A water resource scheduling intelligent management and control and data visualization management system

Abstract

The application discloses a water resource scheduling intelligent management and control and data visualization management system, relates to the field of water conservancy informatization technology, and comprises a cascade topological coupling graph construction module, a graph attention hydraulic deduction module, a hydraulic coupling constraint generation module, a rolling time domain scheduling module and a twin pre-performance correction module; the cascade topological coupling graph construction module is configured to construct each cascade reservoir object in a cascade reservoir group and a river channel interval object connecting adjacent cascade reservoirs into a cascade topological coupling graph; through construction of the cascade topological coupling graph, the upstream and downstream relationships of the cascade reservoir group in the hydraulic system are converted from implicit physical connection into explicit graph structure representation, so that the hydraulic propagation time lag and evolution decay coefficient become quantifiable properties of graph edges, and a structural basis is provided for subsequent hydraulic collaborative deduction.

Description

Technical Field

[0001] This invention belongs to the field of water conservancy information technology, specifically a smart management and data visualization system for water resource scheduling. Background Technology

[0002] A cascade reservoir system refers to a water conservancy project system consisting of multiple reservoirs connected in a series, arranged upstream and downstream, within the same river basin. These reservoirs are interconnected via river channels, forming a unified water conservancy system. The joint operation of flood control and power generation within the cascade reservoir system is a core aspect of river basin water resources management, aiming to maximize power generation benefits while ensuring downstream flood control safety.

[0003] In the actual operation and scheduling of cascade reservoir groups, the following key technical aspects exist: First, inflow prediction, which involves forecasting the inflow to each reservoir in future periods based on upstream water conditions and rainfall in the river basin; second, hydraulic evolution calculation, which involves simulating the attenuation and time lag effects of the outflow during its propagation through the river channel; third, scheduling scheme development, which involves formulating outflow plans for each reservoir based on hydraulic prediction results and engineering constraints; and fourth, scheme evaluation and verification, which involves conducting effect simulations and comprehensive evaluations of candidate scheduling schemes.

[0004] In existing technologies, the aforementioned technical steps are typically completed by separate, independent technical systems. Inflow forecasting often employs hydrological forecasting models such as the unit hydrograph method or the Xin'anjiang model; hydraulic evolution calculations often use the Muskingum channel evolution method; scheduling scheme development relies heavily on the experience and judgment of scheduling personnel or employs multi-objective optimization algorithms; and scheme evaluation is mostly presented in the form of numerical reports, lacking intuitive visualization methods.

[0005] Existing technologies have the following three shortcomings in the joint dispatching of flood control and power generation in cascade reservoir groups:

[0006] First, the hydraulic coupling relationship is not adequately represented. Existing technologies typically treat each cascade reservoir as an independent scheduling object, using its own independent prediction and optimization models. This ignores the hydraulic upstream-downstream coupling relationship formed between the cascade reservoirs through river channels. In reality, the discharge flow of the upstream reservoir directly affects the inflow flow of the downstream reservoir, and there is a definite hydraulic propagation time lag and flow attenuation relationship between the two. Existing technologies fail to effectively utilize this hydraulic coupling relationship for the coordinated scheduling of cascade reservoirs, resulting in hydraulic inconsistencies in the scheduling scheme and systematic deviations between the simulation results and actual operation. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a smart management and data visualization management system for water resource scheduling, aiming to solve the following technical problems: First, by constructing a cascade topological coupling graph, the hydraulic upstream and downstream relationships of a cascade reservoir group are explicitly encoded into a graph structure, enabling hydraulic coupling and collaborative simulation among cascade reservoirs; Second, through a pre-simulation deviation correction module, the digital twin pre-simulation results are compared with historical measured data to correct the evolution attenuation coefficient in the edge attributes of the graph structure online, so that the model parameters adapt to changes in actual operating conditions.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] Specifically, a smart management and data visualization system for water resource scheduling is proposed, including a cascade topology coupling graph construction module, a graph attention hydraulic inference module, a hydraulic coupling constraint generation module, a rolling time-domain scheduling module, and a twin pre-simulation correction module;

[0010] The cascade topology coupling graph construction module is configured to construct a cascade topology coupling graph by combining each cascade reservoir object in the cascade reservoir group with the river channel objects connecting adjacent cascade reservoirs. The node attributes of the cascade topology coupling graph include the upstream water level and outflow value of each cascade reservoir object at the current scheduling time. The edge attributes of the cascade topology coupling graph include the hydraulic propagation time delay parameter and evolution attenuation coefficient of each river channel object. The hydraulic propagation time delay parameter is the number of time periods required for the outflow from the upstream cascade reservoir object to propagate to the downstream cascade reservoir object. The evolution attenuation coefficient is the attenuation ratio of the outflow during the propagation process in the river channel.

[0011] The graph attention hydraulic extrapolation module is configured to perform graph attention message passing operations on the cascade topology coupling graph. The graph attention message passing operation is as follows: for each edge in the cascade topology coupling graph, the water level gradient value is calculated by the upstream graph node connected by the edge and the downstream graph node. The attention weight value of the edge is calculated by the water level gradient value, the hydraulic propagation time delay parameter of the edge, and the evolution attenuation coefficient. The node features of the neighboring graph nodes are weighted and aggregated using the attention weight value as the aggregation coefficient to obtain an aggregated feature vector. The aggregated feature vector is input into a gated recurrent network for time-series encoding, and the extrapolated upstream water level sequence and the extrapolated reservoir flow sequence of each graph node in the future scheduling time domain are output. The time domain length of the future scheduling time domain is determined according to the flood forecast lead time.

[0012] The hydraulic coupling constraint generation module is configured to calculate the hydraulic coupling tightness value between each cascade reservoir object based on the edge attributes of the cascade topological coupling graph and the inferred dam front water level sequence, and to compile the hydraulic coupling tightness value with the upper limit of the reservoir capacity and the upper limit of the discharge capacity of each cascade reservoir object into a hydraulic coupling constraint matrix. The hydraulic coupling tightness value is a quantification of the correlation between the change in the outflow of the upstream cascade reservoir object and the change in the dam front water level of the downstream cascade reservoir object.

[0013] The rolling time-domain scheduling module is configured to decode the hydraulic coupling constraint matrix into the time-limited reservoir capacity and time-limited outflow of each cascade reservoir object within each rolling scheduling cycle. Using the time-limited reservoir capacity and time-limited outflow as search boundaries, and taking the flood control water level requirements, upstream water level change rate requirements, and downstream safe discharge requirements of each cascade reservoir object as hard constraints, and taking the maximum power generation requirement as the optimization direction, a set of candidate scheduling schemes is generated. The optimization direction of the maximum power generation requirement is to maximize the cumulative power generation of all cascade reservoir objects in the future scheduling time domain. Each candidate scheduling scheme in the set of candidate scheduling schemes includes the downstream discharge flow sequence of each cascade reservoir object in the future scheduling time domain.

[0014] The twin pre-simulation correction module is configured to input the outflow sequence of each candidate scheduling scheme into the digital twin of the cascade reservoir group to perform hydraulic state extrapolation and obtain the pre-simulation dam front water level sequence, calculate the time-period deviation vector between the pre-simulation dam front water level sequence and the historical measured water level sequence, correct the evolution attenuation coefficient of each river section object in the cascade topology coupling graph according to the time-period deviation vector, and send the corrected evolution attenuation coefficient back to the cascade topology coupling graph construction module to update the edge attributes. The digital twin of the cascade reservoir group is a digital three-dimensional simulation model constructed based on the topographic data and hydraulic structure parameters of each cascade reservoir object.

[0015] The technical solutions provided by the embodiments of this disclosure have at least the following beneficial effects:

[0016] By constructing a cascade topology coupling graph, the hydraulic upstream and downstream relationships of the cascade reservoir group are transformed from implicit physical connections to explicit graph structure representations. This makes hydraulic propagation delay and evolution attenuation coefficients quantifiable attributes of the graph edges, providing a structural foundation for subsequent hydraulic collaborative simulations. Through the graph attention hydraulic simulation module, hydraulic message passing is performed on the topology using the graph attention mechanism. Physical constraints such as water level gradient, hydraulic delay, and flow attenuation are integrated into the calculation of attention weights, realizing hydraulic state simulation based on physical perception. Through the twin simulation correction module, the deviation between the digital twin simulation results and historical measured data is compared. The evolution attenuation coefficient is corrected online using the least squares method with physical constraints, enabling the model parameters to adapt to the time-varying characteristics of river hydraulic properties and improving the stability of model predictions during long-term scheduling.

[0017] The adjacency feature splicer concatenates the upstream and downstream graph nodes' upstream water level and outflow values ​​into a four-dimensional adjacency feature vector, preserving complete hydraulic state information for the upstream and downstream. A physical perception attention calculator merges the negative exponential function value of hydraulic propagation delay with the linear function value of evolution decay coefficient into a physical bias term. This physical bias term is then multiplied by the linear projection inner product of the adjacency feature vector to obtain the attention weight value. This ensures that the attention weight is simultaneously modulated by both physical constraints and feature associations, avoiding the problem of purely data-driven attention mechanisms neglecting hydraulic physical laws. A message aggregator performs three-layer graph structure information transmission operations and normalizes each layer, enabling hydraulic information to propagate and diffuse layer by layer along the cascade topology, achieving collaborative hydraulic state extrapolation for the entire cascade reservoir system. Attached Figure Description

[0018] Figure 1 This is a system flowchart of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This application provides a smart management and data visualization management system for water resource scheduling, including a cascade topology coupling graph construction module, a graph attention hydraulic inference module, a hydraulic coupling constraint generation module, a rolling time domain scheduling module, and a twin pre-simulation correction module;

[0021] The cascade topology coupling graph construction module is configured to construct a cascade topology coupling graph by combining each cascade reservoir object in the cascade reservoir group with the river channel objects connecting adjacent cascade reservoirs. The node attributes of the cascade topology coupling graph include the upstream water level and outflow value of each cascade reservoir object at the current scheduling time. The edge attributes of the cascade topology coupling graph include the hydraulic propagation time delay parameter and evolution attenuation coefficient of each river channel object. The hydraulic propagation time delay parameter is the number of time periods required for the outflow from the upstream cascade reservoir object to propagate to the downstream cascade reservoir object. The evolution attenuation coefficient is the attenuation ratio of the outflow during the propagation process in the river channel.

[0022] The graph attention hydraulic extrapolation module is configured to perform graph attention message passing operations on the cascade topology coupling graph. The graph attention message passing operation is as follows: for each edge in the cascade topology coupling graph, the water level gradient value is calculated by the upstream graph node connected by the edge and the downstream graph node. The attention weight value of the edge is calculated by the water level gradient value, the hydraulic propagation time delay parameter of the edge, and the evolution attenuation coefficient. The node features of the neighboring graph nodes are weighted and aggregated using the attention weight value as the aggregation coefficient to obtain an aggregated feature vector. The aggregated feature vector is input into a gated recurrent network for time-series encoding, and the extrapolated upstream water level sequence and the extrapolated reservoir flow sequence of each graph node in the future scheduling time domain are output. The time domain length of the future scheduling time domain is determined according to the flood forecast lead time.

[0023] The hydraulic coupling constraint generation module is configured to calculate the hydraulic coupling tightness value between each cascade reservoir object based on the edge attributes of the cascade topological coupling graph and the inferred dam front water level sequence, and to compile the hydraulic coupling tightness value with the upper limit of the reservoir capacity and the upper limit of the discharge capacity of each cascade reservoir object into a hydraulic coupling constraint matrix. The hydraulic coupling tightness value is a quantification of the correlation between the change in the outflow of the upstream cascade reservoir object and the change in the dam front water level of the downstream cascade reservoir object.

[0024] The rolling time-domain scheduling module is configured to decode the hydraulic coupling constraint matrix into the time-limited reservoir capacity and time-limited outflow of each cascade reservoir object within each rolling scheduling cycle. Using the time-limited reservoir capacity and time-limited outflow as search boundaries, and taking the flood control water level requirements, upstream water level change rate requirements, and downstream safe discharge requirements of each cascade reservoir object as hard constraints, and taking the maximum power generation requirement as the optimization direction, a set of candidate scheduling schemes is generated. The optimization direction of the maximum power generation requirement is to maximize the cumulative power generation of all cascade reservoir objects in the future scheduling time domain. Each candidate scheduling scheme in the set of candidate scheduling schemes includes the downstream discharge flow sequence of each cascade reservoir object in the future scheduling time domain.

[0025] The twin pre-simulation correction module is configured to input the outflow sequence of each candidate scheduling scheme into the digital twin of the cascade reservoir group to perform hydraulic state extrapolation and obtain the pre-simulation dam front water level sequence, calculate the time-period deviation vector between the pre-simulation dam front water level sequence and the historical measured water level sequence, correct the evolution attenuation coefficient of each river section object in the cascade topology coupling graph according to the time-period deviation vector, and send the corrected evolution attenuation coefficient back to the cascade topology coupling graph construction module to update the edge attributes. The digital twin of the cascade reservoir group is a digital three-dimensional simulation model constructed based on the topographic data and hydraulic structure parameters of each cascade reservoir object.

[0026] It should be noted that the construction process of the cascade topology coupling graph is as follows: For each cascade reservoir object in the cascade reservoir group, a graph node is created. The upstream water level value in the node attribute of the graph node is obtained by reading the latest monitoring value in the water level monitoring data sequence of the cascade reservoir object, and the outflow value is obtained by reading the latest monitoring value in the flow monitoring data sequence of the cascade reservoir object. For each river channel section object connecting adjacent cascade reservoirs, a directed edge is created in the cascade topology coupling graph. The direction of the directed edge is from the upstream cascade reservoir object to the downstream cascade reservoir object. The hydraulic propagation time delay parameter in the edge attribute of the directed edge is obtained by dividing the river length of the river channel object by the average flow velocity and then rounding it. The average flow velocity is obtained by statistically analyzing the historical flow monitoring data and cross-sectional area of ​​the river channel object. The evolution attenuation coefficient is obtained by statistically analyzing the ratio of the peak value of the historical outflow sequence to the peak value of the historical inflow sequence of the river channel object.

[0027] The calculation process of the attention weight value is as follows: For each edge, calculate the absolute value of the difference between the upstream graph node's water level value and the downstream graph node's water level value connected by the edge to obtain the water level gradient value; calculate the product of the negative exponential function value of the hydraulic propagation time delay parameter and the linear function value of the evolution attenuation coefficient to obtain the physical bias term, wherein the negative exponential function value is the negative hydraulic propagation time delay parameter of the natural constant e divided by the arithmetic mean of the hydraulic propagation time delay parameters of all river channel objects in the cascade reservoir group to the power of ...

[0028] The calculation process for the hydraulic coupling tightness value is as follows: For each pair of graph nodes with a hydraulic upstream-downstream relationship, the outflow values ​​of each adjacent time period in the deduced reservoir flow sequence of the upstream graph node are subtracted to obtain the upstream outflow change sequence; the deduced dam front water level sequence of the downstream graph node is shifted according to the number of time periods corresponding to the hydraulic propagation time delay parameters of the corresponding river section objects between the pair of graph nodes to obtain the shifted water level sequence; the water level values ​​of each adjacent time period in the shifted water level sequence are subtracted to obtain the downstream water level change sequence; the upstream outflow change sequence and the downstream water level change sequence are cross-correlated, and the peak value of the cross-correlation function is taken as the hydraulic coupling tightness value between the pair of graph nodes;

[0029] The correction process of the twin pre-simulation correction module is as follows: The discharge flow sequence of each candidate scheduling scheme is input into the digital twin of the cascade reservoir group. Based on the digital elevation model raster data, reservoir bank boundary vector data, and discharge flow sequence of each candidate scheduling scheme, the digital twin of the cascade reservoir group calculates the reservoir water surface elevation sequence and the overflow weir jet trajectory parameter sequence of each cascade reservoir object in a time-step manner. The pre-simulation dam inlet water level sequence is obtained from the reservoir water surface elevation sequence. The corresponding time between the pre-simulation dam inlet water level sequence and the historical measured water level sequence is calculated. The deviation value of a segment is composed of the deviation values ​​of all time periods, forming a time period deviation vector. Using the mean square error of the time period deviation vector as the loss function, and with the constraint that the evolutionary attenuation coefficient lies within a physically reasonable range calculated using the Manning formula based on the river length data, average river slope data, and roughness data of the corresponding river section object, a gradient descent method is used to solve a constrained least squares optimization problem to obtain the corrected evolutionary attenuation coefficient. The corrected evolutionary attenuation coefficient is then fed back to the tiered topology coupling graph construction module to replace the original evolutionary attenuation coefficient in the edge attributes of the corresponding river section object.

[0030] As an optional embodiment, the graph attention hydraulic inference module includes an adjacency feature splicer, a physical perception attention calculator, a message aggregator, and a time sequence encoder;

[0031] The adjacency feature splicer is a functional component configured to perform feature splicing on the node attributes of the two graph nodes connected by each edge in the ladder topological coupling graph to obtain an adjacency feature vector.

[0032] The physical perception attention calculator is a functional component configured to calculate attention weight values ​​based on the adjacent feature vector, the hydraulic propagation time delay parameter of the river section object corresponding to the edge, and the evolution attenuation coefficient. The calculation order of the attention weight value is as follows: first, calculate the physical bias term by the negative exponential function value of the hydraulic propagation time delay parameter and the linear function value of the evolution attenuation coefficient; then, calculate the feature association term by the inner product value of the adjacent feature vector after linear projection; finally, obtain the attention weight value by multiplying the physical bias term and the feature association term.

[0033] The message aggregator is a functional component configured to perform a weighted summation of the hidden state features of all neighboring graph nodes of each graph node using the attention weight value as the aggregation coefficient to obtain an aggregated feature vector.

[0034] The time encoder is a functional component configured to input the aggregated feature vectors of each graph node into a gated recurrent network in time-period order for time-series encoding, and output the inferred dam front water level sequence and the inferred reservoir flow sequence.

[0035] It should be noted that the splicing method of the adjacency feature splicer is as follows: for each edge, the upstream water level value and outflow value of the upstream graph node connected to the edge are arranged in order with the downstream water level value and outflow value of the downstream graph node to obtain a four-dimensional adjacency feature vector. The arrangement order of the four-dimensional adjacency feature vector is: upstream water level value, upstream outflow value, downstream water level value, and downstream outflow value.

[0036] The specific calculation steps of the physical perception attention calculator are as follows: First, calculate the negative exponential function value of the hydraulic propagation time delay parameter. The calculation formula is e^(-T) divided by T_mean raised to the power of T, where T is the hydraulic propagation time delay parameter of that side, and T_mean is the arithmetic mean of the hydraulic propagation time delay parameters of all river channel objects in the cascade reservoir group. Second, calculate the linear function value of the evolution attenuation coefficient. The calculation formula is a multiplied by k plus b, where k is the evolution attenuation coefficient of that side, a is the attenuation weight coefficient, and b is the bias constant. The attenuation weight coefficient a and the bias constant b are obtained through historical training... The training dataset is analyzed using linear regression to determine the evolution decay coefficient and attention weight values. The third step involves calculating the physical bias term by multiplying the results of the first and second steps. The fourth step involves calculating the feature association term by multiplying the four-dimensional adjacency feature vector with the first weight matrix to obtain the first linear projection vector, multiplying the four-dimensional adjacency feature vector with the second weight matrix to obtain the second linear projection vector, and calculating the dot product of the first and second linear projection vectors as the feature association term. The fifth step involves multiplying the physical bias term from the third step with the feature association term from the fourth step to obtain the attention weight value.

[0037] The message aggregator performs a three-layer graph structure information transfer operation. The graph structure information transfer operation is as follows: using the attention weight value of the current layer as the aggregation coefficient, the hidden state features output by all neighboring graph nodes of each graph node in the previous layer are weighted and summed to obtain the aggregated feature vector of the graph node in the current layer. After each layer of graph structure information transfer operation, the arithmetic mean and standard deviation of the aggregated feature vectors of all graph nodes in the current layer are calculated. The arithmetic mean is subtracted from each aggregated feature vector and then divided by the standard deviation to complete the normalization process.

[0038] The gated recurrent network in the time encoder includes an update gate, a reset gate, and a candidate state generator. The update gate calculates an update gate vector based on the aggregated feature vector of the current time period and the hidden state vector of the previous time period. The reset gate calculates a reset gate vector based on the aggregated feature vector of the current time period and the hidden state vector of the previous time period. The candidate state generator performs element-wise multiplication on the hidden state vector of the previous time period based on the reset gate vector to achieve selective forgetting, and then concatenates it with the aggregated feature vector of the current time period to generate a candidate hidden state vector. Finally, the hidden state vector of the current time period is obtained by linear interpolation of the hidden state vector of the previous time period and the candidate hidden state vector based on the update gate vector.

[0039] As an optional embodiment, the operation of the gated recurrent network is performed according to the following formula:

[0040] Calculation of the updated gate vector: The aggregated feature vector of the current time period is concatenated with the hidden state vector of the previous time period to obtain the concatenated vector. The concatenated vector is multiplied by the update gate weight matrix and then the update gate bias vector is added. The sum is input into the Sigmoid activation function to obtain the updated gate vector. The output value of the Sigmoid activation function is the reciprocal of the input value after negative exponentiation plus one.

[0041] Calculation of the reset gate vector: Multiply the concatenated vector by the reset gate weight matrix, add the reset gate bias vector, and input the sum into the Sigmoid activation function to obtain the reset gate vector;

[0042] Calculation of candidate hidden state vector: The reset gate vector is multiplied element-wise with the hidden state vector of the previous time period to obtain the reset hidden state vector. The reset hidden state vector is concatenated with the aggregated feature vector of the current time period to obtain the reset concatenated vector. The reset concatenated vector is multiplied with the candidate weight matrix and then the candidate bias vector is added. The sum is input into the hyperbolic tangent activation function to obtain the candidate hidden state vector. The output value of the hyperbolic tangent activation function is the difference between the positive and negative exponents of the input value divided by the sum of the positive and negative exponents.

[0043] The calculation of the hidden state vector for the current time period is as follows: the updated gate vector is multiplied element-wise with the hidden state vector of the previous time period to obtain the first weighted vector; the updated gate vector is multiplied element-wise with the candidate hidden state vector to obtain the second weighted vector; and the first weighted vector and the second weighted vector are added together to obtain the hidden state vector for the current time period.

[0044] It should be noted that the dimensions of the update gate weight matrix, reset gate weight matrix, and candidate weight matrix are the hidden state vector dimension multiplied by the sum of the input feature vector dimension and the hidden state vector dimension. The input feature vector dimension is the dimension of the aggregated feature vector, and the hidden state vector dimension is the number of cascade reservoir objects in the cascade reservoir group multiplied by the embedding representation dimension of each graph node. The embedding representation dimension is determined by the product of the node attribute dimension and the average degree of the neighborhood of the graph node. The initial values ​​of the update gate bias vector, reset gate bias vector, and candidate bias vector are all zero vectors.

[0045] The computation steps of the gated recurrent network in each time period are as follows: First, the aggregated feature vector of the current time period is concatenated with the hidden state vector of the previous time period along the feature dimension to obtain a concatenated vector; then, the concatenated vector is multiplied by the update gate weight matrix and the reset gate weight matrix respectively, and the update gate bias vector and the reset gate bias vector are added respectively, and then the update gate vector and the reset gate vector are obtained by passing the Sigmoid activation function; next, the reset gate vector is multiplied element-wise with the hidden state vector of the previous time period to obtain the reset hidden state vector, which is concatenated with the aggregated feature vector of the current time period, multiplied by the candidate weight matrix and the candidate bias vector is added, and the candidate hidden state vector is obtained by passing the hyperbolic tangent activation function; finally, the update gate vector is multiplied element-wise with the hidden state vector of the previous time period, and the update gate vector is multiplied element-wise with the candidate hidden state vector, and the two products are added to obtain the hidden state vector of the current time period;

[0046] The time-series encoder inputs the hidden state vector of each graph node for each time period into the water level fully connected layer and the flow fully connected layer, respectively. The water level fully connected layer is a linear transformation layer with an input dimension equal to the dimension of the hidden state vector and an output dimension equal to the number of time periods in the future scheduling time domain, outputting the inferred dam water level value of the graph node for that time period. The flow fully connected layer is a linear transformation layer with an input dimension equal to the dimension of the hidden state vector and an output dimension equal to the number of time periods in the future scheduling time domain, outputting the inferred reservoir flow value of the graph node for that time period.

[0047] By using a gated recurrent network to control the retention ratio of historical information through an update gate and a reset gate to achieve selective forgetting of historical information, and by using a candidate state generator to fuse the aggregated feature vector of the current period, the extrapolation process can both utilize the memory of historical hydraulic states and adapt to the hydraulic changes of the current period, thus improving the accuracy of time series extrapolation. By using two independent fully connected layers to output the extrapolated upstream water level and reservoir flow rate respectively, the joint prediction of water level and flow rate is achieved, ensuring the consistency between the two hydraulic variables.

[0048] As an optional embodiment, the hydraulic coupling constraint generation module includes a coupling tightness calculator, which is a functional component configured to calculate the hydraulic coupling tightness value between each pair of graph nodes with hydraulic upstream and downstream relationships through cross-correlation operations. The coupling tightness calculator performs the following operations:

[0049] The outflow values ​​of each adjacent time period in the deduced outflow sequence of the upstream node are subtracted to obtain the sequence of changes in upstream outflow.

[0050] The projected upstream water level sequence of the downstream graph nodes is shifted according to the number of time periods corresponding to the hydraulic propagation time delay parameters of the corresponding river channel objects between the graph nodes, to obtain the shifted water level sequence.

[0051] Subtracting the water level values ​​of each adjacent time period in the translated water level sequence yields the downstream water level change sequence.

[0052] The upstream outflow change sequence and the downstream water level change sequence are cross-correlated. The cross-correlation operation is as follows: the downstream water level change sequence is slid forward and backward by five time periods. At each sliding position, the Pearson correlation coefficient between the upstream outflow change sequence and the slid-down downstream water level change sequence is calculated. The maximum value of the Pearson correlation coefficient at all sliding positions is taken as the hydraulic coupling tightness value between the nodes of the pair.

[0053] It should be noted that the specific steps of the first-order difference operation are as follows: For the t-th time period in the outflow sequence or water level sequence, calculate the difference between the value of the t-th time period and the value of the t-1 time period to obtain the change in the t-th time period, where the value of t ranges from 2 to the total number of time periods in the future scheduling time domain, and the change in the 1st time period is set to 0.

[0054] The specific steps of the translation operation are as follows: When the number of time periods corresponding to the hydraulic propagation time delay parameter is an integer, the entire sequence of the projected upstream water level of the downstream node is shifted downstream by that integer number of time periods, and the empty time periods at the beginning of the sequence are filled with zeros; when the number of time periods corresponding to the hydraulic propagation time delay parameter is not an integer, the water level value of each time period after translation is calculated using the linear interpolation method. The calculation method of the linear interpolation method is as follows: Let the integer part of the non-integer time delay parameter be n and the decimal part be d, then the water level value of the t-th time period after translation is the sum of the water level value of the original sequence t-n time period multiplied by 1 minus d and the water level value of the original sequence t-n+1 time period multiplied by d;

[0055] The calculation steps for the Pearson correlation coefficient are as follows: First, calculate the arithmetic mean and standard deviation of the upstream outflow change series; second, calculate the arithmetic mean and standard deviation of the downstream water level change series after sliding; third, subtract the upstream arithmetic mean from each element in the upstream series, and subtract the downstream arithmetic mean from each element in the downstream series to obtain two centered series; fourth, multiply the two centered series element-wise and then sum them; fifth, divide the summation result of the fourth step by the length of the two series minus 1 to obtain the covariance; sixth, divide the covariance of the fifth step by the product of the standard deviations of the two series to obtain the Pearson correlation coefficient.

[0056] In the cross-correlation operation, five time periods are slid forward and backward to obtain a total of eleven sliding positions. Each sliding position corresponds to a Pearson correlation coefficient. The maximum value among these eleven Pearson correlation coefficients is taken as the hydraulic coupling tightness value.

[0057] By using time-delay translation in the cross-correlation operation, the upstream outflow change sequence and the downstream water level change sequence are aligned on the time axis, eliminating the interference of hydraulic propagation time delay on the correlation analysis. By detecting the peak of the Pearson correlation coefficient, the influence of upstream outflow change on downstream water level change is accurately quantified, and the obtained hydraulic coupling tightness value has a clear physical meaning and statistical significance. The cross-correlation operation at eleven sliding positions covers the entire time delay range from five negative time periods to five positive time periods, and can identify both instantaneous coupling and delayed coupling modes.

[0058] As an optional embodiment, the hydraulic coupling constraint generation module further includes a constraint matrix assembler, which is a functional component configured to assemble the hydraulic coupling tightness value, the upper limit of reservoir capacity, and the upper limit of discharge capacity into a hydraulic coupling constraint matrix; the rolling time-domain scheduling module includes a constraint decoder and a scheme searcher, the constraint decoder is a functional component configured to perform singular value decomposition on the hydraulic coupling constraint matrix to obtain effective constraint directions and corresponding boundary values, and the scheme searcher is a functional component configured to generate a set of candidate scheduling schemes using the boundary values ​​as search boundaries.

[0059] The constraint decoder is configured to perform singular value decomposition on the hydraulic coupling constraint matrix to obtain a left singular vector matrix, a diagonal singular value matrix, and a right singular vector matrix. The constraint direction corresponding to the row in the diagonal singular value matrix where the diagonal element value is greater than the maximum diagonal element value multiplied by the truncation ratio is determined as the effective constraint direction. The boundary value of the effective constraint direction is used as the time-limited storage capacity limit and time-limited outflow limit of each cascade reservoir object. The truncation ratio is determined based on the number of cascade reservoir objects in the cascade reservoir group and the storage capacity difference coefficient. The storage capacity difference coefficient is the ratio of the standard deviation to the mean of the storage capacity limit of each cascade reservoir object.

[0060] The scheme searcher is configured to generate a set of candidate scheduling schemes using a variable step-size grid search strategy. The step size of the variable step-size grid search strategy is determined as follows: for each cascade reservoir object, the reservoir capacity margin between its current time period upper limit value and its current time period actual reservoir capacity value is calculated. When the reservoir capacity margin value is greater than the average reservoir capacity margin value of the cascade reservoir group, a large step size is used for grid division. When the reservoir capacity margin value is less than or equal to the average reservoir capacity margin value of the cascade reservoir group, a small step size is used for grid division. The large step size is the reservoir capacity margin value divided by the number of coarse grid divisions, and the small step size is the reservoir capacity margin value divided by the number of fine grid divisions.

[0061] It should be noted that the singular value decomposition operation is implemented using the Jacobi iterative algorithm. The iterative steps of the Jacobi iterative algorithm are as follows: select the off-diagonal element with the largest absolute value in the diagonal singular value matrix, construct a Givens rotation matrix to perform zero-elimination operation on the off-diagonal element, and repeat the selection and zero-elimination operation until the absolute values ​​of all off-diagonal elements are less than the convergence accuracy threshold. The truncation ratio is determined as follows: calculate the ratio of the standard deviation to the mean of the upper limit of the reservoir capacity of all cascade reservoir objects in the cascade reservoir group to obtain the reservoir capacity difference coefficient. When the number of cascade reservoir objects is greater than five and the reservoir capacity difference coefficient is greater than 0.3, the truncation ratio is 0.15; otherwise, the truncation ratio is 0.05.

[0062] The number of coarse grid divisions and the number of fine grid divisions are determined based on the upper limit of the number of candidate scheduling schemes. The number of coarse grid divisions is equal to one-fifth of the upper limit of the number of candidate scheduling schemes, rounded down. The number of fine grid divisions is equal to three-quarters of the upper limit of the number of candidate scheduling schemes, rounded down. The upper limit of the number of candidate scheduling schemes is determined based on the duration limit of the rolling scheduling cycle and the calculation time of a single scheme evaluation.

[0063] The actual reservoir capacity value for the specified time period is the current time period water level value in the simulated upstream dam water level sequence output by the graph attention hydraulic simulation module, obtained by querying the reservoir capacity characteristic line data table of the cascade reservoir object; the average reservoir capacity margin value of the cascade reservoir group is the sum of the reservoir capacity margin values ​​of all cascade reservoir objects divided by the total number of cascade reservoir objects.

[0064] The hydraulic coupling constraint matrix is ​​decomposed into independent constraint directions by singular value decomposition. Effective constraint directions are selected by truncation ratio, which reduces the dimensionality of the constraint space and reduces the interference of redundant constraints on scheduling optimization. A variable step size grid search strategy is adopted to improve search efficiency in areas with large reservoir capacity margin and to ensure search accuracy in areas with small reservoir capacity margin, thus achieving an adaptive balance between search efficiency and search accuracy.

[0065] As an optional embodiment, the rolling time-domain scheduling module further includes a selector, which is a functional component configured to select a target scheduling scheme from a set of candidate scheduling schemes. The selector is configured to calculate a comprehensive evaluation value for each candidate scheduling scheme. The comprehensive evaluation value is calculated as the sum of four terms: a weighted term of the quotient of the flood control weight coefficient and the highest water level in front of the dam, a weighted term of the quotient of the flood control weight coefficient and the peak flow value at the downstream control section, a weighted term of the product of the power generation weight coefficient and the estimated power generation value, and a penalty term of the quotient of the power generation weight coefficient and the water abandonment value. The selector determines the candidate scheduling scheme with the largest comprehensive evaluation value as the target scheduling scheme.

[0066] It should be noted that the specific calculation steps of the comprehensive evaluation value are as follows: First, for each candidate scheduling scheme, extract the highest water level value in front of the dam for each time period within the future scheduling time domain from the discharge flow sequence and the projected dam-front water level sequence of each cascade reservoir object in the scheme. The highest water level value in front of the dam is the maximum value among the projected dam-front water level values ​​of all cascade reservoir objects in that time period. Second, based on the discharge flow sequence of each cascade reservoir object, obtain the flow value of the downstream flood control section in each time period through hydraulic evolution calculation, and take the maximum value among the flow values ​​of the downstream flood control section in each time period within the future scheduling time domain as the peak flow value. Third, calculate the power generation output for each time period based on the discharge flow sequence and the projected dam-front water level sequence of each cascade reservoir object. The power generation output is calculated by multiplying the downstream flow rate by the power generation head and then by the power generation efficiency coefficient. The power generation head is calculated by subtracting the downstream tailwater level from the projected upstream water level. The power generation output values ​​for each time period are summed to obtain the estimated power generation value. The fourth step is to sum the differences between the projected reservoir flow sequence and the downstream flow sequence for each cascade reservoir to obtain the water wastage value. The fifth step is to calculate the comprehensive evaluation value, which is calculated by dividing the flood control weight coefficient by the highest upstream water level, adding the flood control weight coefficient by the downstream control section peak flow, adding the power generation weight coefficient by the estimated power generation value, and subtracting the power generation weight coefficient by the water wastage value. The sixth step is to sort all candidate scheduling schemes from largest to smallest comprehensive evaluation value and select the candidate scheduling scheme with the largest comprehensive evaluation value as the target scheduling scheme.

[0067] The flood control weight coefficient and power generation weight coefficient are determined as follows: based on the flood control scheduling regulations of the basin where the cascade reservoir group is located, the proportion of the number of times flood control constraints are triggered in the historical scheduling cycle to the total number of scheduling times is statistically analyzed. When the proportion is greater than 0.5, the flood control weight coefficient is 0.6 and the power generation weight coefficient is 0.4. When the proportion is less than or equal to 0.5, the flood control weight coefficient is 0.4 and the power generation weight coefficient is 0.6.

[0068] By combining flood control weighting and power generation weighting terms in the comprehensive evaluation value, the dispatching scheme achieves a balance between the two objectives of flood control safety and power generation benefits. By introducing a water abandonment penalty term, the extreme scheme of abandoning a large amount of water in pursuit of maximizing power generation is avoided. By dynamically determining the flood control weight coefficient and power generation weight coefficient based on the flood control constraint trigger ratio in historical dispatching data, the weight allocation conforms to the actual flood control needs of the basin.

[0069] As an optional embodiment, the twin pre-simulation correction module further includes a topology corrector, which is a functional component configured to automatically adjust the number of graph nodes and connectivity relationships of the tiered topology coupling graph based on the time-period deviation vector; the topology corrector is configured to perform the following judgments and operations:

[0070] When the absolute value of the deviation in the time period deviation vector exceeds the topology splitting threshold for three or more consecutive time periods, a topology splitting operation is triggered. The topology splitting operation splits the edge between the corresponding upstream graph node and downstream graph node into two new edges, and inserts a virtual interval graph node between the two new edges. The dam-front water level value in the node attribute of the virtual interval graph node is the arithmetic mean of the dam-front water level values ​​of the upstream graph node and the downstream graph node. The reservoir capacity value in the node attribute of the virtual interval graph node is obtained by querying the reservoir capacity feature line data table based on the dam-front water level value of the virtual interval graph node. The evolution decay coefficient in the edge attribute of the two new edges is half of the evolution decay coefficient of the original edge.

[0071] When the absolute value of the deviation for all time periods in the time period deviation vector is lower than the topology merging threshold, a topology merging operation is triggered. The topology merging operation merges the corresponding upstream graph node, downstream graph node, and the edge connecting the two graph nodes into a new graph node. The dam front water level value in the node attributes of the new graph node is the weighted average of the dam front water level values ​​of the two original graph nodes. The weight of the weighted average is the proportion of the reservoir capacity value of each original graph node to the sum of the reservoir capacity values ​​of the two original graph nodes.

[0072] It should be noted that the determination method for the topology splitting threshold and the topology merging threshold is as follows: extract all time period deviation vectors from historical scheduling data, calculate the standard deviation and mean of all time period deviation vectors, calculate the ratio of the standard deviation to the mean to obtain the coefficient of variation, when the coefficient of variation is greater than 1, the topology splitting threshold is 2 times the standard deviation and the topology merging threshold is 0.3 times the standard deviation, when the coefficient of variation is less than or equal to 1, the topology splitting threshold is 1.5 times the standard deviation and the topology merging threshold is 0.2 times the standard deviation;

[0073] After the topology splitting operation, the virtual interval graph node inherits the reservoir capacity feature line data table of the downstream graph node as its own reservoir capacity feature line data table, and the outflow value in the node attributes of the virtual interval graph node is initialized to zero; after the topology splitting operation is completed, the graph attention hydraulic inference module re-executes the graph attention message passing operation to update the hidden state features of all graph nodes.

[0074] After the topology merging operation, the new graph node inherits all adjacent edges from the upstream and downstream graph nodes. Edges that originally pointed to the upstream and downstream graph nodes now point to the new graph node, and edges that originally pointed to other graph nodes from the upstream and downstream graph nodes are retained to point to other graph nodes from the new graph node. The storage capacity value in the node attribute of the new graph node is the sum of the storage capacity values ​​of the two original graph nodes.

[0075] The automatic splitting operation of the topology modifier increases the number of graph nodes in areas where the model deviates significantly from the actual hydraulic response, thereby improving the model resolution in those areas. The automatic merging operation reduces the number of graph nodes in areas where the model deviates slightly from the actual hydraulic response, thereby reducing the computational complexity of the model. The topology correction threshold is dynamically determined based on the coefficient of variation of the deviation vector over historical periods, enabling the splitting and merging decisions to adapt to the differences in hydraulic characteristics across different watersheds.

[0076] As an optional embodiment, the evolution attenuation coefficient correction process of the twin pre-simulation correction module adopts a physically constrained iterative least squares method, which is executed according to the following steps:

[0077] Based on the channel length, average gradient, and roughness data of each river section, the Manning formula is used to calculate the baseline value of the expected evolutionary attenuation coefficient for that river section. The Manning formula states that the average flow velocity equals the reciprocal of the roughness multiplied by the square of the hydraulic radius, then multiplied by the square root of the gradient. The baseline value of the expected evolutionary attenuation coefficient equals the average flow velocity multiplied by the hydraulic propagation time delay parameter, then divided by the channel length. Based on the baseline value of the expected evolutionary attenuation coefficient and the fluctuation range, the upper and lower bounds of the physically reasonable interval are determined. The upper bound is the baseline value of the expected evolutionary attenuation coefficient plus the fluctuation range, and the lower bound is the baseline value of the expected evolutionary attenuation coefficient minus the fluctuation range. The fluctuation range is determined based on the standard deviation of the historical evolutionary attenuation coefficient sequence of the river section.

[0078] Initialize the current evolution decay coefficient to the evolution decay coefficient in the current edge attribute output by the ladder topology coupling graph construction module;

[0079] Calculate the Jacobian matrix of the time period deviation vector with respect to the current evolution decay coefficient, where each element of the Jacobian matrix is ​​the partial derivative of the deviation value of the corresponding time period in the time period deviation vector with respect to the current evolution decay coefficient;

[0080] The gradient descent direction is calculated by multiplying the transpose of the Jacobian matrix with the time-period deviation vector, and the gradient descent direction is the negative direction of the product.

[0081] Update the current evolution decay coefficient along the gradient descent direction with an iterative step size;

[0082] When the updated current evolution decay coefficient exceeds the physical reasonable range, the current evolution decay coefficient is truncated to the nearest boundary value of the physical reasonable range;

[0083] Repeat the Jacobian matrix calculation, gradient descent direction calculation, update, and truncation steps described above until the mean square error of the time-period deviation vector is less than the convergence threshold or the maximum number of iterations is reached.

[0084] It should be noted that the roughness value in the Manning formula is obtained by back-calculating the historical measured water level and flow data of the river section object. The back-calculation method is as follows: calculate the cross-sectional area and hydraulic radius based on the measured flow and water level values, substitute the square of the third of the hydraulic radius and the square root of the slope into the Manning formula, and solve for the roughness value, which is equal to the square of the third of the hydraulic radius multiplied by the square root of the slope and divided by the average flow velocity. Take the arithmetic mean of the roughness values ​​obtained from back-calculating multiple measured data points as the roughness value of the river section object.

[0085] The fluctuation range is determined as follows: the evolution attenuation coefficient of the river section object in each historical rolling scheduling cycle is extracted from historical scheduling data to form a historical evolution attenuation coefficient sequence, and the standard deviation of the historical evolution attenuation coefficient sequence is calculated. The fluctuation range is the standard deviation multiplied by 2. When the number of samples in the historical evolution attenuation coefficient sequence is less than 30, the fluctuation range is a fixed value of 0.15.

[0086] The iteration step size adopts an adaptive step size strategy. The initial iteration step size is set to 1.0. When the ratio of the decrease in mean square error between two consecutive iterations is greater than 0.5, the current iteration step size is maintained. When the ratio is less than 0.5, the iteration step size is halved. The convergence threshold is determined as follows: when the accuracy level of the water level measuring device is 0.01 meters, the convergence threshold is 0.0001; when the accuracy level is 0.05 meters, the convergence threshold is 0.0025. The convergence threshold is the square of the water level measuring accuracy value. The maximum number of iterations is determined as follows: the number of iterations for correcting the evolution attenuation coefficient is extracted from historical scheduling data, the maximum number of iterations is taken, a 20% margin is added, and then rounded up.

[0087] The baseline value of the expected evolution attenuation coefficient is calculated based on the actual roughness, slope, and cross-sectional parameters of the river channel using the Manning formula, thus setting a reasonable range for the evolution attenuation coefficient based on physical laws. Through truncated gradient descent iteration, the parameters are guided to optimize in the direction of reducing deviation while the physically reasonable range is used as a hard constraint to prevent the parameters from being updated to a numerical region without physical meaning. The adaptive step size strategy maintains a large step size to accelerate convergence when the convergence rate is fast, and automatically reduces the step size to improve accuracy when the convergence rate slows down.

[0088] As an optional embodiment, the twin pre-simulation correction module further includes a constraint tightness feedback unit, which is a functional component configured to calculate the constraint adjustment amount based on the deviation direction between the pre-simulated dam upstream water level sequence and the historical measured water level sequence, and feed it back to the next rolling scheduling cycle to correct the hydraulic coupling tightness value; the constraint tightness feedback unit is configured as follows:

[0089] For each pair of graph nodes with a hydraulic upstream-downstream relationship, the arithmetic mean of the predicted water level sequence corresponding to the pair of graph nodes is obtained by taking the arithmetic mean of all time periods of the predicted water level sequence, and the measured water level is obtained by taking the arithmetic mean of all time periods of the historical measured water level sequence corresponding to the pair of graph nodes.

[0090] When the average water level in the preliminary calculation is greater than the average water level in the actual calculation, the system deviation direction is determined to be water level overestimation. A negative constraint adjustment amount is calculated. The negative constraint adjustment amount is the negative adjustment coefficient multiplied by the absolute value of the difference between the average water level in the preliminary calculation and the average water level in the actual calculation. When the hydraulic coupling constraint generation module calculates the hydraulic coupling tightness value of the pair of graph nodes in the next rolling scheduling cycle, it adds the negative constraint adjustment amount to the current hydraulic coupling tightness value to obtain the corrected hydraulic coupling tightness value.

[0091] When the average water level in the preliminary calculation is less than the average water level in the actual calculation, the system deviation direction is determined to be an underestimation of the water level. A positive constraint adjustment amount is calculated. The positive constraint adjustment amount is the positive adjustment coefficient multiplied by the absolute value of the difference between the average water level in the actual calculation and the average water level in the preliminary calculation. When the hydraulic coupling constraint generation module calculates the hydraulic coupling tightness value of the pair of graph nodes in the next rolling scheduling cycle, it adds the positive constraint adjustment amount to the current hydraulic coupling tightness value to obtain the corrected hydraulic coupling tightness value.

[0092] When the average pre-simulated water level is equal to the average measured water level, the constraint adjustment amount is set to zero.

[0093] It should be noted that the negative and positive adjustment coefficients are determined as follows: All scheduling cycles in which the average pre-scheduled water level is greater than the average measured water level are extracted from historical scheduling data. The regression coefficients between the pre-scheduled deviation and the coupling tightness correction in these scheduling cycles are calculated using the least squares method. The absolute value of the regression coefficient is used as the negative adjustment coefficient. Similarly, all scheduling cycles in which the average pre-scheduled water level is less than the average measured water level are extracted from historical scheduling data, and the positive adjustment coefficient is calculated using the same method. When historical scheduling data is insufficient, both the negative and positive adjustment coefficients start with an initial empirical value of 0.1 and are dynamically adjusted based on actual feedback within the first 10 rolling scheduling cycles.

[0094] The feedback delay of the constraint tightness feedback device is one rolling scheduling cycle, that is, the constraint adjustment amount calculated in the current rolling scheduling cycle will take effect in the next rolling scheduling cycle; after receiving the constraint adjustment amount, the hydraulic coupling constraint generation module synchronously updates the hydraulic coupling tightness value of all current graph node pairs. The synchronous update method is as follows: add the hydraulic coupling tightness value of each graph node pair to the corresponding constraint adjustment amount. When the sum is less than zero, take zero; when the sum is greater than 1, take 1. The updated hydraulic coupling tightness value is reassembled into a hydraulic coupling constraint matrix and then input into the rolling time domain scheduling module.

[0095] The constraint tightness feedback device enables the reverse transmission of pre-simulation deviations to scheduling constraints. When the pre-simulation water level is systematically overestimated, the hydraulic coupling tightness value is increased to tighten the scheduling constraints in the next cycle. When the pre-simulation water level is systematically underestimated, the hydraulic coupling tightness value is decreased to relax the scheduling constraints in the next cycle, forming a two-way closed loop of pre-simulation, evaluation, feedback, and correction. The magnitude of the constraint adjustment is proportional to the amount of pre-simulation deviation, so that the magnitude of constraint correction matches the degree of deviation, avoiding over-correction or under-correction.

[0096] As an optional embodiment, the digital twin of the cascade reservoir group includes a topographic geometry model, a reservoir water surface model, and a downstream water tongue model;

[0097] The terrain geometry model is a functional component configured to construct a triangular mesh terrain surface based on the digital elevation model raster data and reservoir bank boundary vector data of each cascade reservoir object. The vertices of the triangular mesh terrain surface include elevation points in the digital elevation model raster and boundary points in the reservoir bank boundary vector data.

[0098] The reservoir water surface model is a functional component configured to update the reservoir water surface at different time periods based on the discharge flow sequence of each candidate scheduling scheme and the projected dam inlet water level sequence. The update process is as follows: Based on the projected dam inlet water level value and the triangular mesh terrain surface for that time period, the reservoir inundation range polygon is determined by comparing the elevation values ​​of each vertex in the triangular mesh terrain surface with the projected dam inlet water level value. Vertices with elevation values ​​lower than the projected dam inlet water level value are marked as inundated vertices, and the convex hull boundary of all inundated vertices constitutes the reservoir inundation range polygon; the reservoir water surface area value for that time period is calculated based on the geometric intersection of the reservoir inundation range polygon and the triangular mesh terrain surface; the elevation values ​​of all terrain surface vertices within the reservoir inundation range polygon are replaced with the projected dam inlet water level value for that time period to obtain the reservoir water surface elevation value for that time period.

[0099] The discharge jet model is a functional component configured to calculate the trajectory of the overflow jet from the spillway based on the discharge flow sequence of each candidate scheduling scheme and the upstream water level of each cascade reservoir. The calculation process is as follows: the overflow weir crest head value is calculated according to the broad-crested weir flow formula, where the discharge flow value equals the flow coefficient multiplied by the weir crest width and then multiplied by the cube of the overflow weir crest head value; the jet velocity value is calculated based on the overflow weir crest head value and the weir crest elevation value in the weir body geometric parameters; the jet angle value is calculated based on the weir surface slope value in the weir body geometric parameters; and the jet trajectory parameters and landing point position parameters are calculated according to the projectile motion trajectory formula based on the jet velocity value, jet angle value, and gravity acceleration parameters.

[0100] It should be noted that: the construction of the triangular network terrain surface adopts the Delaunay triangulation algorithm. The constraints of the Delaunay triangulation algorithm are: maximizing the minimum interior angle of all triangles in the triangular network, and each side in the reservoir bank boundary line vector data must appear as a side of the triangular network; the spatial resolution of the digital elevation model raster data is determined according to the reservoir capacity of the cascade reservoir object. When the reservoir capacity is greater than 100 million cubic meters, the spatial resolution is 10 meters, and when the reservoir capacity is less than or equal to 100 million cubic meters, the spatial resolution is 5 meters.

[0101] The vertex-by-vertex elevation comparison also includes a connectivity assessment of the submerged vertices, retaining only the set of submerged vertices connected to the central reservoir area, and excluding isolated submerged areas caused by topographic depressions; the reservoir surface area is calculated by summing the areas of all submerged triangular units, and the area of ​​each submerged triangular unit is calculated using Heron's formula from its three side lengths. Heron's formula is as follows: if the three side lengths of a triangle are a, b, and c, the semi-perimeter s is equal to a plus b plus c divided by 2, then the area of ​​the triangle is equal to the square root of s multiplied by s minus a multiplied by s minus b multiplied by s minus c.

[0102] The flow coefficient in the broad-crested weir flow formula is determined according to the weir structure type. When the weir is a WES practical weir, the flow coefficient is 0.502; when the weir is a broad-crested weir, the flow coefficient is 0.36. The flow coefficient value is calibrated based on hydraulic model test data. The formula for calculating the jet velocity is: jet velocity equals the flow coefficient multiplied by the 0.5 power of the overflow weir crest head value, multiplied by the 0.5 power of twice the gravitational acceleration. The jet angle value is equal to the weir slope value. The horizontal displacement component of the projectile motion trajectory formula is the jet velocity multiplied by the cosine of the jet angle, multiplied by the flight time. The vertical displacement component is the jet velocity multiplied by the sine of the jet angle, multiplied by the flight time, minus 0.5 multiplied by the gravitational acceleration multiplied by the square of the flight time. The flight time is obtained by solving the quadratic equation when the vertical displacement component is zero.

[0103] A three-dimensional surface consistent with the actual terrain was constructed by Delaunay triangulation of the terrain geometry model, providing a geometric basis for the accurate calculation of the reservoir water surface. By comparing the elevation of each vertex of the reservoir water surface model and determining the inundation range polygon, the dynamic updating of the reservoir water surface elevation and area with the change of the water level in front of the dam was realized. By using the broad-crested weir formula and projectile motion trajectory calculation of the spillway tongue model, the jet trajectory and landing point of the overflow weir were accurately simulated, providing a visual basis for the engineering safety assessment of the scheduling scheme.

[0104] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A smart management and data visualization system for water resource scheduling, characterized in that, It includes a cascade topology coupling graph construction module, a graph attention hydraulic inference module, a hydraulic coupling constraint generation module, a rolling time-domain scheduling module, and a twin pre-simulation correction module; The cascade topology coupling graph construction module is configured to construct a cascade topology coupling graph by combining each cascade reservoir object in the cascade reservoir group with the river channel objects connecting adjacent cascade reservoirs. The node attributes of the cascade topology coupling graph include the upstream water level and outflow value of each cascade reservoir object at the current scheduling time. The edge attributes of the cascade topology coupling graph include the hydraulic propagation time delay parameter and evolution attenuation coefficient of each river channel object. The hydraulic propagation time delay parameter is the number of time periods required for the outflow from the upstream cascade reservoir object to propagate to the downstream cascade reservoir object. The evolution attenuation coefficient is the attenuation ratio of the outflow during the propagation process in the river channel. The graph attention hydraulic extrapolation module is configured to perform graph attention message passing operations on the cascade topology coupling graph. The graph attention message passing operation is as follows: for each edge in the cascade topology coupling graph, the water level gradient value is calculated by the upstream graph node connected by the edge and the downstream graph node. The attention weight value of the edge is calculated by the water level gradient value, the hydraulic propagation time delay parameter of the edge, and the evolution attenuation coefficient. The node features of the neighboring graph nodes are weighted and aggregated using the attention weight value as the aggregation coefficient to obtain an aggregated feature vector. The aggregated feature vector is input into a gated recurrent network for time-series encoding, and the extrapolated upstream water level sequence and the extrapolated reservoir flow sequence of each graph node in the future scheduling time domain are output. The time domain length of the future scheduling time domain is determined according to the flood forecast lead time. The hydraulic coupling constraint generation module is configured to calculate the hydraulic coupling tightness value between each cascade reservoir object based on the edge attributes of the cascade topological coupling graph and the inferred dam front water level sequence, and to compile the hydraulic coupling tightness value with the upper limit of the reservoir capacity and the upper limit of the discharge capacity of each cascade reservoir object into a hydraulic coupling constraint matrix. The hydraulic coupling tightness value is a quantification of the correlation between the change in the outflow of the upstream cascade reservoir object and the change in the dam front water level of the downstream cascade reservoir object. The rolling time-domain scheduling module is configured to decode the hydraulic coupling constraint matrix into the time-limited reservoir capacity and time-limited outflow of each cascade reservoir object within each rolling scheduling cycle. Using the time-limited reservoir capacity and time-limited outflow as search boundaries, and taking the flood control water level requirements, upstream water level change rate requirements, and downstream safe discharge requirements of each cascade reservoir object as hard constraints, and taking the maximum power generation requirement as the optimization direction, a set of candidate scheduling schemes is generated. The optimization direction of the maximum power generation requirement is to maximize the cumulative power generation of all cascade reservoir objects in the future scheduling time domain. Each candidate scheduling scheme in the set of candidate scheduling schemes includes the downstream discharge flow sequence of each cascade reservoir object in the future scheduling time domain. The twin pre-simulation correction module is configured to input the outflow sequence of each candidate scheduling scheme into the digital twin of the cascade reservoir group to perform hydraulic state extrapolation and obtain the pre-simulation dam front water level sequence, calculate the time-period deviation vector between the pre-simulation dam front water level sequence and the historical measured water level sequence, correct the evolution attenuation coefficient of each river section object in the cascade topology coupling graph according to the time-period deviation vector, and send the corrected evolution attenuation coefficient back to the cascade topology coupling graph construction module to update the edge attributes. The digital twin of the cascade reservoir group is a digital three-dimensional simulation model constructed based on the topographic data and hydraulic structure parameters of each cascade reservoir object.

2. The intelligent management and data visualization management system for water resource scheduling according to claim 1, characterized in that, The graph attention hydraulic inference module includes an adjacency feature splicer, a physical perception attention calculator, a message aggregator, and a temporal encoder; The adjacency feature splicer is a functional component configured to perform feature splicing on the node attributes of the two graph nodes connected by each edge in the ladder topological coupling graph to obtain an adjacency feature vector. The physical perception attention calculator is a functional component configured to calculate attention weight values ​​based on the adjacent feature vector, the hydraulic propagation time delay parameter of the river section object corresponding to the edge, and the evolution attenuation coefficient. The calculation order of the attention weight value is as follows: first, calculate the physical bias term by the negative exponential function value of the hydraulic propagation time delay parameter and the linear function value of the evolution attenuation coefficient; then, calculate the feature association term by the inner product value of the adjacent feature vector after linear projection; finally, obtain the attention weight value by multiplying the physical bias term and the feature association term. The message aggregator is a functional component configured to perform a weighted summation of the hidden state features of all neighboring graph nodes of each graph node using the attention weight value as the aggregation coefficient to obtain an aggregated feature vector. The time encoder is a functional component configured to input the aggregated feature vectors of each graph node into a gated recurrent network in time-period order for time-series encoding, and output the inferred dam front water level sequence and the inferred reservoir flow sequence.

3. The intelligent management and data visualization management system for water resource scheduling according to claim 2, characterized in that, The operation of the gated cyclic network is performed according to the following formula: Calculation of the updated gate vector: The aggregated feature vector of the current time period is concatenated with the hidden state vector of the previous time period to obtain the concatenated vector. The concatenated vector is multiplied by the update gate weight matrix and then the update gate bias vector is added. The sum is input into the Sigmoid activation function to obtain the updated gate vector. The output value of the Sigmoid activation function is the reciprocal of the input value after negative exponentiation plus one. Calculation of the reset gate vector: Multiply the concatenated vector by the reset gate weight matrix, add the reset gate bias vector, and input the sum into the Sigmoid activation function to obtain the reset gate vector; Calculation of candidate hidden state vector: The reset gate vector is multiplied element-wise with the hidden state vector of the previous time period to obtain the reset hidden state vector. The reset hidden state vector is concatenated with the aggregated feature vector of the current time period to obtain the reset concatenated vector. The reset concatenated vector is multiplied with the candidate weight matrix and then the candidate bias vector is added. The sum is input into the hyperbolic tangent activation function to obtain the candidate hidden state vector. The output value of the hyperbolic tangent activation function is the difference between the positive and negative exponents of the input value divided by the sum of the positive and negative exponents. The calculation of the hidden state vector for the current time period is as follows: the updated gate vector is multiplied element-wise with the hidden state vector of the previous time period to obtain the first weighted vector; the updated gate vector is multiplied element-wise with the candidate hidden state vector to obtain the second weighted vector; and the first weighted vector and the second weighted vector are added together to obtain the hidden state vector for the current time period.

4. The intelligent management and data visualization management system for water resource scheduling according to claim 1, characterized in that, The hydraulic coupling constraint generation module includes a coupling tightness calculator, which is a functional component configured to calculate the hydraulic coupling tightness value between each pair of graph nodes with hydraulic upstream and downstream relationships through cross-correlation operations. The coupling tightness calculator performs the following operations: The outflow values ​​of each adjacent time period in the deduced outflow sequence of the upstream node are subtracted to obtain the sequence of changes in upstream outflow. The projected upstream water level sequence of the downstream graph nodes is shifted according to the number of time periods corresponding to the hydraulic propagation time delay parameters of the corresponding river channel objects between the graph nodes, to obtain the shifted water level sequence. Subtracting the water level values ​​of each adjacent time period in the translated water level sequence yields the downstream water level change sequence. The upstream outflow change sequence and the downstream water level change sequence are cross-correlated. The cross-correlation operation is as follows: the downstream water level change sequence is slid forward and backward by five time periods. At each sliding position, the Pearson correlation coefficient between the upstream outflow change sequence and the slid-down downstream water level change sequence is calculated. The maximum value of the Pearson correlation coefficient at all sliding positions is taken as the hydraulic coupling tightness value between the nodes of the pair.

5. The intelligent management and data visualization management system for water resource scheduling according to claim 4, characterized in that, The hydraulic coupling constraint generation module further includes a constraint matrix assembler, which is a functional component configured to assemble the hydraulic coupling tightness value, reservoir capacity upper limit value, and discharge capacity upper limit value into a hydraulic coupling constraint matrix; the rolling time-domain scheduling module includes a constraint decoder and a scheme searcher, which is a functional component configured to perform singular value decomposition on the hydraulic coupling constraint matrix to obtain effective constraint directions and corresponding boundary values, and the scheme searcher is a functional component configured to generate a set of candidate scheduling schemes using the boundary values ​​as search boundaries. The constraint decoder is configured to perform singular value decomposition on the hydraulic coupling constraint matrix to obtain a left singular vector matrix, a diagonal singular value matrix, and a right singular vector matrix. The constraint direction corresponding to the row in the diagonal singular value matrix where the diagonal element value is greater than the maximum diagonal element value multiplied by the truncation ratio is determined as the effective constraint direction. The boundary value of the effective constraint direction is used as the time-limited storage capacity limit and time-limited outflow limit of each cascade reservoir object. The truncation ratio is determined based on the number of cascade reservoir objects in the cascade reservoir group and the storage capacity difference coefficient. The storage capacity difference coefficient is the ratio of the standard deviation to the mean of the storage capacity limit of each cascade reservoir object. The scheme searcher is configured to generate a set of candidate scheduling schemes using a variable step-size grid search strategy. The step size of the variable step-size grid search strategy is determined as follows: for each cascade reservoir object, the reservoir capacity margin between its current time period upper limit value and its current time period actual reservoir capacity value is calculated. When the reservoir capacity margin value is greater than the average reservoir capacity margin value of the cascade reservoir group, a large step size is used for grid division. When the reservoir capacity margin value is less than or equal to the average reservoir capacity margin value of the cascade reservoir group, a small step size is used for grid division. The large step size is the reservoir capacity margin value divided by the number of coarse grid divisions, and the small step size is the reservoir capacity margin value divided by the number of fine grid divisions.

6. The intelligent management and data visualization management system for water resource scheduling according to claim 5, characterized in that, The rolling time-domain scheduling module also includes a selector, which is a functional component configured to select a target scheduling scheme from a set of candidate scheduling schemes. The selector is configured to calculate a comprehensive evaluation value for each candidate scheduling scheme. The comprehensive evaluation value is calculated as the sum of four factors: a weighted term of the quotient of the flood control weight coefficient and the highest water level in front of the dam, a weighted term of the quotient of the flood control weight coefficient and the peak flow value at the downstream control section, a weighted term of the product of the power generation weight coefficient and the estimated power generation value, and a penalty term of the quotient of the power generation weight coefficient and the water abandonment value. The selector determines the candidate scheduling scheme with the largest comprehensive evaluation value as the target scheduling scheme.

7. The intelligent management and data visualization management system for water resource scheduling according to claim 1, characterized in that, The twin pre-simulation correction module also includes a topology corrector, which is a functional component configured to automatically adjust the number of graph nodes and connection relationships of the tiered topology coupling graph based on the time-period deviation vector; the topology corrector is configured to perform the following judgments and operations: When the absolute value of the deviation in the time period deviation vector exceeds the topology splitting threshold for three or more consecutive time periods, a topology splitting operation is triggered. The topology splitting operation splits the edge between the corresponding upstream graph node and downstream graph node into two new edges, and inserts a virtual interval graph node between the two new edges. The dam-front water level value in the node attribute of the virtual interval graph node is the arithmetic mean of the dam-front water level values ​​of the upstream graph node and the downstream graph node. The reservoir capacity value in the node attribute of the virtual interval graph node is obtained by querying the reservoir capacity feature line data table based on the dam-front water level value of the virtual interval graph node. The evolution decay coefficient in the edge attribute of the two new edges is half of the evolution decay coefficient of the original edge. When the absolute value of the deviation for all time periods in the time period deviation vector is lower than the topology merging threshold, a topology merging operation is triggered. The topology merging operation merges the corresponding upstream graph node, downstream graph node, and the edge connecting the two graph nodes into a new graph node. The dam front water level value in the node attributes of the new graph node is the weighted average of the dam front water level values ​​of the two original graph nodes. The weight of the weighted average is the proportion of the reservoir capacity value of each original graph node to the sum of the reservoir capacity values ​​of the two original graph nodes.

8. The intelligent management and data visualization management system for water resource scheduling according to claim 7, characterized in that, The evolution attenuation coefficient correction process of the twin pre-simulation correction module adopts a physically constrained iterative least squares method, which is executed according to the following steps: Based on the channel length, average gradient, and roughness data of each river section, the Manning formula is used to calculate the baseline value of the expected evolutionary attenuation coefficient for that river section. The Manning formula states that the average flow velocity equals the reciprocal of the roughness multiplied by the square of the hydraulic radius, then multiplied by the square root of the gradient. The baseline value of the expected evolutionary attenuation coefficient equals the average flow velocity multiplied by the hydraulic propagation time delay parameter, then divided by the channel length. Based on the baseline value of the expected evolutionary attenuation coefficient and the fluctuation range, the upper and lower bounds of the physically reasonable interval are determined. The upper bound is the baseline value of the expected evolutionary attenuation coefficient plus the fluctuation range, and the lower bound is the baseline value of the expected evolutionary attenuation coefficient minus the fluctuation range. The fluctuation range is determined based on the standard deviation of the historical evolutionary attenuation coefficient sequence of the river section. Initialize the current evolution decay coefficient to the evolution decay coefficient in the current edge attribute output by the ladder topology coupling graph construction module; Calculate the Jacobian matrix of the time period deviation vector with respect to the current evolution decay coefficient, where each element of the Jacobian matrix is ​​the partial derivative of the deviation value of the corresponding time period in the time period deviation vector with respect to the current evolution decay coefficient; The gradient descent direction is calculated by multiplying the transpose of the Jacobian matrix with the time-period deviation vector, and the gradient descent direction is the negative direction of the product. Update the current evolution decay coefficient along the gradient descent direction with an iterative step size; When the updated current evolution decay coefficient exceeds the physical reasonable range, the current evolution decay coefficient is truncated to the nearest boundary value of the physical reasonable range; Repeat the Jacobian matrix calculation, gradient descent direction calculation, update, and truncation steps described above until the mean square error of the time-period deviation vector is less than the convergence threshold or the maximum number of iterations is reached.

9. The intelligent management and data visualization management system for water resource scheduling according to claim 1, characterized in that, The twin pre-simulation correction module also includes a constraint tightness feedback unit, which is a functional component configured to calculate the constraint adjustment amount based on the deviation direction between the pre-simulated upstream dam water level sequence and the historical measured water level sequence, and feed it back to the next rolling scheduling cycle to correct the hydraulic coupling tightness value; the constraint tightness feedback unit is configured as follows: For each pair of graph nodes with a hydraulic upstream-downstream relationship, the arithmetic mean of the predicted water level sequence corresponding to the pair of graph nodes is obtained by taking the arithmetic mean of all time periods of the predicted water level sequence, and the measured water level is obtained by taking the arithmetic mean of all time periods of the historical measured water level sequence corresponding to the pair of graph nodes. When the average water level in the preliminary calculation is greater than the average water level in the actual calculation, the system deviation direction is determined to be water level overestimation. A negative constraint adjustment amount is calculated. The negative constraint adjustment amount is the negative adjustment coefficient multiplied by the absolute value of the difference between the average water level in the preliminary calculation and the average water level in the actual calculation. When the hydraulic coupling constraint generation module calculates the hydraulic coupling tightness value of the pair of graph nodes in the next rolling scheduling cycle, it adds the negative constraint adjustment amount to the current hydraulic coupling tightness value to obtain the corrected hydraulic coupling tightness value. When the average water level in the preliminary calculation is less than the average water level in the actual calculation, the system deviation direction is determined to be an underestimation of the water level. A positive constraint adjustment amount is calculated. The positive constraint adjustment amount is the positive adjustment coefficient multiplied by the absolute value of the difference between the average water level in the actual calculation and the average water level in the preliminary calculation. When the hydraulic coupling constraint generation module calculates the hydraulic coupling tightness value of the pair of graph nodes in the next rolling scheduling cycle, it adds the positive constraint adjustment amount to the current hydraulic coupling tightness value to obtain the corrected hydraulic coupling tightness value. When the average pre-simulated water level is equal to the average measured water level, the constraint adjustment amount is set to zero.

10. A smart management and data visualization management system for water resource scheduling according to claim 1, characterized in that, The digital twin of the cascade reservoir group includes a topographic geometry model, a reservoir water surface model, and a discharge tongue model; The terrain geometry model is a functional component configured to construct a triangular mesh terrain surface based on the digital elevation model raster data and reservoir bank boundary vector data of each cascade reservoir object. The vertices of the triangular mesh terrain surface include elevation points in the digital elevation model raster and boundary points in the reservoir bank boundary vector data. The reservoir water surface model is a functional component configured to update the reservoir water surface at each time period according to the discharge flow sequence of each candidate scheduling scheme and the projected upstream water level sequence. The update process is as follows: based on the projected upstream water level value and the triangular mesh terrain surface for that time period, the reservoir water inundation range polygon is determined by comparing the elevation values ​​of each vertex in the triangular mesh terrain surface with the projected upstream water level value. Vertices with elevation values ​​lower than the projected upstream water level value are marked as inundated vertices, and the convex hull boundary of all inundated vertices constitutes the reservoir water inundation range polygon. The reservoir water surface area value for that time period is calculated based on the geometric intersection of the reservoir water inundation range polygon and the triangular mesh terrain surface. Replace the elevation values ​​of all topographic surface vertices within the polygon of the reservoir inundation area with the projected water level in front of the dam for that period to obtain the reservoir surface elevation value for that period. The discharge jet model is a functional component configured to calculate the trajectory of the overflow jet from the spillway based on the discharge flow sequence of each candidate scheduling scheme and the upstream water level of each cascade reservoir. The calculation process is as follows: the overflow weir crest head value is calculated according to the broad-crested weir flow formula, where the discharge flow value equals the flow coefficient multiplied by the weir crest width and then multiplied by the cube of the overflow weir crest head value; the jet velocity value is calculated based on the overflow weir crest head value and the weir crest elevation value in the weir body geometric parameters; the jet angle value is calculated based on the weir surface slope value in the weir body geometric parameters; and the jet trajectory parameters and landing point position parameters are calculated according to the projectile motion trajectory formula based on the jet velocity value, jet angle value, and gravity acceleration parameters.

Images