Efficient solution method for flood control optimization dispatching of complex reservoir groups in river basins

Abstract

The present invention provides an efficient solution method for flood control optimization scheduling of a complex reservoir group in a river basin, including: processing abnormal fluctuations of reservoir inflow based on the VMD-Enhanced CNN model to obtain a denoised inflow process; constructing a reservoir group flood control scheduling model, decomposing the system based on the equivalent projection theory, selecting the outflow as a coordination variable, and establishing a subsystem equivalent model; using the Ray distributed computing framework to parallel process the subsystem equivalent model, executing the EP-DOA-CUDA algorithm to solve, and obtaining the system's global optimal coordination variables; using the global optimal coordination variables as boundary conditions, optimizing the lower-level subsystem based on the IVYA optimization algorithm, and obtaining the optimal scheduling plan for the reservoir group. The present invention solves the problems of abnormal fluctuations of inflow, high model complexity, and low computational efficiency, and improves the accuracy and timeliness of flood control scheduling.

Description

Technical Field

[0001] The invention belongs to the technical field of flood control dispatching, and in particular to an efficient solution method for flood control optimization dispatching of a complex reservoir group in a river basin. Background Art

[0002] Inflow is a key input parameter in reservoir operation. The inflow of most reservoirs cannot be obtained through direct monitoring, and usually needs to rely on the water balance method for reverse calculation. However, due to the superposition of multiple factors such as water level fluctuation errors, uncertainty of water level storage capacity curves, outflow flow measurement errors, and dynamic storage capacity changes, the reversed inflow often shows a sawtooth fluctuation. These abnormal fluctuations can cause dispatchers to misjudge the inflow of the reservoir, which in turn affects the accuracy of dispatch decisions and the effectiveness of reservoir management. Therefore, how to improve the accuracy of reversed inflow and reduce these errors and fluctuations has become an important issue in reservoir operation optimization.

[0003] With the construction and commissioning of a large number of large reservoirs, the reservoir groups of the main and tributary rivers in the basin have gradually formed a large-scale flood control system. The continuous expansion of the scale of reservoirs has led to the continuous increase in the complexity of the joint dispatching model of reservoir group flood control. Despite the rapid development of computing technology, the joint dispatching of reservoir groups in the basin still faces many challenges in the modeling and solution process. Specifically, the flood control dispatching problem of reservoir groups is essentially a multi-constrained, strong-constrained, nonlinear and high-dimensional optimization problem. There is a significant contradiction between the inefficient solution of this problem and the high timeliness requirements of flood control decision-making. Therefore, how to simplify the model complexity and optimize the existing algorithm to improve the computational efficiency by dimensionality reduction methods under the premise of ensuring the dispatching accuracy, so as to adapt to the dynamically changing decision-making needs in the flood control dispatching of reservoir groups, is a key issue that the academic community has long paid attention to, and it is also a hot area and frontier direction in the research of flood control decision-making of reservoir groups.

[0004] To solve the above problems, researchers have explored a variety of mathematical models and calculation strategies. The theoretical methods of reservoir flood control dispatch have gone through three major development stages: conventional dispatch, traditional optimization dispatch and intelligent optimization dispatch. Conventional dispatch methods, such as dispatch charts and dispatch rules, rely on the statistical laws of historical information and are semi-empirical dispatch methods. Traditional optimization methods, such as linear programming, nonlinear programming and dynamic programming, aim to find the global optimal solution to reservoir problems under the complex constraints of the system. With the rapid development of computer technology, intelligent optimization dispatch methods such as particle swarm optimization, pseudo-physics algorithm, simulated annealing algorithm and other heuristic algorithms have gradually been widely used in the field of multi-reservoir dispatch. Although the above research has solved the reservoir dispatch problem to a certain extent, with the increase in the scale of reservoirs, the difficulty of solving the dispatch model increases exponentially. Traditional optimization methods often face the "curse of dimensionality" problem in the solution process, resulting in low computational efficiency. In addition, constrained by the high-dimensional decision space, intelligent optimization methods are prone to fall into local optimality and it is difficult to obtain the global optimal solution. To this end, researchers introduced a coordinated optimization method, which decomposes the complex joint optimization problem into an upper-level problem and multiple lower-level problems, solves each sub-problem independently, and achieves global optimization through information exchange between the upper and lower levels. The coordinated optimization method effectively reduces the dimensionality of the scheduling model and reduces the computational burden, and has been widely used in different reservoir scheduling scenarios. However, traditional coordinated optimization methods are all implemented in an iterative manner, which has certain limitations. Especially when the reservoir is large, the iterative process has shortcomings such as slow convergence, increased communication burden, and poor scalability. Non-iterative methods to achieve coordinated optimization have gradually become the key to solving these problems.

[0005] In order to further improve the modeling and solution rate of complex flood control systems, combining non-iterative coordinated optimization methods with parallel computing strategies is a more effective way. In recent years, parallel computing has been widely introduced and combined with optimization algorithms due to its good search capabilities and fast execution speed to improve the computational efficiency of reservoir optimization scheduling models. Compared with other optimization algorithms, the DOA method has the advantages of smaller parameters, faster search efficiency, and strong global search capabilities. Combining it with parallel computing can greatly speed up the search speed. Therefore, integrating parallel computing with non-iterative coordinated optimization methods to form a two-layer optimization framework can not only improve the solution speed, but also improve the overall optimization accuracy. Summary of the invention

[0006] Purpose of the invention: To provide an efficient solution method for flood control optimization scheduling of a complex reservoir group in a river basin, so as to solve the above-mentioned problems existing in the prior art.

[0007] Technical solution: An efficient solution method for flood control optimization and dispatching of complex reservoirs in a river basin, including:

[0008] S1. Collect basic data and calculate the original flood inflow flow. Process the original flood inflow flow of each reservoir based on the VMD-Enhanced CNN model to obtain the denoised reservoir inflow flow process.

[0009] S2. Construct a flood control dispatch model for reservoir groups, decompose the flood control system into an upper system and multiple lower subsystems, select the reservoir outflow as the coordination variable and the reservoir water storage as the internal variable based on the equivalent projection theory, and construct a subsystem equivalent model;

[0010] S3, using a distributed computing framework to process the equivalent models of the subsystems in parallel, solving each equivalent model based on the EP-DOA-CUDA algorithm, and integrating the optimization results to obtain the global optimal coordination variables of the system;

[0011] S4. Taking the system's global optimal coordination variables as boundary conditions, the lower-level subsystems are optimized based on the IVYA optimization algorithm to obtain the optimal scheduling plan for the reservoir group.

[0012] According to one aspect of the present application, S1 further comprises:

[0013] S11. Collect and study the basic information and historical dispatching process of each reservoir and flood control section in the basin, and calculate the original flood inflow process of each reservoir based on the water balance principle;

[0014] S12, using VMD method to decompose the original reservoir inflow flood flow signal into several intrinsic mode functions;

[0015] S13, using an adaptive sliding window method based on flood flow characteristics to divide the modal signal into time series segments of adaptive length, and construct training samples for the VMD-Enhanced CNN model;

[0016] S14. Construct a VMD-Enhanced CNN model, including a frequency-adaptive convolutional layer, a multimodal input structure, a time-frequency feature fusion layer, a fully connected layer, and a multi-stage learning strategy, and optimize the network weights through the back-propagation algorithm.

[0017] S15, use the trained VMD-Enhanced CNN model to denoise each modality signal;

[0018] S16, adding up the denoised modal signals to obtain the denoised reservoir inflow process.

[0019] According to one aspect of the present application, S12 further includes:

[0020] S12a, taking the original flood flow signal of each reservoir as an input signal;

[0021] S12b, defining the objective function of VMD, and decomposing the input signal into multiple intrinsic mode signals;

[0022] S12c, constraining the objective function through a frequency selection mechanism so that VMD decomposes the intrinsic mode signal into multiple components with different center frequencies and bandwidths;

[0023] S12d. Based on the variational method, a gradient descent algorithm is used to iteratively update the intrinsic mode signal and frequency until convergence, thereby obtaining multiple intrinsic mode functions.

[0024] According to one aspect of the present application, S13 further includes:

[0025] S13a, divide the main stages of the original flood process of each reservoir, including flood stage, flood peak stage and water withdrawal stage;

[0026] S13b, preliminarily setting different sliding window lengths according to the traffic characteristics of each stage;

[0027] S13c, calculating the rate of change and acceleration of traffic flow in adjacent time periods, evaluating traffic fluctuations in real time, and adjusting the window length accordingly;

[0028] S13d, based on the dynamically adjusted window length, the sliding window method is used to segment the traffic signal to obtain the input samples required by the VMD-Enhanced CNN;

[0029] S13e, prepare corresponding labels for each time series segment to ensure the representativeness and diversity of the samples.

[0030] According to one aspect of the present application, S2 further comprises:

[0031] S21. Construct a multi-objective flood control dispatch model for a reservoir group. The objective function includes: minimizing the comprehensive risk of the reservoir group and the comprehensive risk of the protection area group. The constraints include: water balance of a single reservoir, river flood calculation, water balance between upstream and downstream nodes, upper and lower limits of reservoir water storage, reservoir discharge capacity, outflow amplitude and boundary conditions.

[0032] S22, decompose the entire system into upper and lower layers from the spatial dimension, and establish corresponding optimization problems and optimization goals for each lower subsystem;

[0033] S23, dividing the variables in each subsystem into coordination variables and internal variables, selecting the reservoir outflow as the coordination variable and the reservoir water storage as the internal variable;

[0034] S24, introduce the target variable, transform the target function of the lower-level system into an equivalent inequality constraint form, and reconstruct the lower-level subsystem;

[0035] S25. The water storage capacity is expressed as the accumulation of the difference between the inflow and outflow through the water balance equation, the internal variables of the lower subsystem are eliminated, the corresponding equivalent constraints are constructed, and the equivalent model of the subsystem is established.

[0036] According to one aspect of the present application, S21 further includes:

[0037] S21a, define the flood control risk of the reservoir, and calculate the ratio of the maximum storage capacity used by the reservoir to absorb flood water to the available flood control storage capacity;

[0038] S21b, introduce reservoir flood control risk coefficient and flood control contribution weight index, aiming to minimize the comprehensive flood control risk of reservoir groups;

[0039] S21c, define the flood risk of the protection area, and calculate the ratio of the maximum flow of the flood control section to the designed safe flow;

[0040] S21d, introduce flood risk coefficient and flood risk weight index of protection area, aiming at minimizing the comprehensive risk of protection area group;

[0041] S21e. Establish the constraints of the flood control and dispatching model of the reservoir group, including the water balance constraints of a single reservoir, the river flood routing constraints, the water balance constraints between upstream and downstream nodes, the upper and lower limit constraints of the reservoir water storage capacity, the reservoir discharge capacity constraints, the reservoir discharge amplitude constraints and boundary conditions.

[0042] According to one aspect of the present application, S3 further includes:

[0043] S31, using the Ray distributed computing framework to distribute the equivalent models of each subsystem to multiple independent CPU processes;

[0044] S32, extracting basic information of each reservoir and flood control section, and using the reservoir inflow process as input of the subsystem equivalent model, and executing the EP-DOA-CUDA algorithm based on the CUDA architecture and Numba library within each CPU process to obtain the optimization results of each equivalent model;

[0045] S33, after each CPU process independently completes the optimization task of the subsystem equivalent model, the main process collects the optimization results of each equivalent model and integrates them;

[0046] S34. Using the optimization results of the subsystem equivalent model as constraints, combined with global constraints such as river flood routing and water balance between upstream and downstream nodes, the optimal coordination variables of the system in the global scope are calculated.

[0047] According to one aspect of the present application, S32 further includes:

[0048] S32a, initializing dream individuals on the CPU side, including the initial position and "dream" state of each individual;

[0049] S32b, efficiently transferring the initialized dream entity from the CPU memory to the GPU memory;

[0050] S32c, design and implement CUDA kernel function, responsible for fitness evaluation of dream individuals, update of individual historical optimal dream state Pbest, and synchronous calculation of dream generation state;

[0051] S32d, configuring CUDA thread blocks and grid parameters, and dynamically adjusting the grid size according to the total number of dream individuals;

[0052] S32e, comparing the current fitness of each dream individual with its historical best dream state Pbest in parallel on the GPU, and for individuals whose fitness is better than the historical best value, updating their Pbest to the current fitness and corresponding position;

[0053] S32f, transferring the updated dream individual data from the GPU memory back to the CPU memory, and determining the global optimal dream state Gbest in the current iteration on the CPU side;

[0054] S32g, monitoring the iteration process of the DOA algorithm on the CPU side, and determining whether to continue the iteration according to a preset termination condition;

[0055] S32h. After the termination condition is met, the final global optimal dream state Gbest is output as the optimization result of the subsystem equivalent model.

[0056] According to one aspect of the present application, S4 further includes:

[0057] S41, taking the system's optimal coordination variables as boundary conditions, uniformly and randomly generate a population, and each individual is a solution in the problem space;

[0058] S42, calculate the fitness value of each individual, and use the objective function to measure the quality of the individual;

[0059] S43, sort the fitness of individuals in the population and select the individuals with the best fitness as elite individuals;

[0060] S44. Promote the improvement of local solutions through information sharing mechanism between elite individuals and other individuals;

[0061] S45, combining global search and local search strategies, enhancing local search through the guidance of elite individuals, and introducing random perturbations for global exploration;

[0062] S46. In each round of iteration, the position of each individual is updated, and the position of the solution is adjusted according to the guidance of the elite individual and the balance between global and local searches;

[0063] S47, determining whether the termination conditions are met, including reaching the maximum number of iterations, the fitness value change is less than a set threshold, or the fitness reaches a preset target value;

[0064] S48. Output the final optimization result, which is the optimal dispatching plan for the reservoir group.

[0065] According to one aspect of the present application, in S4, the specific implementation process of the IVYA optimization algorithm includes:

[0066] Update individual positions through information sharing mechanism between elite individuals and other individuals: Xnew_i = Xnew_i + α×(Xe - Xnew_i), where α is a parameter that controls the intensity of cooperation; Xnew_i represents the i-th individual in the population; Xe represents the elite individual in the population;

[0067] Combined with the global search strategy: Xglobal_i = X_i + β×(X_r - X_i), where X_r is an individual randomly selected from the population, and β is a parameter that controls the global search step size;

[0068] Combined with the local search strategy: Xlocal_i = Xi + δ×(Xe - Xi), where δ is the control parameter of the local search step size;

[0069] In each round of iteration, elite guidance and random perturbations are considered comprehensively to update individual positions: Xnew_i = Xi + λ×(Xe - Xi) + u×(Xr - Xi), where λ and u control the intensity of elite guidance and global search perturbations, respectively.

[0070] Beneficial effects: The VMD-CNN fusion method proposed in the present invention effectively solves the problem of abnormal "sawtooth" fluctuations in the inflow volume inverted by water balance, improves the accuracy of the inflow volume inverted, and reduces fluctuations and errors caused by water level fluctuation errors, reservoir capacity curve accuracy, and measurement errors. The established multi-objective flood control scheduling model for reservoir groups introduces multi-dimensional risk assessment parameters based on the importance of reservoirs and potential losses of protection areas to accurately determine the weight distribution of reservoirs and protection areas in the comprehensive risk of the system, thereby improving the scientific nature of the flood control scheduling scheme. The non-iterative coordinated optimization method based on equivalent projection theory is adopted to effectively achieve dimensionality reduction of the scheduling model, and solves the problems of high computational complexity and low efficiency of iterative solution of traditional iterative coordinated optimization methods in multi-objective flood control scheduling of reservoir groups, thereby significantly improving the solution efficiency. The EP-DOA-CUDA two-layer heterogeneous parallel optimization strategy is adopted to achieve efficient parallel solution of the scheduling model, thereby further improving the calculation speed and convergence performance. Meeting the dual requirements of accuracy and timeliness for joint flood control dispatching of reservoir groups; IVYA optimization algorithm, with its excellent global search capability and fast convergence characteristics, can effectively solve complex multi-dimensional optimization problems and provide high-precision and efficient optimization solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 Flow chart of the method of the present invention.

[0072] Figure 2 This is a flow chart of S1 of the present invention.

[0073] Figure 3 This is a flow chart of S2 of the present invention.

[0074] Figure 4 This is a flow chart of S3 of the present invention.

[0075] Figure 5 This is a flow chart of S4 of the present invention. DETAILED DESCRIPTION

[0076] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. However, it is obvious to those skilled in the art that the present invention can be implemented without one or more of these details. In other examples, in order to avoid confusion with the present invention, some technical features known in the art are not described. The order of the related in the present invention is not restrictive, that is, it can be adjusted by those skilled in the art. The order in the present invention is a case-based writing method, not a restrictive description.

[0077] The present invention organically integrates reservoir inflow restoration, nonlinear modeling, non-iterative coordinated optimization and parallel computing, and proposes an efficient solution method for flood control optimization scheduling of complex reservoir groups in a river basin, which solves the problem of abnormal "sawtooth" fluctuations in the inflow flow inferred from the water balance, realizes the decomposition and coordination of the complex reservoir group system in a non-iterative manner, organically combines the process-level parallel computing of the CPU with the thread-level parallel computing of the GPU, greatly shortens the calculation time of the model, and provides an effective means to meet the dynamically changing decision-making needs in flood control scheduling of reservoir groups.

[0078] like Figure 1 As shown, the following technical solution is proposed.

[0079] According to one aspect of the present application, a method for efficiently solving flood control optimization scheduling of a complex reservoir group in a river basin is provided, which is characterized by comprising the following:

[0080] S1. Collect basic information and historical dispatching process of each reservoir and flood control section in the study basin, decompose the original reservoir inflow signal into multiple intrinsic mode functions based on VMD method, divide the modal signal into time series segments of adaptive length using adaptive sliding window method based on flood flow characteristics, construct training samples of VMD-Enhanced CNN model, design VMD-Enhanced CNN model, use the trained VMD-Enhanced CNN model to denoise each modal signal, and reconstruct the original signal by addition or inverse transformation to obtain the reservoir inflow process;

[0081] S2. Construct a flood control dispatching model for a reservoir group. Decompose the flood control system into an upper system and multiple lower subsystems from a spatial dimension. Based on the equivalent projection theory, select the reservoir outflow and water storage capacity with actual physical significance as coordination variables and internal variables respectively. Eliminate the internal variables in the subsystem constraints and convert them into equivalent constraints expressed by coordination variables, and then construct a subsystem equivalent model.

[0082] S3, using the Ray distributed computing framework to distribute the equivalent models of each subsystem to multiple CPU processes for parallel processing, and executing the EP-DOA-CUDA algorithm in parallel based on the CUDA architecture and Numba library to solve the equivalent models of each subsystem, collect and integrate the optimization results of each equivalent model, and obtain the global optimal coordination variables of the system;

[0083] S4. The obtained global optimal coordination variables of the system are passed to the lower subsystem as boundary conditions, and the lower subsystem is optimized based on the IVYA optimization algorithm to obtain the optimal scheduling plan for the reservoir group.

[0084] Inflow is a key input parameter in reservoir operation. However, the inflow of most reservoirs cannot be obtained through direct monitoring, and usually needs to rely on the water balance method for reverse calculation. Due to the superposition of multiple factors such as water level fluctuation error, uncertainty of water level storage capacity curve, outflow flow measurement error and dynamic storage capacity change, the reverse calculated inflow often shows a sawtooth fluctuation. These abnormal fluctuations will cause dispatchers to misjudge the reservoir inflow, which will affect the accuracy of scheduling decisions and the effectiveness of reservoir management. Therefore, in order to solve the "sawtooth" fluctuation problem in the reverse calculation of inflow from water balance, improve the reverse calculation accuracy and reduce errors, the VMD and CNN fusion method is an effective solution. VMD combines complex inflow information into a single CNN. The signal is decomposed into multiple intrinsic mode functions, which effectively extracts the local frequency characteristics of the signal and removes high-frequency noise, thereby smoothing the signal and reducing abnormal fluctuations caused by water level fluctuations, uncertainty in reservoir capacity curves and superposition of measurement errors. CNN, with its powerful nonlinear feature extraction capabilities and deep learning structure, can automatically denoise and optimize the modal signal after VMD decomposition, further improving the accuracy and robustness of the signal. The combination of the two overcomes the defects of complex parameter selection and limited denoising effect in the denoising process of traditional methods, and significantly enhances the smoothness and reliability of the inflow inversion results, providing accurate and efficient technical support for reservoir scheduling optimization. The flow chart of the flood inflow denoising method based on the VMD-CNN fusion method is shown in the figure. Figure 2 shown.

[0085] According to one aspect of the present application, S1 is further:

[0086] S11. Determine the study area and collect basic information of each reservoir in the study area, including: dead water level of the reservoir, flood limit water level, normal water storage level, flood control high water level, design flood level, water level storage capacity curve, discharge capacity curve, water level area curve; collect basic information of each section in the study area, including: warning water level, warning flow, guaranteed water level, guaranteed flow, water level flow curve; collect historical dispatching process of each reservoir and section in the study area, including: reservoir water level change process, reservoir water storage change process, reservoir outflow process, section flow process; based on the water balance principle, calculate the original inflow flood process and interval inflow of each reservoir;

[0087] In this embodiment, relevant information can be obtained through public data, aiming to fully understand the basic situation and hydrological characteristics of the study area, thereby providing sufficient data basis and information support for the design and construction of subsequent models.

[0088] S12, using VMD method to decompose the original flood flow signal of each reservoir into multiple intrinsic mode functions with different frequency characteristics;

[0089] S13, the main different stages of the original flood process of each reservoir are divided, and each modal signal is divided into several time series segments of adaptive length using an adaptive sliding window method based on flood flow characteristics to provide training samples for the VMD-Enhanced CNN model;

[0090] S14. Design the VMD-Enhanced CNN model architecture. Use frequency adaptive mechanism in the convolution layer to dynamically adjust the size and weight of the convolution kernel. Introduce a multimodal input structure. Input multiple frequency modal signals obtained by VMD into the convolution layer as independent channels. Use the pooling layer to reduce the dimension and enhance the feature stability. Design the time-frequency feature fusion layer to effectively fuse the time domain and frequency domain features. Use a multi-stage learning strategy in the fully connected layer to gradually optimize the denoising and traffic prediction tasks. Design a multi-task loss function to take into account both denoising and prediction tasks. Optimize the network weights through the back-propagation algorithm to obtain a trained VMD-Enhanced CNN model.

[0091] S15, using the trained VMD-Enhanced CNN model to denoise each modal signal to obtain a denoised modal signal;

[0092] In this embodiment, VMD-Enhanced CNN uses 32 convolution kernels with a size of 3×1 through pre-learned convolution kernels to extract local features of the signal in the convolution layer. The pooling layer uses a 2×1 pooling window with a step size of 2 to reduce the dimension of the features to remove redundant information and reduce computational complexity. The fully connected layer integrates the extracted features, contains 128 neurons, and generates the final denoising result. For each input time series segment, VMD-Enhanced CNN outputs a corresponding denoised segment. After splicing all denoised segments, a denoised complete modal signal is generated. This method effectively retains the local features and long-term trends of the modal signal, while removing high-frequency interference introduced by measurement errors or environmental noise, obtaining a smoother and more accurate modal signal, and providing reliable basic data for subsequent signal reconstruction.

[0093] S16. Add up the denoised modal signals to restore their time domain representation. By adding up all the denoised modal signals, a complete reconstructed signal is obtained to restore a smooth and denoised reservoir inflow process.

[0094] In this embodiment, the values ​​of the three modal signals after denoising at each moment are IMF1(t), IMF2(t) and IMF3(t), respectively. By adding these values ​​one by one, the reconstructed signal R(t) = IMF1(t) + IMF2(t) + IMF3(t) is obtained, thereby restoring the smooth trend and local dynamic characteristics of the original signal. This addition process effectively removes noise and restores a more accurate reservoir inflow process, providing more reliable basic data for subsequent analysis.

[0095] According to one aspect of the present application, S12 is further:

[0096] S12a, the original flood flow signal of each reservoir is used as the input signal;

[0097] S12b, defining the objective function of VMD, and decomposing the input signal into multiple intrinsic mode signals;

[0098] S12c, introducing a frequency selection mechanism to constrain the objective function, VMD decomposes the intrinsic mode signal into multiple components with different center frequencies and bandwidths;

[0099] S12d. Based on the variational method, a gradient descent algorithm is used to iteratively update the intrinsic mode signal and frequency. In each iteration, the frequency distribution of the intrinsic mode signal is adjusted, and the objective function is gradually optimized until convergence to obtain multiple intrinsic mode functions.

[0100] In this embodiment, the initial inflow sequence of the reservoir is calculated by the water balance method. The sequence contains 168 sampling points with a time interval of 1 hour, corresponding to a 7-day flood process. The original flood sequence is used as the input signal, and the VMD algorithm is applied. The number of decomposition modes is set to 3, the bandwidth parameter is set to 200, and the maximum number of iterations is set to 500. In the process of optimizing the objective function, VMD decomposes the original signal into three intrinsic mode functions with different frequency characteristics. The first mode function extracts the high-frequency signal part with a frequency range of 0.3-0.5 Hz, mainly removing the high-frequency interference caused by measurement errors or equipment noise. The second mode function is concentrated in 0.1-0.3 Hz, reflecting the local oscillation characteristics in the flood process, such as flow changes caused by short-term rainfall. The third mode function captures low-frequency components below 0.1 Hz, showing the overall trend and long-term change law of the flood sequence. The decomposed mode signals respectively show smooth characteristics, remove high-frequency noise, and significantly improve the interpretability of the flood signal, providing more reliable basic data for subsequent denoising and CNN model input.

[0101] In this embodiment, the VMD method is used to decompose the reservoir inflow signal. Compared with traditional methods such as empirical mode decomposition, wavelet decomposition and sliding average method, VMD shows significant advantages in its accuracy of time-frequency decomposition, noise resistance and adaptive characteristics to non-stationary signals. It avoids mode aliasing and endpoint effect problems through variational optimization, and has strong robustness to signal noise, making the decomposition result more stable and reliable. Therefore, VMD has unique advantages in complex hydrological signal processing, especially in solving the "sawtooth" abnormal fluctuation problem in inflow back-calculation, and is an ideal tool for improving signal analysis accuracy and robustness.

[0102] According to one aspect of the present application, S13 is further:

[0103] S13a, divide the main stages of the original flood process of each reservoir, including flood stage, flood peak stage and water withdrawal stage;

[0104] S13b, according to the flow characteristics of each stage, preliminarily set different sliding window lengths, use a smaller window in the flood stage, increase the window in the flood peak stage to capture rapid changes, and appropriately extend the window in the receding stage to avoid redundant information;

[0105] S13c, calculate the change rate and acceleration of the flow in adjacent time periods, evaluate the flow fluctuation in real time, and adjust the window length accordingly. When the flow changes drastically, increase the window length, and when the change is stable, reduce the window length;

[0106] S13d, based on the dynamically adjusted window length, the sliding window method is used to segment the traffic signal, intercept and organize the time series fragments, and obtain the input samples required by the VMD-Enhanced CNN;

[0107] S13e. Prepare corresponding labels for each time series segment, check the balance of the data, and ensure the representativeness and diversity of the samples.

[0108] In this embodiment, for each modal signal, the reservoir inflow signal is divided into different flood stages, and an adaptive sliding window segmentation method is used to dynamically adjust the window length for different flood stages. First, according to the flow change trend, the flood process is divided into a flood stage, a flood peak stage and a water retreat stage, and an initial window length is set for each stage: the flood stage is set to 2 hours, the flood peak stage is set to 4 hours, and the water retreat stage is set to 3 hours. Subsequently, in each stage, the flow change rate and acceleration are calculated to monitor the severity of the flow fluctuation in real time, and the window length is dynamically adjusted. In the flood stage and water retreat stage where the flow changes are relatively stable, the window length is appropriately reduced, and in the flood peak stage where the flow fluctuation is large, the window length is automatically increased to ensure that the rapidly changing flow information can be fully captured. The step size is uniformly set to 1 hour, and the window slides for 1 hour each time to capture a new time segment. Through this dynamic adjustment mechanism, the window length of each time period is guaranteed to match the flow change characteristics.

[0109] In this embodiment, an adaptive sliding window segmentation method based on flood flow characteristics is used to segment time series segments. Different from the traditional fixed window length method, this method adjusts the length of the sliding window in real time according to the changing characteristics of the reservoir inflow, combined with the flow change rate and acceleration, and adaptively changes the window length by dynamically monitoring the severity of flow fluctuations in different flood stages. This dynamic adjustment mechanism enables the segmentation window of each stage to more accurately capture the local characteristics of the flow at that stage, avoiding information loss or redundancy caused by fixed window length in traditional methods, and can more flexibly adapt to changes in different flood flows, providing more accurate training samples.

[0110] According to one aspect of the present application, S14 is further:

[0111] S14a, construct a frequency adaptive convolution layer, adopt a frequency adaptive mechanism, dynamically adjust the size and weight of the convolution kernel, and each convolution layer uses multiple convolution kernels to perform convolution operations on time series segments;

[0112] S14b, defining a multimodal input structure, using a multi-channel input structure to input signals of different frequency components into the convolution layer in parallel, and generating a set of feature maps;

[0113] S14c, applying a pooling layer, using a maximum pooling method or an average pooling method to reduce the dimension of the feature map set, to obtain a feature map set after dimension reduction;

[0114] S14d, introducing a time-frequency feature fusion layer to effectively fuse the time domain and frequency domain features extracted by the convolution and pooling layers, and obtain a feature map set that fuses the feature information of the time domain and the frequency domain;

[0115] S14e, execute the fully connected layer, adopt a multi-stage learning strategy, gradually flatten the reduced feature map set into a one-dimensional vector, and then integrate the information through multiple fully connected layers;

[0116] S14f, design a multi-task loss function to optimize both the denoising task and the traffic prediction task, and balance the two through a weighted loss function;

[0117] S14g. Calculate the gradient through the back-propagation algorithm, use the optimization algorithm to update the network weights, and continue training until the loss function converges.

[0118] In this embodiment, after the sliding window segmentation is completed, the obtained time series fragments are used as the input training samples of VMD-EnhancedCNN. When constructing the CNN architecture, the three modal signals IMF1, IMF2, and IMF3 obtained by VMD decomposition are first input into the convolution layer as independent channels using a multi-channel input structure. The convolution layer uses a frequency adaptive mechanism to dynamically adjust the size and weight of the convolution kernel according to the frequency characteristics of the signal to adapt to different frequency components. IMF1 corresponds to high-frequency components and has a smaller convolution kernel; IMF3 corresponds to low-frequency components and has a larger convolution kernel. Each modal signal generates a feature map through convolution operation. Then, the pooling layer uses a 2×1 maximum pooling window to reduce the dimension of the feature map to reduce the amount of calculation and remove redundant information. Then, the time-frequency feature fusion layer fuses the time domain and frequency domain features extracted by the convolution and pooling layers to obtain a comprehensive feature map. After feature fusion, the feature map is flattened into a one-dimensional vector and input into the fully connected layer. The fully connected layer contains 128 neurons. The denoising and traffic prediction tasks are gradually optimized through multi-stage learning. The loss function uses a weighted multi-task loss function to optimize the denoising effect and prediction accuracy.

[0119] In this embodiment, a VMD-Enhanced CNN model is used to process complex signals. VMD-Enhanced CNN can capture nonlinear patterns and time series dependencies in signals while automatically extracting local features of signals through its deep convolutional layer. Especially in the case of multimodal signal input, VMD-Enhanced CNN can effectively learn and fuse the relationship between different frequency components. In addition, when processing modal signals after VMD decomposition, VMD-Enhanced CNN uses a frequency adaptive convolutional layer to dynamically adjust the size and weight of the convolution kernel, so that the network can adaptively process high-frequency and low-frequency components, thereby improving the denoising effect and prediction accuracy. Compared with traditional statistical methods and shallow machine learning algorithms, VMD-Enhanced CNN not only gets rid of the dependence on manual feature selection, but also greatly improves the denoising and reconstruction capabilities of signals, solves the limitations of traditional methods in complex time series signal processing, and provides more efficient and accurate solutions for fields such as reservoir inflow inversion.

[0120] In summary, this embodiment effectively solves the problem of "sawtooth" abnormal fluctuation of reservoir inflow in water balance reverse prediction by using VMD-CNN fusion method, and significantly improves the accuracy of reverse prediction of inflow. Traditional methods are easily affected by factors such as water level fluctuation error, reservoir capacity curve accuracy and measurement error, resulting in unstable and inaccurate prediction results. VMD accurately decomposes the original signal into different frequency components, making each modal signal purer and reducing noise interference. Subsequently, CNN further learns and fuses the temporal dependence and implicit features of each modal signal through its powerful feature automatic extraction ability and nonlinear mapping ability, thereby effectively improving the signal denoising effect and flow prediction accuracy. The combination of VMD and CNN avoids the dependence on artificial feature selection in traditional methods, and at the same time improves the robustness of the model, which can better cope with complex nonlinear and time-varying signals, not only improving the accuracy of inflow prediction, reducing fluctuations caused by factors such as measurement errors, but also providing accurate and efficient technical support for reservoir scheduling optimization, and providing more reliable solutions for practical applications in related fields.

[0121] In the process of flood control and dispatching of reservoir groups, two core goals must be comprehensively considered: the safety of the reservoir itself and the safety of the downstream protection area. Due to the high complexity of the multi-objective flood control optimization and dispatching problem of reservoir groups, it is difficult to directly solve it using a single mathematical method. For this reason, researchers have introduced a coordinated optimization method, which decomposes the global system into an upper system and a series of lower subsystems. Each subsystem performs optimization calculations independently and gradually approaches the global optimal solution through information exchange between the upper and lower systems. Lagrange duality theory is one of the most commonly used coordinated optimization methods in flood control and dispatching of reservoir groups. It decomposes the global system into multiple subsystems and introduces Lagrange multipliers to coordinate the solutions of each subsystem to gradually approach the global optimal solution. However, this method requires frequent information exchange between the upper and lower systems, and gradually adjusts the solutions of each subsystem through iterative updates of Lagrange multipliers. This iterative solution method has certain limitations, especially when the reservoir is large in scale. In the iterative process, the shortcomings of slow convergence, increased communication burden and poor scalability are apparent. Achieving coordinated optimization in a non-iterative manner has gradually become the key to solving these problems. The process of the non-iterative coordinated optimization method based on equivalent projection theory is as follows: Figure 3 shown.

[0122] According to one aspect of the present application, S2 is further:

[0123] S21. Construct a multi-objective flood control dispatch model for a reservoir group. The objective function includes: minimizing the comprehensive risk of the reservoir group and the comprehensive risk of the protection area group. The constraints include: water balance constraints of a single reservoir, river flood routing constraints, water balance constraints between upstream and downstream nodes, upper and lower limit constraints of reservoir water storage, reservoir discharge capacity constraints, outflow amplitude constraints and boundary conditions.

[0124] S22. Decompose the entire system into upper and lower layers from the spatial dimension, and establish corresponding optimization problems and optimization goals for each lower subsystem. The mathematical expression of the model is simplified as follows:

[0125] min F u (x [R] , y u ) +∑F l r (x r , y l r )

[0126] s C u (x [R] , y u )≤0

[0127] C l r (x r , y lr ) ≤0

[0128] Where: u and l represent the upper system and lower system respectively; R is the number of subsystems, r∈R, [R]:={1,2,…,R},x [R] :={x1,x2…,x R}; F u (.) and F l (.) represent the objective functions of the upper system and the lower system respectively, C u (.) and C l r (.) represent the constraints of the upper system and the lower system respectively.

[0129] S23, the variables in each subsystem are divided into coordination variables and internal variables, the reservoir outflow with actual physical significance is selected as the coordination variable x, and the reservoir water storage capacity is selected as the internal variable y. The upper system is responsible for calculating and transmitting the coordination variables, but cannot directly control the internal variables of the lower subsystem. The lower subsystem will use the optimal coordination variables calculated by the upper system as boundary conditions to further optimize the values ​​of its internal variables;

[0130] S24. In order to ensure that the internal variables exist only in the constraints of the lower system, the target variable T is introduced. r , transform the objective function of the lower-level system into an equivalent inequality constraint form, and reconstruct the lower-level subsystem. Through this transformation, the original model can be restated as follows:

[0131] min F u (x [R] , y u ) +∑T r

[0132] s C u (x [R] , y u )≤0

[0133] C l r (x r , y l r ) ≤0

[0134] F l r (x r , y l r ) ≤T r ≤T

[0135] Where: T ris the target variable of the lower system r, reflecting the subsystem's response to the given coordination variable x r The maximum target value that can be achieved under the condition of F; T is a constant and can be greater than F l r (x r , y l r )

[0136] S25. In order to eliminate the internal variables of the lower subsystem, the water storage capacity is expressed as the accumulation of the difference between the inflow and outflow through the water balance equation, and then the water storage capacity is converted into a function form of the inflow and outflow, so that the change of storage capacity is directly related to the inflow and outflow. The water balance constraints of a single reservoir, the upper and lower limit constraints of the reservoir water storage capacity, the reservoir discharge capacity constraints and boundary conditions are processed, and the corresponding equivalent constraints are constructed to establish the equivalent model of the subsystem:

[0137] 1) Water balance equivalent constraint

[0138] V i,t+1 = V i,1 +[∑ t m=2 (Q i,m -qc i,m )+((Q i,1 -qc i,1 )+(Q i,t+1 -qc i,t+1 )) / 2]×Δt

[0139] Where: Q i,m and qc i,m are the inflow and outflow of the reservoir in the mth period (m 3 / s).

[0140] 2) Equivalence constraints on upper and lower limits of reservoir water storage

[0141] V i,min -V i,1 ≤[∑ t-1 m=2 (Q i,m -qc i,m )+((Q i,1 -qc i,1 )+(Q i,t+1 -qc i,t+1 )) / 2]×Δt≤V i,max -V i,1

[0142] 3) Equivalent constraints on reservoir discharge capacity

[0143] 0≤qc i (t) ≤fv~q {[∑ t-1 m=2 (Q i,m -qc i,m )+ ((Q i,1 -qc i,1 )+(Q i,t+1 -qc i,t+1 )) / 2] ×Δt}

[0144] 4) Boundary condition equivalent constraints

[0145] V i,T+1 = V i,obj

[0146] V i,T+1 = V i,1 +[∑ T m=2 (Q i,m -qc i,m )+((Q i,1 -qc i,1 )+(Q i,T+1 -qc i,T+1 )) / 2]×Δt

[0147] According to one aspect of the present application, S21 is further:

[0148] S21a, define the reservoir flood risk level S R,i , which means the ratio of the maximum storage capacity used by the reservoir to absorb floods during the dispatch period to the available flood control storage capacity. The calculation formula is as follows:

[0149] V i,Available = V i,Total – V i,Current

[0150] S R,i =max{V i (t)} / V i,Available， max{V i (t)}< V i,Available

[0151] S R,i =1+exp((max{V i (t)}- V i,Available ) / V i,Available ) max{V i (t)}> V i,Available

[0152] Where: S R,i is the flood risk of the i-th reservoir; max{V i(t)} is the maximum storage capacity of the i-th reservoir used to absorb flood water during the dispatch period (m³); V i,Available is the available flood control storage capacity of the i-th reservoir; V i,Total is the designed flood control storage capacity of the i-th reservoir (m³), V i,Current is the initial storage capacity of the i-th reservoir (m³), T is the dispatch period (h);

[0153] S21b. Due to the different storage capacity and water inflow conditions of reservoirs, different reservoirs face significant differences in flood control pressure during flood control scheduling. In order to more accurately evaluate the ability of reservoirs to cope with floods and their importance, the reservoir flood control risk coefficient δ and flood control contribution weight index α are introduced:

[0154] δ i = (W i +P i ×F i ) / V i,Available

[0155] α i = (δ i / ∑ n i=1 δ i ) ×100%

[0156] Where: i W is the reservoir flood risk factor, which comprehensively reflects the impact of reservoir inflow floods, future rainfall and idle storage capacity, and measures the reservoir's flood absorption capacity and potential pressure; i is the inflow of water into the i-th reservoir (m 3 );P i is the subsequent rainfall of the i-th reservoir (mm); F i is the controlled area of ​​the i-th reservoir (km 2 ); α i is the flood risk weight index of the i-th reservoir.

[0157] The goal is to minimize the comprehensive flood control risk of the reservoir group, and its expression is as follows:

[0158] S * R,i =α i ×S R,i

[0159] minF1=∑ n i=1 S * R,i

[0160] Where: S * R,i is the comprehensive flood control risk of the i-th reservoir.

[0161] S21c, define the flood risk of the protection area as S F,i , which means the ratio of the maximum flow of the flood control section to the designed safe flow during the dispatch period. The calculation formula is as follows:

[0162] Q i,max = max[∑ k∈ΩR,i qc' k,t +Q k→i,t ]

[0163] S F,i =Q i,max / Q i,s , Q i,max i,s

[0164] S F,i =1+exp((Q i,max -Q i,s ) / Q i,s ), Q i,max >Q i,s

[0165] Where: S F,i is the flood risk of the ith protection zone; Ω R,i is the set of upstream reservoirs that have direct hydraulic connection with protection area i; qc' k,t is the response of the outflow of the kth reservoir in the downstream protection area (m 3 / s); Q k→i,t is the interval flow from the reservoir dam site to protection zone i that has direct hydraulic connection with protection zone i (m 3 / s); Q i,max is the maximum flow rate of protection zone i during the scheduling period; Q i,s is the design safety flow of flood control area i (m 3 / s).

[0166] S21d. Since the flooding losses faced by different protection areas are significantly different, in order to more accurately reflect the importance of the protection area, the flood risk coefficient λ and flood risk weight index β of the protection area are introduced:

[0167] λ i =ω1×Pop i / ∑ m i=1 Pop i +ω2×E i / ∑ m i=1 E i+ ω3×C i / ∑ m i=1 ​C i

[0168] β i =(λ i / ∑ m i=1 λ i ) ×100%

[0169] Where: i is the flood risk coefficient of the protection zone, which comprehensively reflects the potential impact and degree of impact on the population, economy and ecosystem of different protection zones when facing flood events; ω is the weight parameter of the flood risk coefficient, which is used to reflect the importance of different influencing factors; Pop i is the population density of the ith protection zone (people / km²); E i is the GDP density of the ith protection zone (CNY / km²); C i is the ecological service value density of the i-th protection zone (CNY / km²); β i It is the flood risk weight index of the protection area.

[0170] The goal is to minimize the comprehensive risk of the protection zone group, and its expression is as follows:

[0171] S * F,i =α i ×S F,i

[0172] minF2=∑ m i=1 S * F,i

[0173] Where: S * F,i It is the comprehensive flood risk level of the protection area group.

[0174] S21e, the scheduling model considers the following equality and inequality constraints:

[0175] (1) Water balance constraints of a single reservoir:

[0176] V i,t+1 = V i,t + ((Q i,t+1 + Q i,t ) / 2- (qc i,t+1 +qc i,t ) / 2) ×Δt, t=1 to T

[0177] Where: V i,t and V i,t+1 are the water storage of reservoir i at the beginning and end of period t (m3 );Q i,t and Q i,t+1 are the inflows at the beginning and end of period t of the i-th reservoir (m 3 / s);qc i,t and qc i,t+1 are the outflows at the beginning and end of period t of the i-th reservoir (m 3 / s); Δt is the scheduling step length (h).

[0178] (2) River flood routing constraints:

[0179] qc' k,t =C0×qc k,t+1 + C1×qc k,+ C2×qc' k,t

[0180] Where: C0, C1, C2 are the parameters of the Muskingum River flood routing model, and C0+C1+C2=1.

[0181] (3) Water balance constraints between upstream and downstream nodes:

[0182] Q i,t =∑ k∈ΩR,i qc' k,t +Q k→i,t

[0183] (4) Upper and lower limits of reservoir water storage:

[0184] V i,min ≤ V i,t ≤ V i,max

[0185] Where: V i,min and V i,max Divided into the lower and upper limits of the storage capacity of the i-th reservoir (m 3 ); V i,t is the water storage capacity of the i-th reservoir at time t (m 3 ).

[0186] (5) Constraints on reservoir discharge capacity:

[0187] 0 ≤ qc i,t ≤ f i,v~q (V i,t )

[0188] Where: f i,v~q (.) is the discharge capacity curve of the i-th reservoir.

[0189] (6) Constraints on the fluctuation range of reservoir discharge during adjacent periods:

[0190] ∣qci,t+1 -qc i,t ∣≤Δq i

[0191] Where: |qc i,t+1 -qc i,t ∣ is the fluctuation range of outflow from the i-th reservoir in adjacent periods (m 3 / s);Δq i is the maximum outflow variation allowed for the i-th reservoir in adjacent periods (m 3 / s).

[0192] (7) Boundary conditions for reservoir operation:

[0193] V i,1 = f i,z~v (Z i,1 ),V i,T+1= V i,obj

[0194] Where: f i,z~v ( . ) is the water level and storage capacity curve of the i-th reservoir; V i,1 and V i,T+1 are the storage capacity at the beginning and end of the i-th reservoir operation period (m 3 ); Z i,1 is the initial water level of the i-th reservoir (m); V i,obj is the target storage capacity of the i-th reservoir at the end of the period (m 3 ).

[0195] In order to further improve the modeling and solution rate of complex flood control systems, combining non-iterative coordinated optimization methods with parallel computing strategies is a more effective way. In recent years, parallel computing has been widely introduced and combined with optimization algorithms due to its good search capabilities and fast execution speed to improve the computational efficiency of reservoir optimization scheduling models. Compared with other optimization algorithms, the DOA method has the advantages of smaller parameters, faster search efficiency, and strong global search capabilities. Combining it with parallel computing can greatly speed up the search speed. Therefore, integrating parallel computing with non-iterative coordinated optimization methods to form a two-layer optimization framework can not only improve the solution speed, but also improve the overall optimization accuracy. The EP-DOA-CUDA two-layer heterogeneous parallel optimization solution strategy is as follows: Figure 4 shown.

[0196] According to one aspect of the present application, S3 is further:

[0197] S31. Based on the “divide and conquer” strategy, the Ray distributed computing framework is used to distribute the equivalent models of each subsystem to multiple independent CPU processes;

[0198] S32, extracting basic information of each reservoir and flood control section, and using it together with the calculated reservoir inflow process as the input of the subsystem equivalent model, and executing the EP-DOA-CUDA algorithm based on the CUDA architecture and Numba library within each CPU process to obtain the optimization results of each equivalent model;

[0199] S33, after each CPU process independently completes the optimization task of the subsystem equivalent model, the main process collects the optimization results of each equivalent model and integrates them;

[0200] S34. Using the optimization results of the subsystem equivalent model as constraints, combined with global constraints such as river flood routing and water balance between upstream and downstream nodes, the optimal coordination variables of the system in the global scope are calculated.

[0201] According to one aspect of the present application, S32 is further:

[0202] S32a, initializing dream individuals on the CPU side, including the initial position and "dream" state of each individual;

[0203] S32b, efficiently transferring the initialized dream entity from the CPU memory to the GPU memory;

[0204] S32c, design and implement CUDA kernel function, responsible for fitness evaluation of dream individuals, update of individual historical optimal dream state Pbest, and synchronous calculation of dream generation state;

[0205] S32d, configuring CUDA thread blocks and grid parameters, and dynamically adjusting the grid size according to the total number of dream individuals;

[0206] S32e, comparing the current fitness of each dream individual with its historical best dream state Pbest in parallel on the GPU, and for individuals whose fitness is better than the historical best value, updating their Pbest to the current fitness and corresponding position, and processing the individual's Pbest update in parallel through multi-threading;

[0207] S32f, transferring the updated dream individual data from the GPU memory back to the CPU memory, comparing the fitness values ​​of all individuals on the CPU side, and determining the global optimal dream state Gbest in the current iteration;

[0208] S32g, monitor the iteration process of the DOA algorithm on the CPU side, and determine whether to continue the iteration according to the preset termination condition. If the termination condition is not met, return to S33b to S33f to perform the next round of dream generation and optimization iteration, and gradually approach the global optimal dream state;

[0209] S32h. After the termination condition is met, the final global optimal dream state Gbest is output as the optimization result of the subsystem equivalent model.

[0210] In this embodiment, in the GPU acceleration process of the EP-DOA-CUDA algorithm, the dream individuals are first initialized on the CPU side, with an individual scale of 1000. The initial position and dream state of each individual are randomly generated, the position dimension is 30, the initialization range of the dream state is [-1,1], the dream state factor is [0.05, 0.3], and the maximum number of iterations is 1000. Then, the initialized dream individual data is transferred from the CPU memory to the GPU memory using an efficient data transmission mechanism to ensure that the data layout meets the GPU parallel computing requirements. On the GPU side, by designing and implementing CUDA kernel functions and combining Numba's just-in-time compilation function, individual fitness evaluation, historical optimal dream state update, and individual state synchronization calculation are performed. The individual fitness evaluation and dream state update are calculated in parallel by CUDA. The number of individuals is 1000, and each thread block is set to contain 256 threads. The calculation grid size is 4 thread blocks to ensure that all individuals can be covered. Subsequently, the GPU processes the fitness of each individual in parallel. The fitness of the current dream is compared with the historical optimal dream fitness. If the current fitness is better than the historical optimal, the historical dream state is updated. The updated individual data is then returned to the CPU through CUDA memory transmission. The global optimal dream state Gbest is evaluated on the CPU side, and the iterative process of the algorithm is monitored. Whether to continue the iteration is determined based on the termination condition. If the termination condition is not met, the algorithm will return to the update to continue execution. If the termination condition is met, the final global optimal dream state is output as the optimization result of the subsystem equivalent model. The whole process makes full use of the parallel computing capability of the GPU. Through efficient memory transmission and computing scheduling, the computing speed and accuracy are greatly improved. Experiments show that the dream optimization algorithm has good results.

[0211] IVYA successfully achieves an effective balance between global search and local search through the competition and cooperation mechanism between elite individuals, thus avoiding falling into the local optimal solution in the solution space. IVYA can accelerate the convergence process through information sharing and elite guidance, while improving the quality of the solution. It is especially suitable for multi-objective optimization problems. In multi-objective scenarios, IVYA can effectively balance between different objectives to ensure that the algorithm finds the Pareto optimal solution. Its strong adaptability and efficient search strategy enable IVYA to handle high-dimensional and complex constrained optimization problems with high computational efficiency and accuracy. Therefore, it has shown significant advantages in fields such as resource scheduling and engineering design. IVYA optimization algorithm solves problems such as Figure 5 shown.

[0212] According to one aspect of the present application, S4 is further:

[0213] S41, randomly generate a population P={X1, X2,…X N}, each individual X 1= {x i1 ,x i2 ,…x im} is an m-dimensional decision vector, representing a solution in the problem space, with the system's optimal coordination variables as boundary conditions, and a uniformly randomly generated population to ensure a wide coverage of the solution space;

[0214] S42. Calculate each individual X i The fitness value f (X i ), using the objective function f (X i ) to measure the merits and demerits of individuals;

[0215] S43, by sorting the fitness of the individuals in the population, select the individuals with the best fitness as elites. The elite individuals have a dominant position in the next iteration and are responsible for guiding the search direction of other individuals. By comparing the fitness value f(X i ), select the optimal solution X e :

[0216] X e =argmin f (X i )

[0217] Where: X e It is the individual with the smallest fitness value in the population

[0218] S44. Through the information sharing mechanism between elite individuals and other individuals, the improvement of local solutions is promoted. The elite individuals and other poor individuals X e To collaborate between them, information sharing can be updated by the following formula:

[0219] Xnew i= X i +α×(X e -X i )

[0220] Where: α is a parameter that controls the intensity of cooperation and determines the degree of influence of elite individuals on other individuals.

[0221] S45. Combining global search and local search strategies, local search is enhanced through the guidance of elite individuals, and random perturbations are introduced for global exploration. Global search can be performed using the following formula:

[0222] Xglobal i= X i +β×(X r -X i )

[0223] Where: X r is an individual randomly selected from the population, and β is a parameter that controls the global search step size.

[0224] Local search updates the current solution through the guidance of the elite individual position:

[0225] Xlocal i= X i +δ×(X e -X i )

[0226] Where: δ is the control parameter of the local search step size.

[0227] S46. In each round of iteration, the position of each individual is updated, and the position of the solution is adjusted according to the guidance of the elite individual and the balance of global and local search. The update process comprehensively considers the guidance of the elite individual and random perturbations:

[0228] Xnew i= X i +λ×(X e -X i )+ u×(X r -X i )

[0229] Where: λ and u control the strength of elite guidance and global search perturbations, respectively, to ensure that the population conducts appropriate search between global and local.

[0230] S47: Determine whether the termination condition is met. Common termination conditions include reaching the maximum number of iterations T. max , the fitness value change is less than the set threshold ε or the fitness reaches the preset target value, the termination condition can be judged by the following formula:

[0231] If max(|f (X i )-f (X e )∣) <ε or t = T max , stop

[0232] Where: t is the current number of iterations, ε is the allowable error range.

[0233] S48. Output the final optimization result.

[0234] In this embodiment, a population is first generated randomly. Each individual is a vector containing 168×n decision variables, where n is the number of reservoirs, representing a solution in the problem space. The initial generation range of the population is determined by the optimal coordination variable of the system, and a uniform random distribution is used to ensure a wide coverage of the solution space. The population size is set to 1000 individuals. Subsequently, the fitness value of each individual is calculated. The fitness value evaluates the individual through the objective function to measure the quality of the individual, and the population is sorted according to the fitness. The individual with the best fitness is selected as the elite individual. The elite individual dominates the search direction in subsequent iterations, and shares information with other poor individuals to cooperate in improving the local solution. The strength of the information sharing mechanism is controlled by the parameter α. Set The α value is set to 0.5. To promote global exploration, the algorithm combines local search with global search strategies. It guides local search through elite individuals and introduces random perturbations for global search. The local search step size and global search step size are controlled by parameters γ and β, respectively. γ is set to 0.3 and β is set to 0.8. In each round of iteration, the individual updates its position according to the guidance of the elite individuals and the global perturbation. The balance between global and local search during the update process is adjusted by parameter δ (δ is set to 0.7). The iterative process continues until the preset termination condition is met, such as reaching the maximum number of iterations of 1000, the fitness value change is less than the set threshold 1e-6, or the fitness value reaches the target value 1e-3. Finally, the optimal solution is output as the optimization result of the problem.

Claims

1. An efficient solution method for flood control optimization scheduling of complex reservoir groups in a river basin, characterized by: include: S1. Collect basic data and calculate the original flood inflow flow. Process the original flood inflow flow of each reservoir based on the VMD-Enhanced CNN model to obtain the denoised reservoir inflow flow process. S2. Construct a flood control dispatch model for a reservoir group, decompose the flood control system into an upper system and multiple lower subsystems, select the reservoir outflow as the coordination variable and the reservoir water storage as the internal variable based on the equivalent projection theory, and construct a subsystem equivalent model; S3, using a distributed computing framework to process the equivalent models of the subsystems in parallel, based on the EP-DOA-CUDA algorithm, taking the reservoir inflow process as input, solving the equivalent models of each subsystem, integrating the optimization results to obtain the global optimal coordination variables of the system; S4, taking the system's global optimal coordination variables as boundary conditions, optimizing the lower subsystem based on the IVYA optimization algorithm, and obtaining the optimal dispatching plan for the reservoir group; S1 further includes: S11. Collect and study the basic information and historical dispatching process of each reservoir and flood control section in the basin, and calculate the original flood inflow process of each reservoir based on the water balance principle; S12, using VMD method to decompose the original reservoir inflow flood flow signal into several intrinsic mode functions; S13, using an adaptive sliding window method based on flood flow characteristics to divide the modal signal into time series segments of adaptive length, and construct training samples for the VMD-Enhanced CNN model; S14. Construct a VMD-Enhanced CNN model, including a frequency-adaptive convolutional layer, a multimodal input structure, a time-frequency feature fusion layer, a fully connected layer, and a multi-stage learning strategy, and optimize the network weights through the back-propagation algorithm. S15, use the trained VMD-Enhanced CNN model to denoise each modality signal; S16, adding up the denoised modal signals to obtain a denoised reservoir inflow process; S2 further includes: S21. Construct a multi-objective flood control dispatch model for a reservoir group. The objective function includes: minimizing the comprehensive risk of the reservoir group and the comprehensive risk of the protection area group. The constraints include: water balance of a single reservoir, river flood calculation, water balance between upstream and downstream nodes, upper and lower limits of reservoir water storage, reservoir discharge capacity, outflow amplitude and boundary conditions. S22, decompose the entire system into upper and lower layers from the spatial dimension, and establish corresponding optimization problems and optimization goals for each lower subsystem; S23, dividing the variables in each subsystem into coordination variables and internal variables, selecting the reservoir outflow as the coordination variable and the reservoir water storage as the internal variable; S24, introduce the target variable, transform the target function of the lower-level system into an equivalent inequality constraint form, and reconstruct the lower-level subsystem; S25. The water storage capacity is expressed as the accumulation of the difference between the inflow and outflow through the water balance equation, the internal variables of the lower subsystem are eliminated, the corresponding equivalent constraints are constructed, and the equivalent model of the subsystem is established.

2. The efficient solution method for flood control optimization scheduling of complex reservoir groups in a river basin according to claim 1 is characterized in that: S12 further includes: S12a, taking the original flood flow signal of each reservoir as an input signal; S12b, defining the objective function of VMD, and decomposing the input signal into multiple intrinsic mode signals; S12c, constraining the objective function through a frequency selection mechanism so that VMD decomposes the intrinsic mode signal into multiple components with different center frequencies and bandwidths; S12d. Based on the variational method, a gradient descent algorithm is used to iteratively update the intrinsic mode signal and frequency until convergence, thereby obtaining multiple intrinsic mode functions.

3. The efficient solution method for flood control optimization scheduling of complex reservoir groups in a river basin according to claim 1 is characterized in that: S13 further comprises: S13a, divide the main stages of the original flood process of each reservoir, including flood stage, flood peak stage and water withdrawal stage; S13b, preliminarily setting different sliding window lengths according to the traffic characteristics of each stage; S13c, calculating the rate of change and acceleration of traffic flow in adjacent time periods, evaluating traffic fluctuations in real time, and adjusting the window length accordingly; S13d, based on the dynamically adjusted window length, the sliding window method is used to segment the traffic signal to obtain the input samples required by the VMD-Enhanced CNN; S13e, prepare corresponding labels for each time series segment to ensure the representativeness and diversity of the samples.

4. The efficient solution method for flood control optimization scheduling of complex reservoir groups in a river basin according to claim 1 is characterized in that: S21 further includes: S21a, define the flood control risk of the reservoir, and calculate the ratio of the maximum storage capacity used by the reservoir to absorb flood water to the available flood control storage capacity; S21b, introduce reservoir flood control risk coefficient and flood control contribution weight index, aiming to minimize the comprehensive flood control risk of reservoir groups; S21c, define the flood risk of the protection area, and calculate the ratio of the maximum flow of the flood control section to the designed safe flow; S21d, introduce flood risk coefficient and flood risk weight index of protection area, aiming at minimizing the comprehensive risk of protection area group; S21e. Establish the constraints of the flood control and dispatching model of the reservoir group, including the water balance constraints of a single reservoir, the river flood routing constraints, the water balance constraints between upstream and downstream nodes, the upper and lower limit constraints of the reservoir water storage capacity, the reservoir discharge capacity constraints, the reservoir discharge amplitude constraints and boundary conditions.

5. The efficient solution method for flood control optimization scheduling of complex reservoir groups in a river basin according to claim 1 is characterized in that: S3 further includes: S31, using the Ray distributed computing framework to distribute the equivalent models of each subsystem to multiple independent CPU processes; S32, extracting basic information of each reservoir and flood control section, and using the reservoir inflow process as input of the subsystem equivalent model, and executing the EP-DOA-CUDA algorithm based on the CUDA architecture and Numba library within each CPU process to obtain the optimization results of each equivalent model; S33, after each CPU process independently completes the optimization task of the subsystem equivalent model, the main process collects the optimization results of each equivalent model and integrates them; S34. Using the optimization results of the subsystem equivalent model as constraints, combined with global constraints including river flood routing and water balance between upstream and downstream nodes, the optimal coordination variables of the system in the global scope are calculated.

6. The efficient solution method for flood control optimization dispatching of complex reservoir groups in a river basin according to claim 5 is characterized in that: S32 further includes: S32a, initializing dream individuals on the CPU side, including the initial position and "dream" state of each individual; S32b, efficiently transferring the initialized dream entity from the CPU memory to the GPU memory; S32c, design and implement CUDA kernel function, responsible for fitness evaluation of dream individuals, update of individual historical optimal dream state Pbest, and synchronous calculation of dream generation state; S32d, configuring CUDA thread blocks and grid parameters, and dynamically adjusting the grid size according to the total number of dream individuals; S32e, comparing the current fitness of each dream individual with its historical best dream state Pbest in parallel on the GPU, and for individuals whose fitness is better than the historical best value, updating their Pbest to the current fitness and corresponding position; S32f, transferring the updated dream individual data from the GPU memory back to the CPU memory, and determining the global optimal dream state Gbest in the current iteration on the CPU side; S32g, monitoring the iteration process of the DOA algorithm on the CPU side, and determining whether to continue the iteration according to a preset termination condition; S32h. After the termination condition is met, the final global optimal dream state Gbest is output as the optimization result of the subsystem equivalent model.

7. The efficient solution method for flood control optimization scheduling of complex reservoir groups in a river basin according to claim 1 is characterized in that: S4 further includes: S41, taking the system's optimal coordination variables as boundary conditions, uniformly and randomly generate a population, and each individual is a solution in the problem space; S42, calculate the fitness value of each individual, and use the objective function to measure the quality of the individual; S43, sort the fitness of individuals in the population and select the individuals with the best fitness as elite individuals; S44. Promote the improvement of local solutions through information sharing mechanism between elite individuals and other individuals; S45, combining global search and local search strategies, enhancing local search through the guidance of elite individuals, and introducing random perturbations for global exploration; S46. In each round of iteration, the position of each individual is updated, and the position of the solution is adjusted according to the guidance of the elite individual and the balance between global and local searches; S47, determining whether the termination conditions are met, including reaching the maximum number of iterations, the fitness value change is less than a set threshold, or the fitness reaches a preset target value; S48. Output the final optimization result, which is the optimal dispatching plan for the reservoir group.

8. The efficient solution method for flood control optimization dispatching of complex reservoir groups in a river basin according to claim 7 is characterized in that: In S4, the specific implementation process of the IVYA optimization algorithm includes: Update individual positions through information sharing mechanism between elite individuals and other individuals: Xnew_i = Xnew_i + α×(Xe -Xi), where α is a parameter that controls the intensity of cooperation; Xi represents the i-th individual in the population; Xe represents the elite individual in the population; Combined with the global search strategy: Xglobal_i = X_i + β×(X_r - X_i), where X_r is an individual randomly selected from the population, and β is a parameter that controls the global search step size; Combined with the local search strategy: Xlocal_i = Xi + δ×(Xe - Xi), where δ is the control parameter of the local search step size; In each round of iteration, elite guidance and random perturbations are considered comprehensively to update individual positions: Xnew_i = X_i + λ×(X_e- X_i) + u×(X_r - X_i), where λ and u control the intensity of elite guidance and global search perturbations, respectively.

Images