformation layers. Each transformation layer contains a linear transformation and a sinusoidal periodic activation function.

[0014] Discrete rainfall data from a multi-source dataset are used as supervision to construct a loss function to measure the deviation between the output of the mapping function and the discrete observations;

[0015] Based on the loss function, the parameters of the mapping function are adjusted using the gradient descent method until convergence, resulting in a continuous spatiotemporal function that can output the rainfall intensity at any spatiotemporal location.

[0016] As a further aspect of the present invention: the adjustment of the parameters of the mapping function based on the loss function and using the gradient descent method specifically includes:

[0017] Based on the bias value output by the loss function, the parameter gradient of the learnable mapping function in the spatiotemporal function of continuous rainfall is calculated;

[0018] Based on the spatiotemporal distribution characteristics of the parameter gradient, the sampling points are adaptively densified in local spatiotemporal regions where the loss value exceeds the threshold, and the parameter gradient contribution of the corresponding region is recalculated.

[0019] Combining the frequency response characteristics of the sinusoidal periodic activation function, the parameters are updated using an adaptive learning rate associated with the current iteration step. After the parameter update, the decrease in the loss function is judged. If the decrease in the loss function is less than a preset value for multiple consecutive iteration steps, the iteration is terminated.

[0020] As a further aspect of the present invention: S3 specifically includes:

[0021] Construct a differential relation expression with the hydraulic state of the river network as the dependent variable and the time coordinate as the independent variable. The differential relation expression contains a learnable mapping structure between the rate of change of the state, the current state, and the external driving terms.

[0022] The output of the spatiotemporal function of continuous rainfall is used as an external driving term, and it is embedded into the source term position in the differential relation expression according to the physical conservation principle.

[0023] In the process of updating the differential relation expression, a residual calculation layer of the hydraulic control equation is introduced. The residual calculation layer takes the current state as input, outputs the deviation value after the state quantity is substituted into the control equation, and superimposes the corresponding deviation value as a correction term onto the rate of change of state to ensure that the evolution process satisfies the physical conservation law.

[0024] As a further aspect of the present invention: the step of using the output of the continuous rainfall spatiotemporal function as an external driving term and embedding it into the source term position in the differential relation expression according to the principle of physical conservation specifically includes:

[0025] The rainfall intensity output value at the current time of the target river section is read from the spatiotemporal function of continuous rainfall. Combined with the catchment area and runoff coefficient of the corresponding river section, the net rain flux entering the river channel at the corresponding time is calculated.

[0026] Net rainfall flux is written into the right side of the differential relation in the form of source terms of the continuous equations in the Saint-Venant equations, in the form of lateral inflow per unit river length.

[0027] Based on the water surface width value corresponding to the current water level of the river section, physical conservation constraints are applied to the net rainfall flux written into the source term to ensure that the water level rise rate caused by lateral inflow matches the river channel's storage capacity.

[0028] As a further aspect of the present invention: S4 specifically includes:

[0029] Using hydrodynamic monitoring data from a multi-source dataset as the fitting benchmark, the deviation between the current dynamic mapping output value and the measured value at the monitoring point is calculated.

[0030] Physical constraint points are randomly sampled across the entire spatiotemporal range. The hydraulic state values ​​at the sampling points are substituted into the hydraulic control equations to calculate the continuity deviation and momentum deviation, thus forming the physical constraint residuals.

[0031] The monitoring point deviation and the physical constraint residual are weighted and combined to construct the comprehensive loss. The parameters of the dynamic mapping are then corrected in reverse based on the loss value. After each correction, the physical constraint residual is recalculated until the comprehensive loss meets the convergence condition.

[0032] As a further aspect of the present invention: the formation process of the physical constraint residual is as follows:

[0033] In the spatiotemporal computation domain, a gradient-based adaptive sampling strategy is adopted. Based on the hydraulic state change rate distribution output by dynamic mapping, the sampling density of physical constraint points is increased in the region where the change rate exceeds the threshold.

[0034] The spatiotemporal coordinates of the sampling points are input into a dynamic mapping to obtain the corresponding water level and flow rate values. The partial derivatives of the water level and flow rate with respect to time and space are calculated using automatic differentiation technology.

[0035] Substitute the water level, flow rate, and their partial derivatives into the continuity and momentum equations in the Saint-Venant equations, calculate the difference between the left and right sides of the equations, and use the corresponding difference as the physical constraint residual.

[0036] As a further aspect of the present invention: S5 specifically includes:

[0037] Based on historical hydrological data, the types and intensity distributions of random disturbance events involving weed clumps and floating objects in the upstream flow are extracted, and the probability density function of the disturbance events is constructed.

[0038] The probability density function is discretely sampled to generate multiple random disturbance scenarios, and each scenario is used as a boundary condition input to the optimized dynamic mapping to deduce the evolution trajectory of the hydraulic state of the river network under different disturbances.

[0039] Using the gate opening and closing sequence as the decision variable, the constraint that the highest water level in the river channel does not exceed the top elevation of the embankment under all random disturbance scenarios, and the objective function that the total energy consumption of the pumping station is minimized, the gate opening and closing sequence that satisfies the constraints is solved by the bibliometric optimization method.

[0040] As a further aspect of the present invention: the method of using the split-bar optimization method to solve for the gate opening and closing sequence that satisfies the constraints specifically includes:

[0041] Based on the hydraulic state evolution trajectory derived from multiple random disturbance scenarios, the highest water level sequence of the river corresponding to each random disturbance scenario is extracted, and an empirical distribution function with the gate opening and closing sequence as the independent variable is constructed.

[0042] Based on the empirical distribution function, the moment uncertainty set is used to describe the offset range between the real distribution and the empirical distribution of the random disturbance scenario, and the original constraints are transformed into the sub-Bruker constraint that must be satisfied for all possible distributions.

[0043] The fractional bar constraint is equivalently transformed into a fractional programming form that includes risk measurement parameters. The corresponding fractional programming is solved iteratively by the bisection search method. In each iteration, the gate opening and closing sequence is updated and the feasibility of all distributions in the moment uncertainty set is verified until the gate opening and closing sequence that satisfies the preset safety constraints and minimizes the total energy consumption of the pumping station is obtained.

[0044] A smart design system for water conservancy projects based on multi-source heterogeneous data fusion includes:

[0045] The data acquisition and preprocessing module is used to acquire meteorological radar rainfall forecast data and river network hydrodynamic monitoring data, and perform spatiotemporal coordinate system 1 and outlier cleaning to obtain an aligned multi-source dataset.

[0046] The continuous rainfall spatiotemporal function construction module is used to construct a continuous rainfall spatiotemporal function that can output the rainfall intensity at any spatiotemporal location by using implicit neural representation and discrete rainfall data from a multi-source dataset as supervision.

[0047] The physical enhancement neural differential equation construction module is used to construct physical enhancement neural differential equations. It embeds the spatiotemporal function of continuous rainfall as an external driving term into the neural differential equation to describe the dynamic mapping of the river network hydraulic state over time. The internal structure of the neural differential equation is ensured to be consistent with physical laws by introducing residual constraints of the hydraulic control equation.

[0048] The dynamic mapping optimization module uses hydrodynamic monitoring data from multiple sources as the fitting target and constructs physical constraints based on the hydrodynamic residuals at random spatiotemporal points of the dynamic mapping. It iteratively optimizes the parameters of the dynamic mapping to obtain the optimized dynamic mapping.

[0049] The module for solving the joint scheduling scheme of pump and gate groups uses the optimized dynamic mapping as an environment simulator. Combined with the probability distribution of random interference events in the upstream flow, it uses a robust optimization algorithm to solve the gate opening and closing sequence that meets the preset safety constraints, which serves as the design scheme for the joint scheduling of pump and gate groups.

[0050] The beneficial effects of this invention are:

[0051] (1) This invention reconstructs discrete radar rainfall data into a continuous spatiotemporal function through implicit neural representation, solving the time scale gap between discrete sampling and continuous hydrodynamic processes in traditional methods. At the same time, a gradient-based adaptive sampling strategy is adopted in the dynamic mapping parameter optimization process to automatically densify constraint points in areas with drastic changes in the physical field, enabling the neural network training to focus on key hydraulic feature areas, improving the prediction accuracy of the model under complex conditions such as flood rise and fall and gate opening and closing influence areas, and providing a more accurate hydraulic state inference basis for pump and gate group scheduling.

[0052] (2) This invention introduces residual constraints from the hydraulic control equations into the neural differential equations, ensuring that the evolution of the dynamic mapping strictly follows the physical conservation laws described by the Saint-Venant equations, thus avoiding physical inconsistencies that may occur in a purely data-driven model. Simultaneously, the sub-Bruker optimization method is used to handle random disturbances such as weed clumps in the upstream flow, transforming safety constraints into sub-Bruker constraints that must be satisfied for all possible distributions. The resulting gate opening and closing sequence ensures that the river level does not exceed limits under the most unfavorable disturbance scenarios, while also minimizing pump station energy consumption, achieving a synergistic optimization of the scheduling scheme's safety and economy. Attached Figure Description

[0053] The invention will now be further described with reference to the accompanying drawings.

[0054] Figure 1 This is a flowchart of the method of the present invention;

[0055] Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Please see Figure 1 As shown, this invention is an intelligent design method for water conservancy projects based on multi-source heterogeneous data fusion, comprising the following steps:

[0058] S1: Acquire meteorological radar rainfall forecast data and river network hydrodynamic monitoring data, and perform spatiotemporal coordinate system I and outlier cleaning processing to obtain an aligned multi-source dataset;

[0059] S2: Using implicit neural representations and discrete rainfall data from multi-source datasets as supervision, a continuous spatiotemporal function for rainfall that can output rainfall intensity at any spatiotemporal location is constructed.

[0060] S3: Construct a physically enhanced neural differential equation, embedding the spatiotemporal function of continuous rainfall as an external driving term into the neural differential equation to describe the dynamic mapping of the river network hydraulic state over time. The internal structure of the neural differential equation is ensured to be consistent with physical laws by introducing residual constraints from the hydraulic control equation.

[0061] S4: Using hydrodynamic monitoring data from multiple sources as the fitting target, and constructing physical constraints based on the hydrodynamic residuals at random spatiotemporal points of the dynamic mapping, the parameters of the dynamic mapping are iteratively optimized to obtain the optimized dynamic mapping;

[0062] S5: The optimized dynamic mapping is used as an environment simulator. Combined with the probability distribution of random interference events in the upstream flow, the gate opening and closing sequence that meets the preset safety constraints is solved by a robust optimization algorithm, which serves as the design scheme for joint scheduling of pump and gate groups.

[0063] In S1, meteorological radar rainfall forecast data and river network hydrodynamic monitoring data are acquired, and spatiotemporal coordinate system I and outlier cleaning are performed to obtain an aligned multi-source dataset, specifically including:

[0064] The meteorological radar rainfall forecast data is collected through a weather radar network deployed in the watershed where the project is located. This radar network scans every 6 minutes, emitting electromagnetic waves and receiving backscattered echoes from precipitation particles in the clouds. Gridded rainfall intensity distribution data is obtained through reflectivity factor inversion. After data acquisition, it is transmitted in real-time to the central data server via a dedicated communication network. The original data format is a raster file containing latitude and longitude coordinates and rainfall intensity values.

[0065] The river network hydrodynamic monitoring data is acquired by an automatic monitoring station network deployed at key river sections. Each monitoring station is equipped with a radar level gauge and an acoustic Doppler current profiler. The level gauge measures the distance from the water surface to the instrument by emitting radar waves to obtain the real-time water level, while the current profiler calculates the vertical velocity distribution by emitting and receiving the echo frequency shift of suspended particles in the water through an acoustic transducer. The above equipment collects data at a sampling interval of 5 minutes. The collected results are packaged by a remote terminal unit and reported to the data management center via a wireless network. The original data format is a structured record containing the station number, timestamp, water level value, and average cross-sectional velocity.

[0066] The acquired multi-source data underwent spatiotemporal coordinate system unification processing. Specifically, the latitude and longitude geographic coordinates used in the meteorological radar data were projected into an engineering independent coordinate system consistent with the river network monitoring data. All data timestamps were unified to Beijing time and resampled to a consistent 5-minute time interval. Subsequently, outlier cleaning was performed, removing abrupt changes exceeding the sensor's range, missing values ​​due to communication interruptions, and ground clutter interference from radar echoes. Short-term missing data were then filled in using linear interpolation, ultimately forming a spatiotemporally aligned multi-source dataset.

[0067] In S2, implicit neural representations are used with discrete rainfall data from a multi-source dataset as supervision to construct a continuous spatiotemporal rainfall function that can output rainfall intensity at any spatiotemporal location. Specifically, this includes:

[0068] First, a learnable continuous mapping function is constructed, taking spatiotemporal coordinates as input and rainfall intensity as output. This mapping function is implemented by embedding features into the input three-dimensional spatiotemporal coordinates (containing two spatial dimensions and one temporal dimension) using multi-resolution hashing. The construction process of multi-resolution hashing involves pre-setting multiple grid resolution levels from coarse to fine, for example, setting five grid resolution levels from 2^10 to 2^14. Each level corresponds to a hash table, and the size of the hash table is set to... Each input spatiotemporal coordinate is processed by calculating its grid vertex coordinates at each resolution level. The hash-encoded feature vector for each vertex is obtained through interpolation, and the feature vectors from all levels are concatenated to form the final embedded feature vector for that spatiotemporal coordinate. This embedded feature vector is then processed by a differentiable network consisting of multiple cascaded fully connected layers. Each fully connected layer performs a linear transformation sequentially, multiplying the input feature vector by the layer weight matrix and adding a bias vector. The result of the linear transformation is then nonlinearly mapped using a sinusoidal periodic activation function. The output of the sinusoidal periodic activation function is set to the sine of the product of the input value and a preset angular frequency coefficient, which gradually increases with the depth of the network layers.

[0069] Secondly, the discrete rainfall data from the multi-source dataset obtained after S1 processing is used as a supervision signal to construct a loss function to measure the deviation between the output value of the mapping function and the discrete observation value. The specific form of the loss function is as follows: calculate the squared difference between the predicted rainfall intensity value output by the mapping function and the measured rainfall intensity value at the spatiotemporal coordinates corresponding to each discrete observation data point, sum the squared differences over all observation points, and then divide by the total number of observation points to obtain the loss function value in the form of mean square error.

[0070] Then, based on the loss function, the parameters of the mapping function are adjusted using gradient descent until convergence, thereby obtaining a continuous rainfall spatiotemporal function capable of outputting rainfall intensity at arbitrary spatiotemporal locations. This parameter adjustment process specifically includes the following sub-steps:

[0071] The first sub-step involves calculating the gradients of all parameters of the learnable mapping function in the spatiotemporal function of continuous rainfall with respect to the loss function value, based on the current bias value output by the loss function, using the backpropagation algorithm. The backpropagation algorithm works as follows: starting from the loss function value, following the order from the output layer to the input layer, the chain rule is applied layer by layer to calculate the partial derivatives of the parameters at each layer, i.e., the parameter gradients.

[0072] The second sub-step involves adaptively densifying sampling points within local spatiotemporal regions where the loss value exceeds a preset threshold, based on the spatiotemporal distribution characteristics of the parameter gradient, and recalculating the parameter gradient contribution within these densified regions. Specifically, after calculating the parameter gradient in each iteration, the loss values ​​corresponding to each spatial location and time point are simultaneously statistically analyzed. Regions with loss values ​​exceeding 1.5 times the global average loss value are marked as key regions. Within these key regions, new spatiotemporal coordinate points are randomly selected at a sampling density twice the original grid density. These new sampling points are input into the current mapping function for forward computation to obtain predicted values, which are then compared with the original discrete observations within the region to calculate the local loss value. This is then backpropagated to obtain the parameter gradient increment corresponding to that region, and this increment is superimposed on the original parameter gradient.

[0073] The third sub-step involves updating the parameters using an adaptive learning rate associated with the current iteration step, taking into account the frequency response characteristics of the sinusoidal periodic activation function. The adaptive learning rate is set initially to 0.1%, and as the iteration step increases, the learning rate decreases by the absolute value of the sine wave (the number of iteration steps divided by 10,000) to match the response sensitivity of the sinusoidal activation function at different training stages. After updating the parameters, the loss function value is recalculated, and the magnitude of its decrease is assessed. If, for 30 consecutive iterations, the decrease in loss value is less than 0.01% of the initial loss value, the convergence condition is considered met, and the iteration process terminates. At this point, the parameters of the mapping function are fixed, resulting in the final spatiotemporal function of continuous rainfall.

[0074] In S3, a physically enhanced neural differential equation is constructed, embedding the spatiotemporal function of continuous rainfall as an external driving term into the neural differential equation to describe the dynamic mapping of the river network hydraulic state over time. The internal structure of the neural differential equation is ensured to be consistent with physical laws by introducing residual constraints from the hydraulic control equations, specifically including:

[0075] First, a differential relation expression is constructed with the river network hydraulic state as the dependent variable and the time coordinate as the independent variable. This differential relation expression is implemented using the framework of neural network constant differential equations, specifically a learnable function describing the mapping relationship between the rate of change of the state, the current state, and the external driving term. The river network hydraulic state consists of the water level and flow values ​​of all calculated cross-sections within the study area, forming a high-dimensional state vector. The external driving term is the rainfall intensity value provided by the continuous rainfall spatiotemporal function output in step S2. The learnable mapping structure is implemented by a multilayer neural network. The input of this network is the hydraulic state vector at the current moment and the external driving term at the current moment, and the output is the derivative of the hydraulic state vector with respect to time, i.e., the rate of change of the state. This neural network contains four hidden layers, each with 256 nodes. The activation function between the hidden layers is the hyperbolic tangent function, and the output layer uses a linear activation function. Through this network structure, the rate of change of the hydraulic state at any moment can be calculated from the current state and the rainfall driving term.

[0076] Secondly, the output of the spatiotemporal function of continuous rainfall is used as an external driving term, and it is embedded into the source term position in the above differential relation expression according to the principle of physical conservation. The specific implementation of this step is as follows:

[0077] The first sub-step involves reading the rainfall intensity output value at the target river segment location at the current moment from the continuous rainfall spatiotemporal function constructed in step S2. Specifically, the current time coordinates and the planar coordinates of the target river segment's center point are input into the continuous rainfall spatiotemporal function. After forward calculation, the function returns the instantaneous rainfall intensity value at that location, in millimeters per hour. Subsequently, calculations are performed using the corresponding catchment area and runoff coefficient values ​​for that river segment. The catchment area is determined based on the watershed extracted from the digital elevation model, and the runoff coefficient is obtained by looking up the underlying land use type in a table. When calculating the net storm flux entering the river channel, the rainfall intensity value is multiplied by the catchment area value, then multiplied by the runoff coefficient value, and the units are converted to obtain the net storm flux value in cubic meters per second.

[0078] The second sub-step involves incorporating the calculated net rainfall flux into the right-hand side of the differential equation, expressed as a lateral inflow per unit river length, in the form of a source term in the Saint-Venant continuity equations. In the Saint-Venant continuity equations, lateral inflow per unit river length is represented as the flow rate flowing into the river per unit length per unit time. For a calculated river segment of 500 meters, the total net rainfall flux is divided by the segment length to obtain the lateral inflow per unit river length for that segment. This value is then directly added as a source term to the right-hand side of the continuity equation for the corresponding segment, ensuring that the differential equation reflects the impact of rainfall runoff when updating water levels.

[0079] The third sub-step involves applying physical conservation constraints to the net rainfall flux written into the source term based on the water surface width value corresponding to the current water level of the river segment. The water surface width value is obtained from the river channel cross-section shape and water level value through a cross-section lookup table. The physical conservation constraint is verified by calculating the water level rise rate caused by lateral inflow per unit river length, i.e., dividing the lateral inflow value per unit river length by the current water surface width value to obtain the theoretical water level rise rate. This theoretical water level rise rate is then compared with the actual output value of the water level change rate in the differential relation expression. If the deviation exceeds 5%, the net rainfall flux is corrected to reduce this deviation and ensure that the water level change caused by lateral inflow matches the actual regulation capacity of the river channel.

[0080] Then, during the update process of the differential relation expression, a residual calculation layer of the hydraulic control equations is introduced. This residual calculation layer takes the hydraulic state vector at the current moment as input, and its internal calculation process is as follows: First, the water level and flow rate values ​​of each cross-section are extracted from the state vector, and the gradient of water level along the flow path and the gradient of flow rate along the flow path are calculated using the spatial difference method. Subsequently, the water level value, flow rate value, water level gradient, flow rate gradient, and time derivative are substituted into the continuity equation and momentum equation in the Saint-Venant equations, and the difference between the left-hand and right-hand terms of the equations is calculated. The formula for calculating the difference in the continuity equation is: the partial derivative of water level with respect to time plus the partial derivative of flow rate with respect to flow path, minus the lateral inflow value per unit river length. The formula for calculating the difference in the momentum equation is: the partial derivative of flow rate with respect to time plus the square of the cross-sectional flow rate divided by the partial derivative of the flow area with respect to flow path, plus the gravitational acceleration multiplied by the flow area multiplied by the partial derivative of water level with respect to flow path, and finally the friction gradient term is added. The calculated difference between the continuity equation and the momentum equation is used as the physical deviation value at that moment.

[0081] Finally, the aforementioned physical deviation values ​​are added as correction terms to the rate of change of state output by the differential relation expression. The correction method is as follows: the physical deviation value is multiplied by a preset correction coefficient, which is 0.1, and then the product is added to the original rate of change of state. Through this superposition correction, the rate of change of state of the differential relation expression can be dynamically adjusted during the evolution process, ensuring that the final output hydraulic state evolution trajectory strictly satisfies the physical conservation laws described by the Saint-Venant equations. After the above steps, the constructed physical-enhanced neural differential equation can be used to describe the dynamic mapping of the river network hydraulic state over time.

[0082] In S4, hydrodynamic monitoring data from a multi-source dataset is used as the fitting target, and physical constraints are constructed based on the hydraulic residuals of the dynamic mapping at random spatiotemporal points. The parameters of the dynamic mapping are iteratively optimized to obtain the optimized dynamic mapping, specifically including:

[0083] First, using the hydrodynamic monitoring data from the multi-source dataset obtained after S1 processing as the fitting benchmark, the deviation between the output value and the measured value of the current dynamic mapping at the monitoring point is calculated. Specifically, for each monitoring station, the measured water level and flow rate values ​​at each historical moment are obtained. The time coordinates of the corresponding moment and the planar coordinates of the station are input into the current dynamic mapping. After forward calculation, the dynamic mapping returns the predicted water level and flow rate values ​​at the station location. The squared difference between the predicted water level and the measured value is calculated for each moment, and then the squared difference between the predicted flow rate and the measured value is calculated. The sum of the squared differences for all moments is then divided by the product of the total number of monitoring points and the total number of moments to obtain the monitoring point deviation.

[0084] Secondly, physical constraint points are randomly sampled across the entire spatiotemporal domain. The hydraulic state values ​​at these sampling points are then substituted into the hydraulic control equations to calculate continuity and momentum deviations, thus forming physical constraint residuals. The specific formation process of these physical constraint residuals is as follows:

[0085] The first sub-step employs a gradient-based adaptive sampling strategy within the spatiotemporal computation domain. Based on the hydraulic state change rate distribution output by the dynamic mapping, the sampling density of physical constraint points is increased in regions where the change rate exceeds a threshold. Specifically, the entire computation time period is discretized into time grids with one-minute intervals, and the computational river network is discretized into spatial grids with fifty-meter intervals, forming an initial uniform set of candidate sampling points. Then, the current dynamic mapping is run once to obtain the partial derivatives of water level and flow rate with time at all candidate points. Regions where the absolute value of the partial derivative exceeds twice the average partial derivative of all candidate points are marked as high-change-rate regions. Within these high-change-rate regions, the density of the original candidate points is doubled, i.e., new midpoints are inserted between the original grid points as new physical constraint points. Finally, a fixed number of points are randomly selected from all candidate points as the physical constraint sampling points for this iteration.

[0086] The second sub-step involves inputting the spatiotemporal coordinates of the sampling points into a dynamic mapping process to obtain the corresponding water level and flow rate values. Then, an automatic differentiation technique is used to calculate the partial derivatives of water level and flow rate with respect to time and space. The automatic differentiation technique is implemented as follows: during the forward computation of the dynamic mapping, the derivative relationship of each computation node relative to the input is recorded simultaneously. When it is necessary to calculate the partial derivative of water level with respect to time at a certain sampling point, starting from the output node corresponding to that sampling point, the computation graph is traced backward to the input time coordinate node. The derivative product along the path is accumulated using the chain rule to obtain the partial derivative value of water level with respect to time. The same method is used to calculate the partial derivative values ​​of water level with respect to spatial coordinates, flow rate with respect to time, and flow rate with respect to spatial coordinates.

[0087] The third sub-step involves substituting the water level, flow rate, and their partial derivatives into the continuity and momentum equations in the Saint-Venant equations system, calculating the difference between the left and right sides of the equations, and using this difference as the physical constraint residual. The difference in the continuity equation is calculated as follows: add the partial derivative of water level with respect to time to the partial derivative of flow rate with respect to the flow path, and subtract the lateral inflow value per unit river length calculated in step S3. The result is the continuity equation residual. The difference in the momentum equation is calculated as follows: add the partial derivative of flow rate with respect to time to the partial derivative of the square of the flow rate divided by the cross-sectional area with respect to the flow path, add the gravitational acceleration multiplied by the cross-sectional area value, multiply by the water level's partial derivative with respect to the flow path, and finally add the preset friction slope value. The result is the momentum equation residual. Two physical constraint values, the continuity equation residual and the momentum equation residual, are calculated for each sampling point.

[0088] Then, the monitoring point deviation calculated in step S4 is combined with the physical constraint residuals to construct the comprehensive loss using a weighted combination. The weighted combination method is as follows: multiply the monitoring point deviation by a first weighting coefficient, which has a value of 0.6; multiply the mean square of the continuous equation residuals of all sampling points by a second weighting coefficient, which has a value of 0.2; multiply the mean square of the momentum equation residuals of all sampling points by a third weighting coefficient, which has a value of 0.2; and add the results of the above three products to obtain the comprehensive loss value.

[0089] Finally, the parameters of the dynamic mapping are corrected in reverse based on the comprehensive loss value, and the physical constraint residuals are recalculated after each correction until the comprehensive loss meets the convergence condition. The reverse correction is implemented using the Adam optimization algorithm, with an initial learning rate set to 0.01%. In each iteration, the gradient of all parameters of the dynamic mapping is calculated based on the comprehensive loss value, and the parameter values ​​are updated in the opposite direction of the gradient. After each parameter update, the calculation process of the physical constraint residuals in this step is re-executed. Based on the updated dynamic mapping, the continuity equation residuals and momentum equation residuals at all sampling points are recalculated, and the comprehensive loss is reconstructed. This process is repeated until the decrease in the comprehensive loss value is less than 0.01% of the initial comprehensive loss value in 30 consecutive iterations. At this point, the iteration is stopped, the current dynamic mapping parameters are fixed, and the optimized dynamic mapping is obtained.

[0090] In S5, the optimized dynamic mapping is used as an environment simulator. Combined with the probability distribution of random disturbance events in the upstream flow, a robust optimization algorithm is used to solve for the gate opening and closing sequence that satisfies preset safety constraints. This serves as the design scheme for the joint scheduling of the pump and gate group, specifically including:

[0091] First, the types and intensity distributions of random disturbance events in the upstream flow were extracted based on historical hydrological data. Historical hydrological records of the engineering area were collected, including video surveillance images, records of sudden water level changes, and manual inspection logs, to identify disturbance events such as weed clumps and floating debris carried downstream by floodwaters. For each recorded event, the time of occurrence, the location of the river segment, and the estimated accumulation volume or degree of blockage were recorded. All events were classified by type, and their occurrence frequency and intensity values ​​were statistically analyzed. The probability density function of each disturbance event was fitted using the kernel density estimation method. Taking weed clump disturbance as an example, its intensity was expressed as a percentage of the equivalent blockage cross-sectional area, ranging from 0% to 30%. The bandwidth parameter of the kernel density estimation was set according to empirical rules as 0.9 times the sample standard deviation divided by the fifth root of the sample size.

[0092] Secondly, the probability density function is discretely sampled to generate multiple random disturbance scenarios. Using the Latin hypercube sampling method, two thousand samples are extracted from the joint probability distribution of each disturbance event. Each sample corresponds to a complete disturbance scenario, specifically including the time point of the disturbance, the river segment location, and the intensity value. These two thousand scenarios are then used as boundary conditions to input the dynamic mapping obtained after optimization in step S4, i.e., the physically enhanced neural differential equation. For each scenario, corresponding initial and boundary conditions are set, and the dynamic mapping is used to perform forward integration to calculate the hydraulic state evolution of the entire river network over the next 72 hours, recording the water level and flow rate at each calculation section at each moment. The integration time step is set to 60 seconds, and the fourth-order Runge-Kutta method is used for numerical solution. Finally, two thousand sets of hydraulic state evolution trajectories are obtained, each containing the highest water level values ​​of all river segments within the prediction period.

[0093] Then, using the gate opening and closing sequence as the decision variable, the constraint that the highest water level in the river channel does not exceed the top elevation of the dike under all random disturbance scenarios, and the objective function of minimizing the total energy consumption of the pumping station, the distributed Bruker bar optimization method is used to solve for the gate opening and closing sequence that satisfies the constraints. The gate opening and closing sequence is represented as the opening value of each gate every 30 minutes within the next 24 hours, containing a total of 20 control periods. The total number of decision variables is the number of gates multiplied by 20. The total energy consumption of the pumping station is calculated based on the operating power and operating time of each pumping station. The power is related to the flow rate and head and is determined by the pumping station characteristic curve.

[0094] The method of using the bibliometric bar optimization to solve for the gate opening and closing sequence that satisfies the constraints specifically includes the following sub-steps:

[0095] The first sub-step involves extracting the highest water level in the river channel for each scenario, based on the hydraulic state evolution trajectories derived from the two thousand random disturbance scenarios. For a given candidate gate opening / closing sequence, the maximum water level of all river segments in each scenario is calculated using dynamic mapping. The highest water level in the k-th scenario is denoted as Hk, where k = 1, 2, ..., 2000. This yields a set of highest water level samples {H1, H2, ..., H2000}, which depends on the current gate opening / closing sequence. Based on these samples, an empirical distribution function is constructed with the gate opening / closing sequence as the independent variable. The empirical distribution function is defined as the cumulative probability at the sample points, i.e., for any threshold h, the empirical distribution value is the proportion of samples less than or equal to h. Simultaneously, the mean of the samples is calculated. and variance The mean is obtained by summing all samples and dividing by the total number of samples, and the variance is obtained by summing the squares of the differences between each sample and the mean and dividing by the total number of samples minus one.

[0096] The second sub-step, based on the empirical distribution function, uses the uncertainty set of moments to describe the offset between the true distribution and the empirical distribution of the random disturbance scenario. Since the distribution of actual disturbance events may deviate from historical samples, the uncertainty set of moments is introduced to cover all possible true distributions. The definition is as follows:

[0097] ;

[0098] in, Represents all probability distributions defined over the real number field. Let represent the set of all probability distributions defined over the real number field. For distribution The highest water level Expected value and These are the empirical mean and empirical variance calculated in the first sub-step, respectively. and The confidence coefficients are set to 0.1 and 1.2, respectively, to control the size of the uncertainty set. This set includes all distributions whose expected values ​​deviate from the empirical mean by no more than 0.1 times the standard deviation and whose variance does not exceed 1.2 times the empirical variance. The original constraint "the highest water level in the river channel does not exceed the levee crest elevation Hmax" is transformed into a fractional Brussels bar constraint that must be satisfied for all possible distributions within the moment uncertainty set. That is, the probability of exceeding the limit in the worst case is required to be 0, which is equivalent to requiring that the supremacy of the highest water level under all possible distributions is not greater than Hmax. Since the moment uncertainty set only constrains the first and second moments and cannot directly limit the supremacy, conditional value of risk is used as a risk measure, transforming the constraint into the conditional value of risk not exceeding Hmax in the worst case. Conditional value of risk is defined as the average loss exceeding a certain threshold, and its fractional Brussels bar form can be expressed as the following fractional programming:

[0099] ;

[0100] in, Represents conditional risk value. This represents the supremum, which is the maximum value among all elements in a set. This represents the infimum, which is the minimum value among all elements in a set. Let represent the set of all real numbers. The confidence level is set to 0.95. express The larger of 0 and 0. This expression transforms the conditional value of risk into a value about the auxiliary variable. For optimization problems, the upper bound of the internal expectation can be estimated using moment information, Chebyshev's inequality, or semidefinite programming.

[0101] The third sub-step transforms the aforementioned sub-Bruker constraint into a fractional programming form that includes risk measurement parameters, and iteratively solves this fractional programming problem using a bisection search method. The specific implementation process is as follows: First, set an upper and lower bound for the target value E0 of the total energy consumption of the pumping station. The lower bound is set to 0, and the upper bound is set to the total energy consumption of all pumping stations operating at full power for 24 hours. In each bisection search iteration, take the median value Emid between the current upper and lower bounds of energy consumption, and solve a feasibility subproblem: Does there exist a set of gate opening and closing sequences such that the total energy consumption of the pumping station does not exceed Emid and satisfies the sub-Bruker constraint? The feasibility subproblem is solved by solving a semidefinite programming problem. This semidefinite programming problem uses the gate sequence as a variable, with the sub-Bruker conditional risk value less than or equal to Hmax as a constraint, and the objective is to minimize a slack variable. If the subproblem is feasible, the upper bound of energy consumption is updated to Emid; otherwise, the lower bound is updated to Emid. Repeat the bisection search until the interval between the upper and lower bounds is less than a preset threshold of 0.01 MWh. In each feasibility verification, the empirical mean and variance need to be recalculated based on the current gate timing, and the moment uncertainty set needs to be updated before solving a semidefinite programming problem. The final gate opening and closing sequence that satisfies all constraints and minimizes the total energy consumption of the pumping station is obtained as the output design scheme for the joint scheduling of the pumping station group. This scheme can cope with the uncertainty of random disturbances such as weeds in the upstream flow, ensuring that the river level does not exceed the limit under various possible disturbances.

[0102] Please see Figure 2 As shown, the intelligent design system for water conservancy projects based on multi-source heterogeneous data fusion includes:

[0103] The data acquisition and preprocessing module is used to acquire meteorological radar rainfall forecast data and river network hydrodynamic monitoring data, and perform spatiotemporal coordinate system 1 and outlier cleaning to obtain an aligned multi-source dataset.

[0104] The continuous rainfall spatiotemporal function construction module is used to construct a continuous rainfall spatiotemporal function that can output the rainfall intensity at any spatiotemporal location by using implicit neural representation and discrete rainfall data from a multi-source dataset as supervision.

[0105] The physical enhancement neural differential equation construction module is used to construct physical enhancement neural differential equations. It embeds the spatiotemporal function of continuous rainfall as an external driving term into the neural differential equation to describe the dynamic mapping of the river network hydraulic state over time. The internal structure of the neural differential equation is ensured to be consistent with physical laws by introducing residual constraints of the hydraulic control equation.

[0106] The dynamic mapping optimization module uses hydrodynamic monitoring data from multiple sources as the fitting target and constructs physical constraints based on the hydrodynamic residuals at random spatiotemporal points of the dynamic mapping. It iteratively optimizes the parameters of the dynamic mapping to obtain the optimized dynamic mapping.

[0107] The module for solving the joint scheduling scheme of pump and gate groups uses the optimized dynamic mapping as an environment simulator. Combined with the probability distribution of random interference events in the upstream flow, it uses a robust optimization algorithm to solve the gate opening and closing sequence that meets the preset safety constraints, which serves as the design scheme for the joint scheduling of pump and gate groups.

[0108] The working principle of this invention is as follows: First, meteorological radar rainfall forecast data and river network hydrodynamic monitoring data are acquired, and spatiotemporal coordinate system I and outlier cleaning are performed to obtain an aligned multi-source dataset. Then, implicit neural representation is used with discrete rainfall data as supervision to construct a continuous rainfall spatiotemporal function that can output rainfall intensity at any spatiotemporal location. Next, a physically enhanced neural differential equation is constructed, embedding the continuous rainfall spatiotemporal function as an external driving term, and ensuring the consistency between the dynamic mapping and physical laws by introducing residual constraints from the hydraulic control equation. Then, hydrodynamic monitoring data is used as the fitting target, and physical constraints are constructed based on the hydraulic residuals of the dynamic mapping at random spatiotemporal points to iteratively optimize the parameters of the dynamic mapping. Finally, the optimized dynamic mapping is used as an environmental simulator. Combined with the probability distribution of random disturbance events in the upstream flow, the gate opening and closing sequence that meets the preset safety constraints is solved by the Bruker optimization algorithm, which serves as the design scheme for the joint scheduling of pump and gate groups. This invention uses implicit neural representation to make discrete rainfall data continuous, constructs physically enhanced neural differential equations to achieve deep integration of data and physical laws, and uses blobs optimization to handle random disturbance uncertainties, thereby obtaining a pump and gate group scheduling scheme that balances safety and economy.

[0109] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for intelligent design of water conservancy projects based on multi-source heterogeneous data fusion, characterized in that, Includes the following steps: S1: Acquire meteorological radar rainfall forecast data and river network hydrodynamic monitoring data, and perform spatiotemporal coordinate system I and outlier cleaning processing to obtain an aligned multi-source dataset; S2: Using implicit neural representations and discrete rainfall data from multi-source datasets as supervision, a continuous spatiotemporal function for rainfall that can output rainfall intensity at any spatiotemporal location is constructed. S3: Construct a physically enhanced neural differential equation; wherein, a neural differential equation is constructed with the hydraulic state of the river network as the dependent variable and the time coordinate as the independent variable; the hydraulic state of the river network consists of the water level and flow rate values ​​of all calculation sections within the study area; the neural differential equation is implemented by a multilayer neural network, the input of which is the hydraulic state vector at the current moment and the external driving term at the current moment, and the output is the derivative of the hydraulic state vector with respect to time, i.e., the rate of change of the hydraulic state; the external driving term is the rainfall intensity value provided by the continuous rainfall spatiotemporal function output in step S2; During the update process of the multilayer neural network, a residual calculation layer of the hydraulic control equation is introduced. The residual calculation layer takes the hydraulic state vector at the current moment as input, calculates the physical deviation value after substituting the hydraulic state vector into the hydraulic control equation, and adds the physical deviation value at the current moment as a correction term to the hydraulic state change rate to ensure that the evolution trajectory of the final output hydraulic state satisfies the physical conservation law. S4: Using the hydrodynamic monitoring data in the multi-source dataset obtained after step S1 as the fitting target, and constructing physical constraints based on the hydraulic residuals of the multilayer neural network at random spatiotemporal points, the parameters of the multilayer neural network are iteratively optimized to obtain the optimized multilayer neural network; the hydraulic residuals are the difference between the continuity equation and the momentum equation in the hydraulic control equations. S5: Using the optimized multilayer neural network as an environmental simulator, and combining it with the probability distribution of random disturbance events in the upstream flow, the gate opening and closing sequence that satisfies the preset safety constraints is solved through the Bruker optimization algorithm. This serves as the design scheme for the joint scheduling of the pump and gate group, specifically including: Based on historical hydrological data, the types and intensity distributions of random disturbance events in the upstream flow are extracted, and the probability density function of each disturbance event is constructed. The probability density function is discretely sampled to generate multiple random disturbance scenarios, and each random disturbance scenario is used as a boundary condition input to the optimized multilayer neural network to deduce the evolution trajectory of the hydraulic state of the river network under different disturbances. Using the gate opening and closing sequence as the decision variable, the safety constraint that the highest water level in the river channel does not exceed the top elevation of the embankment under all random disturbance scenarios, and the objective function of minimizing the total energy consumption of the pumping station, the partial Bruker bar optimization method is used to solve the gate opening and closing sequence. Specifically, this includes: extracting the highest water level in the river channel corresponding to each random disturbance scenario based on the evolution trajectory of the hydraulic state derived from multiple random disturbance scenarios, summarizing the highest water level sequence in the river channel, and constructing an empirical distribution function with the gate opening and closing sequence as the independent variable based on the highest water level sequence in the river channel; using the empirical distribution function as the basis, using the moment uncertainty set to describe the offset range between the real distribution and the empirical distribution of the random disturbance scenarios, and transforming the original safety constraint into a partial Bruker bar constraint that must be satisfied for all possible distributions; transforming the partial Bruker bar constraint into a fractional programming form containing risk measurement parameters, and iteratively solving the corresponding fractional programming using the bisection search method, updating the gate opening and closing sequence in each iteration and verifying its feasibility for all distributions in the moment uncertainty set, until the gate opening and closing sequence that satisfies the preset safety constraints and minimizes the total energy consumption of the pumping station is obtained.

2. The intelligent design method for water conservancy projects based on multi-source heterogeneous data fusion according to claim 1, characterized in that, S2 specifically includes: A learnable continuous mapping function is constructed with spatiotemporal coordinates as input and rainfall intensity as output. The learnable continuous mapping function uses multi-resolution hash coding to embed features of the input coordinates and processes the embedded features through multiple cascaded differentiable transformation layers. Each transformation layer contains a linear transformation and a sinusoidal periodic activation function. Discrete rainfall data from a multi-source dataset are used as supervision to construct a loss function to measure the deviation between the output of the learnable continuous mapping function and the discrete observations; Based on the loss function, the parameters of the learnable continuous mapping function are adjusted using the gradient descent method until convergence, resulting in a continuous rainfall spatiotemporal function that can output rainfall intensity at any spatiotemporal location.

3. The intelligent design method for water conservancy projects based on multi-source heterogeneous data fusion according to claim 2, characterized in that, The adjustment of the mapping function parameters based on the loss function and using gradient descent specifically includes: Based on the bias value output by the loss function, the parameter gradient of the learnable mapping function in the spatiotemporal function of continuous rainfall is calculated; Based on the spatiotemporal distribution characteristics of the parameter gradient, the sampling points are adaptively densified in local spatiotemporal regions where the loss value exceeds a preset threshold, and the parameter gradient contribution of the corresponding region is recalculated. By leveraging the frequency response characteristics of the sinusoidal periodic activation function, an adaptive learning rate associated with the current iteration step is used to update the parameters. After the parameters are updated, the magnitude of the decrease in the loss function is judged. If the magnitude of the decrease in multiple consecutive iterations is less than the preset value, the iteration is terminated.

4. The intelligent design method for water conservancy projects based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The S3 further includes: The external driving term is input into the neural differential equation according to the principle of physical conservation, specifically including: The rainfall intensity value at the current time of the target river segment is read from the spatiotemporal function of continuous rainfall. Combined with the corresponding catchment area value and runoff coefficient value of the river segment, the net rain flux entering the river channel at the current time is calculated. Divide the total net rainfall flux by the river length to obtain the lateral inflow per unit river length for that river segment; Divide the lateral inflow value per unit river length by the water surface width value corresponding to the current water level of the river section to obtain the theoretical water level rise rate; The theoretical water level rise rate is compared with the actual output value of the water level change rate in the neural differential equation. If the deviation between the two water level change rates exceeds the deviation threshold, the net rainwater flux is corrected to reduce the deviation of the water level change rate and ensure that the water level change caused by lateral inflow matches the actual regulation capacity of the river.

5. The intelligent design method for water conservancy projects based on multi-source heterogeneous data fusion according to claim 4, characterized in that, The step of calculating the physical deviation value after substituting the hydraulic state vector into the hydraulic control equation, and then adding the physical deviation value as a correction term to the hydraulic state change rate, further includes: The hydraulic governing equations are the Saint-Venant equations. The water level and flow rate values ​​of each cross section are extracted from the hydraulic state vector, and the gradient of the water level and the gradient of the flow rate along the process are calculated using the spatial difference method. Then, the water level value, flow rate value, water level gradient, flow rate gradient, and time derivative are substituted into the continuity equation and momentum equation in the Saint-Venant equations to calculate the difference between the continuity equation and the momentum equation. The calculated difference between the continuity equation and the momentum equation is used as the physical deviation value at that moment; finally, the above physical deviation value is added as a correction term to the hydraulic state change rate.

6. The intelligent design method for water conservancy projects based on multi-source heterogeneous data fusion according to claim 1, characterized in that, S4 specifically includes: Using hydrodynamic monitoring data from a multi-source dataset as a fitting benchmark, the deviation between the output value and the measured value of the multilayer neural network at the monitoring point is calculated. Physical constraint points are randomly sampled within the entire spatiotemporal range. The hydraulic state values ​​of the river network at the sampling points are substituted into the hydraulic control equations to calculate the continuity deviation and momentum deviation, forming hydraulic residuals, which serve as physical constraint residuals. The deviation between the output value and the measured value at the monitoring point is weighted and combined with the physical constraint residual to construct a comprehensive loss value. The parameters of the multilayer neural network are then reverse-corrected based on the comprehensive loss value, and the physical constraint residual is recalculated after each correction until the comprehensive loss value meets the convergence condition.

7. The intelligent design method for water conservancy projects based on multi-source heterogeneous data fusion according to claim 6, characterized in that, The formation process of the physical constraint residual is as follows: In the spatiotemporal computation domain, a gradient-based adaptive sampling strategy is adopted. Based on the distribution of the rate of change of the hydraulic state of the river network output by the multilayer neural network, the sampling density of physical constraint points is increased in the region where the rate of change of the hydraulic state exceeds the rate of change threshold. Then, the spatiotemporal coordinates of the sampling points are input into the multilayer neural network to obtain the corresponding water level and flow rate values, and the partial derivatives of water level and flow rate with respect to time and space are calculated using automatic differentiation technology; Substitute the water level, flow rate, and their partial derivatives into the continuity and momentum equations in the Saint-Venant equations, calculate the difference between the left and right sides of the equations, and use the corresponding difference as the physical constraint residual.

8. A smart design system for water conservancy projects based on multi-source heterogeneous data fusion, characterized in that: The method for executing the intelligent design method for water conservancy projects based on multi-source heterogeneous data fusion as described in any one of claims 1-7 includes: The data acquisition and preprocessing module is used to acquire meteorological radar rainfall forecast data and river network hydrodynamic monitoring data, and perform spatiotemporal coordinate system 1 and outlier cleaning to obtain an aligned multi-source dataset. The continuous rainfall spatiotemporal function construction module is used to construct a continuous rainfall spatiotemporal function that can output the rainfall intensity at any spatiotemporal location by using implicit neural representation and discrete rainfall data from a multi-source dataset as supervision. A physically enhanced neural differential equation construction module is used to construct physically enhanced neural differential equations, wherein a neural differential equation is constructed with the hydraulic state of the river network as the dependent variable and the time coordinate as the independent variable; the hydraulic state of the river network consists of the water level and flow rate values ​​of all calculation sections within the study area; the neural differential equation is implemented by a multilayer neural network, the input of which is the hydraulic state vector at the current moment and the external driving term at the current moment, and the output is the derivative of the hydraulic state vector with respect to time, i.e., the rate of change of the hydraulic state; the external driving term is the rainfall intensity value provided by the continuous rainfall spatiotemporal function output in step S2; During the update process of the multilayer neural network, a residual calculation layer of the hydraulic control equation is introduced. The residual calculation layer takes the hydraulic state vector at the current moment as input, calculates the physical deviation value after substituting the hydraulic state vector into the hydraulic control equation, and adds the physical deviation value at the current moment as a correction term to the hydraulic state change rate to ensure that the evolution trajectory of the final output hydraulic state satisfies the physical conservation law. The multi-layer neural network optimization module uses hydrodynamic monitoring data from a multi-source dataset as the fitting target and constructs physical constraints based on the hydraulic residuals of the multi-layer neural network at random spatiotemporal points. Iterative optimization of the parameters of the multi-layer neural network is then performed to obtain the optimized multi-layer neural network. The hydraulic residuals are the difference between the continuity equation and the momentum equation in the hydraulic control equations. The module for solving the joint scheduling scheme of the pump and gate group uses an optimized multilayer neural network as an environmental simulator. Combined with the probability distribution of random disturbance events in the upstream flow, it uses a bibliometric optimization algorithm to solve for the gate opening and closing sequence that satisfies preset safety constraints. This serves as the design scheme for the joint scheduling of the pump and gate group, specifically including: Based on historical hydrological data, the types and intensity distributions of random disturbance events in the upstream flow are extracted, and the probability density function of each disturbance event is constructed. Multiple random interference scenarios are generated by discretely sampling the probability density function. Each random disturbance scenario is used as a boundary condition input to the optimized multilayer neural network to deduce the evolution trajectory of the hydraulic state of the river network under different disturbances. Using the gate opening and closing sequence as the decision variable, the safety constraint that the highest water level in the river channel does not exceed the top elevation of the embankment under all random disturbance scenarios, and the objective function of minimizing the total energy consumption of the pumping station, the partial Bruker bar optimization method is used to solve the gate opening and closing sequence. Specifically, this includes: extracting the highest water level in the river channel corresponding to each random disturbance scenario based on the evolution trajectory of the hydraulic state derived from multiple random disturbance scenarios, summarizing the highest water level sequence in the river channel, and constructing an empirical distribution function with the gate opening and closing sequence as the independent variable based on the highest water level sequence in the river channel; using the empirical distribution function as the basis, using the moment uncertainty set to describe the offset range between the real distribution and the empirical distribution of the random disturbance scenarios, and transforming the original safety constraint into a partial Bruker bar constraint that must be satisfied for all possible distributions; transforming the partial Bruker bar constraint into a fractional programming form containing risk measurement parameters, and iteratively solving the corresponding fractional programming using the bisection search method, updating the gate opening and closing sequence in each iteration and verifying its feasibility for all distributions in the moment uncertainty set, until the gate opening and closing sequence that satisfies the preset safety constraints and minimizes the total energy consumption of the pumping station is obtained.

Images