A plain river network dike area flood control and waterlogging removal multi-objective particle swarm optimization scheduling system

Abstract

The present application relates to the field of process control, in particular to a plain river network dike area flood control multi-objective particle swarm optimization scheduling system. The specific implementation steps include: building a coupled hydrodynamic model module, calculating the water exchange between one-dimensional hydrodynamic model and two-dimensional hydrodynamic model, and generating training data set through multiple simulations, training and generating neural network proxy model; using particle swarm optimization algorithm, generating initial population based on the topological structure information and heuristic rules of plain river network; dynamically adjusting the feasible region constraint of optimization algorithm according to external working condition information, and inputting the candidate scheduling scheme meeting the constraint into the neural network proxy model to output the predicted scheduling state; based on the predicted scheduling state and weight coefficient, the fitness value of each candidate scheduling scheme is calculated, the optimal scheduling scheme is output through iteration optimization, and the decision-making efficiency, robustness and safety of the scheduling system in the face of complex working conditions and emergencies are improved.

Description

Technical Field

[0001] This invention relates to the field of process control, specifically to a multi-objective particle swarm optimization scheduling system for flood control and drainage in plain river network polder areas. Background Technology

[0002] To safeguard the lives and property of people in plain river network areas, significant progress has been made in flood control and drainage scheduling technology for water conservancy projects. Existing technologies, based on key cross-sections such as warning and guaranteed water levels, have developed a set of effective static scheduling procedures and diagrams. In terms of engineering systems, water conservancy hubs such as sluice gates, pumping stations, and control gates distributed throughout the river network constitute a solid physical foundation for flood control and drainage. Regarding information acquisition, hydrological telemetry networks and meteorological forecasting systems can provide relatively timely water and rainfall information, providing data input for scheduling decisions.

[0003] However, when faced with basin-wide, high-intensity extreme rainfall events, the deep-seated contradictions of the existing dispatching system are increasingly exposed, and these challenges are particularly prominent from the perspective of process control. Some dispatching decisions exhibit typical decentralized characteristics, with the start and stop logic of each dispatching unit relying solely on its local state variables without considering the systemic impact of its dispatching behavior on the downstream and the global state. This uncoordinated concurrent drainage can easily cause severe superposition of flood peaks in local areas, leading to a rapid backwater effect on external river levels and turning local drainage needs into regional flood control crises. Some dispatching decisions also exhibit significant lag. Traditional dispatching models are mostly passive responses based on historical experience and established rules, lacking the ability to accurately predict and dynamically respond to future water situation trends, making it difficult to seize the valuable pre-discharge dispatching window. With the increasing number of water conservancy projects, the entire dispatching system presents itself as a high-dimensional, nonlinear, and strongly coupled complex process control object. The sharp increase in decision variables causes the computational complexity of seeking the global optimal solution to grow exponentially, the so-called dimensionality curse, which far exceeds the capabilities of traditional optimization algorithms and manual dispatching.

[0004] Therefore, this invention proposes a multi-objective particle swarm optimization scheduling system for flood control and drainage in plain river network polder areas. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-objective particle swarm optimization scheduling system for flood control and drainage in plain river network polder areas, enabling rapid, coordinated, and globally optimal intelligent scheduling of all water conservancy facilities within the region under complex and dynamic flood conditions. Specifically, this is achieved by constructing a hierarchical decision-making framework integrating physical simulation, surrogate learning, and intelligent optimization, and deeply fusing high-fidelity simulation based on one-dimensional-two-dimensional coupled hydrodynamics, rapid derivation using a neural network surrogate model, and a multi-objective particle swarm optimization algorithm guided by topological rules and dynamic weights.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A multi-objective particle swarm optimization scheduling system for flood control and drainage in plain river network polder areas includes:

[0008] The coupled hydrodynamic model module calculates the flow exchange between a one-dimensional hydrodynamic model used to simulate underground pipe network flow and a two-dimensional hydrodynamic model used to simulate surface runoff.

[0009] The proxy model generation module calls the coupled hydrodynamic model module to perform multiple simulations to generate a training dataset, and trains the training dataset to generate a neural network proxy model.

[0010] The multi-objective weight adaptive module dynamically determines a set of weight coefficients representing the relative importance of multiple scheduling objectives based on the received real-time warning level.

[0011] The particle swarm optimization scheduling module generates initial candidate scheduling schemes as the initial population of the particle swarm based on the topological structure information of the plain river network and preset heuristic rules. During the iterative optimization process, the feasible region constraints of the optimization algorithm are dynamically adjusted according to the real-time received external project information. The candidate scheduling schemes are input into the neural network surrogate model and the predicted scheduling state is output. Based on the predicted scheduling state and weight coefficients, the fitness value of each candidate scheduling scheme is calculated, and the optimal scheduling scheme is output through iterative optimization of the particle swarm algorithm.

[0012] Preferably, the coupled hydrodynamic model module calculates the flow exchange between a one-dimensional hydrodynamic model simulating underground pipe network flow and a two-dimensional hydrodynamic model simulating surface runoff, including: the one-dimensional hydrodynamic model is a stormwater flood management model solved based on the Saint-Venant equations; the two-dimensional hydrodynamic model is a flood runoff model solved based on two-dimensional unsteady shallow water equations; the calculation method of the flow exchange includes: calculating the vertical exchange by using the orifice flow formula and the weir flow formula to calculate the overflow caused by the head difference between the underground pipe network nodes and the two-dimensional surface grid; and calculating the lateral exchange by using the riverbank of the one-dimensional hydrodynamic model as the internal boundary of the two-dimensional hydrodynamic model, and using the weir flow formula to calculate the lateral runoff flow generated when the river level exceeds the bank elevation.

[0013] Preferably, the proxy model generation module includes: calling the coupled hydrodynamic model module to simulate operating conditions covering different rainfall scenarios and scheduling scheme combinations, and generating a training dataset; the neural network proxy model includes an input layer, a hidden layer, and an output layer; the input layer receives a scheduling scheme vector consisting of the opening degree of each sluice gate and the start / stop state of the pumping station; the output layer outputs a multi-dimensional predicted scheduling state vector consisting of the highest water level of key sections in the region and the flooded area.

[0014] Preferably, the multi-objective weight adaptive module includes: an internal storage of a scenario-policy mapping rule base; wherein each scenario is characterized by a warning level defined by real-time rainfall intensity and external river tide level data; each policy corresponds to a set of preset weight coefficient combinations for the two objectives of flood control and drainage; the module receives real-time rainfall intensity and real-time external river tide level data representing the current warning level from external sources in real time; the received real-time data is matched one by one with multiple preset scenarios in the rule base to determine the interval in which the current real-time data falls, defined by multiple closest preset scenarios; when the real-time data precisely matches a preset scenario, the module directly outputs the preset policy weight corresponding to that scenario; when the real-time data falls between two preset scenarios, the module performs linear interpolation calculation on the preset policy weights corresponding to the two preset scenarios to generate and output a set of dynamic weight coefficients applicable to the current scenario.

[0015] Preferably, the particle swarm optimization scheduling module generates initial candidate scheduling schemes as the initial population of the particle swarm based on the topological structure information of the plain river network and preset heuristic rules. This includes: acquiring the topological structure information of a plain river network represented by nodes and edges, containing the upstream and downstream relationships of various water conservancy facilities; applying preset heuristic rules to perform path traversal and priority sorting on the topological structure; the heuristic rules include spatial priority rules, which determine the spatial execution order of scheduling operations based on the upstream and downstream relationships in the topological structure; the heuristic rules also include temporal priority rules, which determine the temporal order of scheduling operations based on the division of the main stream and tributaries in the topological structure; and generating a set of candidate scheduling schemes with random perturbations in specific operation parameters that conform to the heuristic rules as the initial population.

[0016] Preferably, the step of dynamically adjusting the feasible region constraints of the optimization algorithm based on real-time received external engineering information during the iterative optimization process, inputting candidate scheduling schemes into the neural network proxy model and outputting a predicted scheduling state includes: receiving external engineering information including real-time fault alarm information of water conservancy facilities; dynamically adjusting the value range of the corresponding decision variables in the optimization algorithm as feasible region constraints; checking the candidate scheduling schemes newly generated by the particle swarm optimization algorithm and performing boundary repair processing on schemes that exceed the feasible region constraints; using the neural network proxy model as the fitness function evaluation engine of the particle swarm optimization algorithm, and transmitting the candidate scheduling schemes generated by the particle swarm optimization algorithm in each iteration as input vectors to the input layer of the neural network proxy model; having the neural network proxy model perform a forward propagation calculation instead of calling the coupled hydrodynamic model to perform a complete and time-consuming physical simulation; and having the output layer of the neural network proxy model output a predicted scheduling state vector that can characterize the multi-dimensional physical state of the highest water level of the key section and the flooded area in the region under the candidate scheduling scheme.

[0017] Preferably, based on the predicted scheduling state and weight coefficients, the fitness value of each candidate scheduling scheme is calculated, and the optimal scheduling scheme is output through iterative optimization using the particle swarm optimization algorithm. This includes: weighting and summing the weight coefficients with the predicted scheduling state vector to calculate the scalar fitness value of each candidate scheduling scheme; treating each candidate scheduling scheme as a particle, updating the particle's velocity and position in the solution space based on the optimal fitness value scheme in the particle's own historical iterations and the optimal fitness value scheme in the historical iterations of the entire particle population, generating a new candidate scheduling scheme; and after satisfying a preset termination condition, outputting the scheme with the optimal fitness value in the current historical iterations of the entire particle population as the optimal scheduling scheme.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] 1. This invention constructs a coupled hydrodynamic model as a high-fidelity physical world simulation engine and employs a neural network surrogate model to learn and replace the coupled hydrodynamic model, achieving rapid and efficient approximate extrapolation of complex working conditions in plain river network polder areas. It can not only comprehensively and dynamically simulate the complex water flow exchange process between surface runoff and underground pipe networks, but also compress the physical simulation time that originally required several hours to the second level, ensuring that subsequent optimization decisions have a solid scientific basis that reflects the real physical world. At the same time, this invention uses the surrogate model as a fast fitness function evaluation engine for the optimization algorithm, improving the system's real-time response capability and forward-looking decision-making capability under rapidly changing flood conditions, meeting the current demand for high timeliness in flood control and drainage scheduling.

[0020] 2. This invention introduces heuristic rules based on the topological structure information of the entire river network and common sense of hydraulics during the initialization stage of the particle swarm optimization algorithm. This enables intelligent guidance of the search starting point of the optimization algorithm. It not only avoids the blind random search of traditional optimization algorithms in a huge solution space, but also concentrates computing resources on the region where the optimal solution is more likely to exist physically from the first step of the algorithm. This topology-aware initialization mechanism significantly improves the convergence speed of the algorithm and the success rate of finding the global optimal solution, effectively solving the "curse of dimensionality" problem in high-dimensional and complex scheduling problems.

[0021] 3. This invention constructs an adaptive module that can dynamically adjust the weights of multiple objectives based on the real-time warning level, and uses these dynamic weights to guide the iterative direction of the particle swarm optimization algorithm. This achieves intelligent trade-offs for multiple and conflicting scheduling objectives, enabling the system to automatically switch core management objectives at different flood stages. It also transforms a difficult multi-objective optimization problem into a single-objective problem with a clear optimization direction in real time. This scenario-adaptive decision-making paradigm ensures that the final output scheduling scheme can maximize the comprehensive disaster reduction efficiency of the entire region under the current specific conditions, improving the intelligence level and practical applicability of scheduling decisions. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall architecture of the multi-objective particle swarm optimization scheduling system for flood control and drainage in plain river network polder areas according to an embodiment of the present invention;

[0023] Figure 2 This is a schematic diagram of the coupling mechanism of the coupled hydrodynamic model in an embodiment of the present invention;

[0024] Figure 3 This is a schematic diagram of the workflow of the particle swarm optimization scheduling module in an embodiment of the present invention. Detailed Implementation

[0025] To enable a more comprehensive and clear understanding of the objectives, technical solutions, and beneficial effects of this invention, the technical solutions of this invention will be described in detail below with reference to specific embodiments. It should be noted that these embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Those skilled in the art can make various equivalent modifications or substitutions based on these embodiments without departing from the concept and scope of this invention, and all such modifications and substitutions should be considered to fall within the scope of protection of this invention.

[0026] This invention provides a multi-objective particle swarm optimization scheduling system for flood control and drainage in plain river network polder areas. Its core lies in constructing a hierarchical and modular intelligent decision-making framework to achieve precise and efficient management of complex water conservancy systems. (Refer to...) Figure 1The system mainly includes the following four closely cooperating functional modules: coupled hydrodynamic model module, surrogate model generation module, multi-objective weight adaptive module, and particle swarm optimization scheduling module.

[0027] Example 1

[0028] A multi-objective particle swarm optimization scheduling system for flood control and drainage in plain river network polder areas includes:

[0029] The coupled hydrodynamic model module calculates the flow exchange between a one-dimensional hydrodynamic model used to simulate underground pipe network flow and a two-dimensional hydrodynamic model used to simulate surface runoff.

[0030] The proxy model generation module calls the coupled hydrodynamic model module to perform multiple simulations to generate a training dataset, and trains the training dataset to generate a neural network proxy model.

[0031] The multi-objective weight adaptive module dynamically determines a set of weight coefficients representing the relative importance of multiple scheduling objectives based on the received real-time warning level.

[0032] The particle swarm optimization scheduling module generates initial candidate scheduling schemes as the initial population of the particle swarm based on the topological structure information of the plain river network and preset heuristic rules. During the iterative optimization process, the feasible region constraints of the optimization algorithm are dynamically adjusted according to the real-time received external project information. The candidate scheduling schemes are input into the neural network surrogate model and the predicted scheduling state is output. Based on the predicted scheduling state and weight coefficients, the fitness value of each candidate scheduling scheme is calculated, and the optimal scheduling scheme is output through iterative optimization of the particle swarm algorithm.

[0033] Furthermore, refer to Figure 2 The coupled hydrodynamic model module calculates the flow exchange between a one-dimensional hydrodynamic model simulating underground pipe network flow and a two-dimensional hydrodynamic model simulating surface runoff, including:

[0034] The one-dimensional hydrodynamic model is a stormwater flood management model (SWMM) based on the Saint-Venant equations. The calculation of water flow in pipes and rivers is based on solving the complete one-dimensional Saint-Venant equations, including the continuity equation and the momentum equation.

[0035] The continuity equation characterizes the law of conservation of mass, which states that the amount of water flowing into and out of a control volume must be balanced.

[0036] The momentum equation characterizes the relationship between forces acting on a water body (such as gravity, pressure, and frictional resistance) and the rate of change of water momentum. It can accurately simulate complex hydraulic phenomena such as water storage, backwater surging, and pressure flow. The frictional loss of the water flow is calculated based on the Manning equation, and its magnitude is directly proportional to the square of the Manning roughness coefficient and the square of the flow velocity, and inversely proportional to the fourth-third power of the hydraulic radius. Through numerical solutions to the Saint-Venant equations, the one-dimensional hydrodynamic model can output the water level and flow rate at any cross-section in pipes and rivers.

[0037] The described two-dimensional hydrodynamic model is a flood overflow model based on two-dimensional unsteady shallow water equations. This model discretizes the surface region into a series of square computational grids and solves for the water exchange between each grid and its four adjacent grids. The core calculations of the model are also based on the continuity equation (ensuring water balance in each grid) and the momentum equation. In the two-dimensional model, the exchange flow between adjacent grids is calculated according to the Manning formula, and its magnitude depends on the head difference between the two grids, the cross-sectional area of ​​the water passage, and the roughness coefficient. By solving the two-dimensional unsteady shallow water equations, the two-dimensional hydrodynamic model can simulate the two-dimensional overflow, evolution, and recession of floodwaters on the surface.

[0038] The coupled hydrodynamic model module also executes a two-way dynamic interaction method with mutual boundary conditions. Within each coupled calculation time step, the system performs the following data exchange and calculations in parallel:

[0039] The weighted average water level value of all two-dimensional grids at the downstream outlet connection section of the one-dimensional river channel model is calculated and output using the two-dimensional hydrodynamic model.

[0040] The one-dimensional hydrodynamic model receives this water level value and uses it as a dynamically updated fixed water level boundary condition for its own calculation in the next time step.

[0041] The total flow rate of the water flowing out of the downstream outlet is calculated and output using the one-dimensional hydrodynamic model.

[0042] The two-dimensional hydrodynamic model receives this flow rate value and uses it as a boundary condition for a flow source term. This condition is then distributed to multiple surface grids at the connecting section as the initial inflow for the surface runoff calculation in the next time step.

[0043] This embodiment uses a real-time bidirectional boundary interaction mechanism to simulate the complex backwater backing effect that cannot be accurately captured by unidirectional coupling or simplified connection methods. This improves the fundamental technical problem that traditional methods cannot simulate this kind of bidirectional feedback, which leads to an overestimation of the system's drainage capacity and thus causes decision-making errors. This is also the key basis for the high-fidelity simulation of this invention.

[0044] The calculation method for the water exchange volume includes:

[0045] The calculation of vertical exchange volume involves the following steps: between the manholes and other nodes in the SWMM model and the corresponding surface grid in the 2D model, when the water level at the pipe network node is higher than the water level at the surface grid, the overflow from the pipe network to the surface is calculated using the orifice outflow formula, and the flow rate is proportional to the square root of the head difference; when the surface water level is higher than the water level at the pipe network node, the return flow from the surface to the pipe network is calculated using the weir flow formula, and the flow rate is proportional to the cube of the head difference.

[0046] The calculation of lateral exchange volume involves treating the riverbank as a broad-crested weir when the water level calculated by the one-dimensional river model exceeds the preset embankment elevation. The lateral overflow flow generated is then calculated using the weir flow formula and input as a source term into the two-dimensional surface overflow model.

[0047] This embodiment solves the fundamental problem that single-dimensional models cannot accurately simulate the multi-scale and cross-domain characteristics of urban flooding by coupling the Saint-Venant equations of the one-dimensional hydrodynamic model and solving the shallow water equations of the two-dimensional hydrodynamic model. It also clearly defines the vertical and lateral exchange mechanism based on the physical formulas of orifice and weir flow, and constructs a simulation engine that can reproduce the real physical world with high fidelity.

[0048] Furthermore, the proxy model generation module includes:

[0049] The coupled hydrodynamic model module is invoked to simulate operating conditions covering different rainfall scenarios and scheduling scheme combinations, generating a training dataset. The neural network proxy model includes an input layer, a hidden layer, and an output layer. The input layer receives a scheduling scheme vector consisting of the opening degree of each sluice gate and the start / stop status of the pumping station. The output layer outputs a multi-dimensional predicted scheduling state vector consisting of the highest water level of key sections in the region and the flooded area.

[0050] The proxy model generation module calls the coupled hydrodynamic model module to perform offline simulations of operating conditions covering different rainfall scenarios and scheduling scheme combinations, generating a large-scale training dataset. The rainfall scenarios cover typical rainstorm events of different intensities, durations, and spatial distributions. The scheduling scheme combinations include various opening settings of sluice gates (e.g., discrete levels or continuous percentages from fully closed to fully open) and the start-stop status and operating power configuration of pumping stations (e.g., binary start-stop status or multi-level power output). Sufficiently diverse and representative operating conditions are generated through system sampling methods (such as Latin hypercube sampling). The input (rainfall scenario, scheduling scheme) and output (highest water level of key sections in the region, flooded area, etc.) of each simulation constitute a training sample. These data will be used for subsequent supervised learning of the neural network.

[0051] The constructed neural network surrogate model includes an input layer, hidden layers, and an output layer. In this embodiment, a multilayer perceptron is preferably used as the surrogate model. The input layer receives a scheduling scheme vector consisting of the opening degree of each sluice gate (e.g., a normalized continuous value between 0 and 1 or a discrete code) and the start / stop status of the pumping station (e.g., a binary variable of 0 / 1, or a discrete code of multiple power outputs). The dimension of the input vector is equal to the total number of scheduling variables for all sluice gates and pumping stations. The output layer outputs a multidimensional predicted scheduling state vector consisting of the highest water level of key sections in the region and the flooded area. These output values ​​are key indicators for the particle swarm optimization algorithm to evaluate the fitness of the scheduling scheme. The dimension of the output vector is equal to the number of key sections to be predicted plus the flooded area. The model includes multiple hidden layers to learn the nonlinear relationship between the input scheduling scheme and the complex hydrodynamic response. The specific number of layers and the number of neurons in each layer can be adjusted according to the model complexity and the amount of data. For example, two to three hidden layers can be used, with the number of neurons decreasing (e.g., 64 neurons in the first hidden layer and 32 neurons in the second hidden layer). These hidden layers use modified linear units as nonlinear activation functions to learn and fit the deep nonlinear mapping relationship between the scheduling scheme and the complex hydrodynamic response.

[0052] The training process employs supervised learning, utilizing the generated training dataset to train the neural network. The optimizer chosen is the Adam adaptive learning rate optimizer, with a learning rate set to 0.001. The loss function is the mean squared error, minimizing the deviation between the surrogate model's predictions and the actual simulation results of the coupled hydrodynamic model. The training process continues until the model's performance on the validation set reaches the convergence criterion or the maximum number of training epochs is reached.

[0053] The proxy model generation module also employs a phased training method when training the neural network proxy model:

[0054] In the first stage, preliminary training is performed: from the training dataset, operational data containing only single water conservancy facility scheduling or simple dual facility linkage are selected to form the first training subset; using the first training subset, the randomly initialized neural network proxy model is trained to learn and fit the basic causal relationship between scheduling behavior and hydrodynamic response.

[0055] The second stage involves in-depth training: after completing the initial training, the neural network agent model that has already been initially trained is further fine-tuned and optimized using a complete training dataset containing all complex multi-facility joint scheduling scenarios.

[0056] The method may further include a cross-regional parameter transfer step for building a proxy model for a new region, wherein some or all of the weight parameters of a neural network proxy model that has been trained in the source region are used as the initial weight parameters for training the proxy model for the new region to be built; and the proxy model for the new region, loaded with the initial weight parameters, is fine-tuned using a small amount of simulated data from the new region.

[0057] This embodiment decomposes the learning process of complex nonlinear mapping relationships by adopting this phased training method. Through cross-regional parameter transfer and fine-tuning mechanisms, it greatly reduces the amount of repetitive simulation and training computation required when building surrogate models for new regions. This improves the technical bottleneck in practical engineering applications, which requires a time-consuming, zero-based model training for each new research region, and significantly enhances the scalability and engineering practicality of the entire solution.

[0058] This embodiment constructs a neural network proxy model to achieve rapid and efficient approximate extrapolation of complex working conditions in plain river network polder areas. It compresses the physical simulation time, which originally required several hours, to the second level, significantly improving the system's real-time response capability and forward-looking decision-making capability under rapidly changing flood conditions, and meeting the current demand for high timeliness in flood control and drainage scheduling.

[0059] Furthermore, the multi-objective weight adaptive module includes:

[0060] The multi-objective weighted adaptive module internally stores a scenario-policy mapping rule base, which encodes historical experience and expert knowledge in a structured manner. Each scenario is characterized by a warning level defined by real-time rainfall intensity (e.g., in millimeters per hour) and external river tide level data (e.g., in meters). These warning levels can be scientifically classified based on regional flood control and drainage standards, historical hydrological data, and expert experience. For example, they are divided into "Blue Warning" (light rainfall, low tide, i.e., real-time rainfall intensity less than 5 mm per hour and external river tide level less than 0.5 meters below the reference level) and "Yellow Warning" (moderate rainfall, mid-tide, i.e., real-time rainfall intensity between 5 mm per hour and 15 mm per hour, or ... The alerts include "Tide levels between 0.5 and 1.5 meters above the datum," "Orange Alert" (heavy rainfall, high tide, i.e., real-time rainfall intensity between 15 and 30 millimeters per hour, or outer river tide levels between 1.5 and 2.5 meters above the datum), and "Red Alert" (extremely heavy rainfall, super high tide, i.e., real-time rainfall intensity exceeding 30 millimeters per hour, or outer river tide levels exceeding 2.5 meters above the datum). Each strategy corresponds to a set of pre-set weighted coefficients for the two main objectives of flood control and drainage, and may also include water resource utilization objectives. The sum of all weighted coefficients is 1. For example, during light rainfall at low tide, the weight of water resource utilization may be higher (e.g., 0.4), while the weights of flood control and drainage are lower (0.3 each). During heavy rainfall at high tide, the weights of flood control and drainage will significantly increase (e.g., flood control 0.8, drainage 0.15), while the weight of water resource utilization will decrease (e.g., 0.05).

[0061] During real-time system operation, the dynamic determination process of the multi-objective weight adaptive module includes: receiving real-time rainfall intensity and real-time river tide level data representing the current warning level from external sources; matching the received real-time data one by one with multiple preset scenarios in the rule base to determine the interval within which the current real-time data falls, defined by multiple closest preset scenarios; when the real-time data precisely matches a preset scenario, the module directly outputs the preset strategy weight corresponding to that scenario; when the real-time data falls between two preset scenarios (e.g., "yellow warning" and "orange warning"), the module performs linear interpolation calculations on the preset strategy weights corresponding to these two preset scenarios to generate and output a set of dynamic weight coefficients applicable to the current scenario. Specifically, the weights are obtained as follows:

[0062] When the system is running in real time, the module first receives real-time rainfall intensity and real-time outer river tide level data that represent the current warning level from the outside. Then, it matches this set of real-time data with multiple preset scenarios in the rule base one by one.

[0063] When the real-time data precisely matches the definition of a preset scenario (for example, the current rainfall intensity is 20 mm / hour, which falls within the range of "orange warning"), the module will directly output the preset strategy weight corresponding to the scenario (for example, directly return flood control weight 0.8 and drainage weight 0.15).

[0064] When the real-time data falls between the quantization intervals of two preset scenarios (for example, the current rainfall intensity is 14.5 mm / hour, between the boundaries of "yellow alert" and "orange alert"), in order to ensure a smooth transition of weight switching and avoid abrupt changes in the scheduling strategy, the module will initiate a linear interpolation calculation mechanism. It will first obtain the preset strategy weights corresponding to these two adjacent preset scenarios ("yellow alert" and "orange alert"), and then perform linear interpolation calculations based on the relative position of the current real-time data between the boundaries of these two scenarios, thereby generating and outputting a set of dynamic weight coefficients most suitable for the current specific scenario.

[0065] This embodiment constructs an adaptive module that dynamically adjusts the weights of multiple objectives based on real-time early warning levels. These dynamic weights are then used to guide the iterative direction of the particle swarm optimization algorithm, enabling intelligent trade-offs for multiple, conflicting scheduling objectives. This allows the system to automatically switch core management objectives at different flood stages and transforms a difficult multi-objective optimization problem into a single-objective problem with a clear optimization direction in real time. This scenario-adaptive decision-making paradigm ensures that the final output scheduling scheme maximizes the overall disaster reduction efficiency of the entire region under the current specific conditions, improving the intelligence level and practical applicability of scheduling decisions.

[0066] Furthermore, refer to Figure 3 The particle swarm optimization scheduling module, based on the topological structure information of the plain river network and preset heuristic rules, generates initial candidate scheduling schemes as the initial population of the particle swarm, including:

[0067] The particle swarm optimization scheduling module executes a dual-intelligent-guided optimization method:

[0068] During the optimization startup phase, an initialization guidance step based on topology rules is executed. This step generates a set of initial candidate scheduling schemes biased in the physically feasible solution region in the solution space based on the topological structure information of the plain river network and preset heuristic rules, which serve as the initial population of the particle swarm.

[0069] During the optimization iteration phase, an fitness guidance step based on dynamic weighting is executed. This step determines the optimization direction for each iteration by weighting and summing the weight coefficients output by the multi-objective weight adaptive module with the prediction results output by the neural network surrogate model.

[0070] This embodiment employs a dual intelligent guidance mechanism, deeply integrating domain expert knowledge and real-time scenario judgment into the inner and outer loops of the optimization algorithm. Initialization guidance, by constraining the starting point of the search, improves the shortcomings of traditional particle swarm optimization algorithms, which suffer from slow convergence speed and wasted computational resources due to blind random initialization when facing high-dimensional complex problems. Fitness guidance, by dynamically adjusting the optimization objective, improves the technical bottleneck in multi-objective conflict scenarios, where the algorithm is prone to getting trapped in local optima due to fuzzy evaluation criteria and unclear search direction.

[0071] To obtain the topological structure information of a plain river network, characterized by nodes (representing water conservancy facilities or key river sections) and edges (representing river channels or pipeline connections), which includes the upstream and downstream relationships of various water conservancy facilities (such as sluice gates, pumping stations, and control gates), including the geographical location, connection relationships, and control range of each facility, this information is usually stored in the form of a graph database or adjacency matrix.

[0072] Using pre-defined heuristic rules, path traversal and priority sorting are performed on the topological structure to generate a physically plausible initial population. Path traversal can be achieved by performing graph traversal algorithms (such as breadth-first search or depth-first search) on the topological map of the hydraulic facilities in the plain river network to identify the upstream and downstream connections between facilities. The heuristic rules include spatial priority rules, which determine the spatial execution order of scheduling operations based on the upstream and downstream relationships in the topological structure. For example, during flood control, upstream facilities are usually prioritized to control inflow, or downstream facilities are prioritized to ensure smooth outflow; during drainage, drainage facilities in low-lying areas may be prioritized. The heuristic rules also include time priority rules, which determine the temporal order of scheduling operations based on the division of main streams and tributaries in the topological structure. For example, before a flood, tributaries may be pre-emptively discharged to free up reservoir capacity; during floods... During peak periods, key scheduling is implemented for main channel facilities to ensure unobstructed drainage. Based on the determined spatial execution order and temporal sequence, a set of candidate scheduling schemes that conform to heuristic rules and have random perturbations in specific operational parameters (such as sluice gate opening, pump station start-up and shutdown time, and operating power) are generated as the initial population. The random perturbation refers to making small-scale random adjustments to specific operational parameters such as sluice gate opening and pump station start-up and shutdown status within their designed operating range based on the initial schemes generated by heuristic rules, thereby increasing the diversity of the initial population.

[0073] This embodiment introduces heuristic rules based on the topology of the entire river network and hydraulic principles during the initialization phase of the particle swarm optimization algorithm. This intelligently guides the search starting point of the optimization algorithm, avoiding the blind random search in the huge solution space common in traditional optimization algorithms. Furthermore, from the very first step of the algorithm's startup, computational resources are concentrated on regions where the physical optimum is more likely to exist. This topology-aware initialization mechanism significantly improves the algorithm's convergence speed and the success rate of finding the global optimum, effectively solving the "curse of dimensionality" problem in high-dimensional, complex scheduling problems.

[0074] Furthermore, in the iterative optimization process, the feasible region constraints of the optimization algorithm are dynamically adjusted based on the real-time received external work information, and candidate scheduling schemes are input into the neural network proxy model and the predicted scheduling state is output, including:

[0075] The system first receives external operational information in real time from external monitoring systems or higher-level scheduling platforms. This external operational information can be a real-time fault alarm signal issued by a water pump due to power supply problems, or a status feedback that a sluice gate cannot reach its maximum opening due to mechanical failure. Upon receiving such information, the particle swarm optimization scheduling module immediately and dynamically adjusts the internally defined feasible domain used to constrain decision variables. Specifically, if a fault shutdown signal for "Main Pump Station No. 2" is received, at the algorithm level, the value range of the decision variable corresponding to the operating state of Pump Station No. 2 in the particle position vector is forcibly modified from its original allowed operating range to a fixed value representing its complete shutdown state, for example, set to [0, 0]. After completing the dynamic update of the feasible domain, the system checks each newly generated candidate scheduling scheme. If it finds that the value of any dimension exceeds the boundary of the currently updated feasible domain, it will adopt a preset repair strategy, such as boundary absorption, to forcibly correct its value to the nearest valid boundary value.

[0076] After completing the aforementioned real-time constraint processing, the verified and repaired, physically feasible candidate scheduling schemes are then transmitted as input vectors to the input layer of the neural network proxy model. The neural network proxy model, acting as the fitness function evaluation engine for the particle swarm optimization algorithm, transmits the candidate scheduling schemes (i.e., the positions of particles in the solution space, representing a set of specific gate pump operation commands) generated by the particle swarm optimization algorithm in each iteration as an input vector to the input layer of the neural network proxy model. The neural network proxy model then performs a forward propagation calculation, replacing the invocation of the coupled hydrodynamic model to perform a complete physical simulation based on partial differential equation solving. Finally, the output layer of the neural network proxy model outputs a predicted scheduling state vector that characterizes the multidimensional physical state of the highest water level at key sections and the flooded area within the region under the candidate scheduling scheme.

[0077] This embodiment uses a neural network surrogate model as the fitness function evaluation engine for the particle swarm optimization algorithm, transforming a computationally expensive physical simulation process into a lightweight neural network forward propagation process. This solves the core technical bottleneck that complex hydrodynamic models cannot be directly used for real-time optimization and scheduling due to their enormous computational time (usually several hours). The system can complete tens of thousands of scheme evaluations and optimizations within minutes, thereby improving the timeliness and practical usability of optimization decisions.

[0078] Furthermore, based on the predicted scheduling state and weight coefficients, the fitness value of each candidate scheduling scheme is calculated, and the optimal scheduling scheme is output through iterative optimization using the particle swarm optimization algorithm, including:

[0079] After obtaining the predicted scheduling state, the system performs a weighted summation of the weight coefficients output by the multi-objective weight adaptive module and the multi-dimensional predicted scheduling state vector output by the neural network surrogate model to calculate the scalar fitness value of each candidate scheduling scheme. This scalar value reflects the merits of the scheme in the current scenario in terms of the overall objective. Generally, the smaller the objective function value (or the larger it is, depending on the optimization direction), the better the fitness.

[0080] Each candidate scheduling scheme is treated as a particle. Based on the particle's own optimal fitness value (individual optimal) and the optimal fitness value of the entire particle population (global optimal), the particle's velocity and position in the solution space are updated, generating a new candidate scheduling scheme. In practice, the operating parameters of the particle swarm optimization algorithm can be adjusted according to actual conditions. For example, the particle population size is typically set between 20 and 100 particles, the inertia weight used to update particle velocity is typically between 0.4 and 0.9, and the individual learning factor (cognitive coefficient) and social learning factor (social coefficient) are typically between 1.5 and 2.5, controlling the degree to which particles approach their own historical optimal position and the population's historical optimal position, respectively.

[0081] After the preset termination conditions are met, the system outputs the optimal fitness value of the current particle population in the history of iterations as the optimal scheduling scheme. The termination conditions may include reaching the maximum number of iterations (e.g., 1000 times), the fitness value converging to a certain threshold (e.g., the improvement rate is less than 0.001 for 20 consecutive iterations), or the computation time reaching the upper limit (e.g., 5 minutes), to ensure that a high-quality solution is obtained under real-time requirements.

[0082] This embodiment calculates a scalar fitness value by weighting and summing dynamically changing weight coefficients with a high-dimensional prediction effect vector. This transforms a mathematically difficult-to-handle multi-objective optimization problem involving multiple conflicting objectives into a single-objective optimization problem with a clear evaluation criterion. Through the inherent evolutionary mechanism of the particle swarm optimization algorithm, which takes into account both individual exploration and global optimization, the algorithm is able to search efficiently and robustly in a huge, nonlinear solution space. This improves upon the technical bottleneck of traditional optimization algorithms when facing high-dimensional, multi-objective conflict problems, which are prone to getting stuck in local optima and failing to find the global optimum due to unclear search direction and fuzzy evaluation criteria. This significantly improves the quality and reliability of the decision.

[0083] This embodiment constructs a hierarchical intelligent decision-making system integrating four core modules: coupled hydrodynamic simulation, surrogate model generation, multi-objective weight adaptation, and particle swarm optimization scheduling. The coupled hydrodynamic model module simulates physical processes with high fidelity; the surrogate model generation module learns and generates a neural network surrogate model that can quickly replace the physical simulation; the multi-objective weight adaptation module dynamically determines the optimal decision objective weights based on real-time warning levels; and the particle swarm optimization scheduling module, guided by topology rules, uses the surrogate model for rapid evaluation and, guided by the dynamic weights, iteratively optimizes and outputs the optimal scheduling scheme. By organically integrating high-fidelity physical simulation, rapid machine learning inference, and multi-objective intelligent optimization algorithms within a unified, end-to-end framework, this system achieves complementary advantages and synergistic effects from multiple advanced technologies. This solves the fundamental problem of existing scheduling technologies, which, due to inherent defects in models, algorithms, and decision logic, cannot simultaneously address the scientific, real-time, global, and adaptive aspects of decision-making.

[0084] Example 2

[0085] This embodiment deploys the above-mentioned plain river network flood control dispatch system in a certain plain river network polder area to realize intelligent and refined management of regional flood control and drainage. This plain river network polder area is low-lying, densely networked with rivers and complex water system, and faces enormous pressure for flood control and drainage. Traditional dispatch methods mainly rely on manual experience, which is difficult to cope with the risk of waterlogging under sudden heavy rainfall and complex tidal combinations. Therefore, there is an urgent need for a flood control dispatch system that can respond quickly, predict accurately, and optimize intelligently.

[0086] When faced with an impending rainstorm, the system will receive and integrate multi-source data in real time, including real-time rainfall forecast data (e.g., hourly rainfall in the next 24 hours, in millimeters / hour), river network water level data (e.g., real-time water level at key sections, in meters), pipeline drainage data (e.g., flow rate at main drainage outlets, in cubic meters / second), and external river tide data (e.g., real-time tide level of external rivers, in meters).

[0087] The simulation results of the coupled hydrodynamic model module will serve as the basic data for the gate and pump scheduling optimization model, providing high-precision predictions of the physical world state. The model will take the opening time and degree of opening of the gate (e.g., gate opening angle, in degrees or opening height, in meters) and the start-up time and operating power of the pump (e.g., number of pump stations in operation or pump power, in kilowatts) as decision variables to be optimized.

[0088] The multi-objective weight adaptive module identifies the current water situation based on real-time rainfall intensity and external river tide data, and dynamically adjusts the weight coefficients of flood control and drainage objectives. For example, when the external river high tide level overlaps with the inland river heavy rainfall, the flood control safety weight will be significantly increased, and the system will prioritize controlling the inland river water level to prevent the external river water from flowing back into the water. Specifically, the system comprehensively considers multiple important objectives and assigns corresponding weights: flood control objective, minimizing the risk of river and pipe network water levels exceeding warning levels to ensure regional flood control safety; drainage objective, reducing the depth and duration of water accumulation in the region to avoid urban flooding disasters; water resource utilization objective, rationally storing and utilizing water resources to improve water resource utilization efficiency. The formulation of the scheduling plan must strictly comply with various physical and operational constraints, including physical constraints of gates and pumps (the opening angle of sluice gates, the operating power of pumps, etc., must not exceed their design range and upper and lower limits), water balance constraints (the inflow and outflow of water in the entire region must be balanced, and changes in the water storage capacity of water storage facilities must be considered), and water level constraints (the water level of rivers and pipe networks must be controlled within a safe range, neither too high to cause flood risk nor too low to affect normal function).

[0089] The system calculates the scalar fitness value of each physically feasible candidate scheduling scheme based on the predicted scheduling state and dynamic weight coefficients through a weighted summation. This fitness value comprehensively evaluates the merits of the scheme in the current scenario. Subsequently, the algorithm enters the particle update phase. Each candidate scheduling scheme is treated as a particle, and its velocity and position in the solution space are updated based on the scheme with the best fitness value in its own historical iterations (individual optimal) and the scheme with the best fitness value in the historical iterations of the entire particle population (global optimal). This generates a new generation of better candidate scheduling schemes. Moreover, in each iteration, it can dynamically adjust the feasible region constraints of its search based on real-time received external working condition information (such as water pump failure) and check and repair the newly generated particles (candidate schemes).

[0090] Finally, after meeting the preset termination conditions (e.g., reaching the maximum number of iterations or fitness convergence), the system will initiate a final physical verification and confirmation process. This process submits key candidate solutions to a high-fidelity physical model for validation, resulting in a reliable optimal scheduling scheme that is both efficiently found and accurately verified, minimizing the flooded area and controlling water levels at key sections within a safe range. Simultaneously, during the optimization process, additional constraints such as pump station operating energy consumption and sluice gate operation frequency can be incorporated into the fitness function or used as criteria for selecting feasible solutions, ensuring the comprehensiveness and feasibility of the final solution.

[0091] After the optimal dispatch plan is generated, the system will push detailed dispatch instructions to the flood control command center in real time. Command center personnel can make decisions based on the system's suggestions and the actual situation on site, and execute dispatch instructions through the remote control system. At the same time, the system provides a visual interface to display real-time water conditions, predicted inundation maps, and comparative effects of different dispatch plans to assist decision-makers in making judgments.

[0092] It should be emphasized that the specific embodiments described herein are merely examples illustrating the core ideas of the present invention, and not limitations thereof. For those skilled in the art, any modifications, combinations, or equivalent substitutions made to the above embodiments without departing from the core principles and spirit disclosed in the present invention should fall within the protection scope claimed by the present invention. The final protection scope of the present invention should be determined by the contents of the appended claims.

Claims

1. A plain river network dike area flood control and waterlogging elimination multi-objective particle swarm optimization scheduling system, characterized in that, The application relates to a flood control system, which comprises the following modules: a coupling water dynamics model module, which calculates water flow exchange between a one-dimensional water dynamics model for simulating water flow in an underground pipe network and a two-dimensional water dynamics model for simulating surface flow; an agent model generation module, which calls the coupling water dynamics model module to generate a training data set through multiple simulations, trains the training data set, and generates a neural network agent model; wherein the agent model generation module adopts a staged training method to generate the neural network agent model, uses simple working condition data to preliminarily train the model, and uses complex working condition data to further train the preliminarily trained model; wherein an input layer of the neural network agent model receives a scheduling scheme vector composed of water gate opening degrees and pump station start-stop states, and an output layer outputs a multi-dimensional predicted scheduling state vector composed of the highest water level of key sections in a region and waterlogging submerged area; a multi-target weight self-adaptive module, which dynamically determines a group of weight coefficients representing the relative importance of multiple scheduling targets according to a received real-time early warning level, through linear interpolation calculation by querying a scenario-strategy mapping rule library; a particle swarm optimization scheduling module, which generates an initial candidate scheduling scheme as an initial population of a particle swarm based on topological structure information of a plain river network and preset heuristic rules; in an iterative optimization process, the feasible region constraint of the optimization algorithm is dynamically adjusted according to real-time received external working condition information, the neural network agent model is used as an adaptive function evaluation engine of the particle swarm algorithm, a candidate scheduling scheme is input into the neural network agent model, and a predicted scheduling state is output; based on the predicted scheduling state and the weight coefficients, the adaptive value of each candidate scheduling scheme is calculated, and the optimal scheduling scheme is output through iterative optimization of the particle swarm algorithm. The coupling water dynamics model module calculates water flow exchange between a one-dimensional water dynamics model for simulating water flow in an underground pipe network and a two-dimensional water dynamics model for simulating surface flow, and comprises the following: the one-dimensional water dynamics model is a storm flood management model based on Saint-Venant equation groups; the two-dimensional water dynamics model is a flood flow model based on two-dimensional unsteady shallow water equations; the calculation method of the water flow exchange includes: vertical exchange calculation, which calculates overflow between underground pipe network nodes and two-dimensional surface grids due to water head difference through an orifice flow formula and a weir formula; lateral exchange calculation, which takes the river bank of the one-dimensional water dynamics model as the internal boundary of the two-dimensional water dynamics model, and calculates lateral flow generated when the river water level exceeds the bank elevation through a weir formula. The agent model generation module comprises the following: calling the coupling water dynamics model module to simulate working conditions covering different rainfall scenarios and scheduling scheme combinations to generate a training data set; the neural network agent model comprises an input layer, a hidden layer and an output layer; the input layer receives a scheduling scheme vector composed of water gate opening degrees and pump station start-stop states; and the output layer outputs a multi-dimensional predicted scheduling state vector composed of the highest water level of key sections in a region and waterlogging submerged area. ​ ​ 2. The flood control and waterlogging relief multi-objective particle swarm optimization scheduling system for a plain river network basin according to claim 1, characterized in that, ​ 3. The flood control and waterlogging relief multi-objective particle swarm optimization scheduling system for a plain river network catchment according to claim 1, characterized in that, ​ 4. The flat river network dike area flood control and waterlogging elimination multi-objective particle swarm optimization scheduling system according to claim 1, characterized in that, The multi-objective weight adaptive module comprises a scenario-strategy mapping rule library stored in the multi-objective weight adaptive module; each scenario is represented by an early warning level defined by real-time rainfall intensity and river tide data; each strategy corresponds to a set of weight coefficient combinations preset for the two targets of flood control and waterlogging drainage; real-time rainfall intensity and real-time river tide data representing the current early warning level are received from outside; the received real-time data are matched with the multiple preset scenarios in the rule library one by one to determine the interval in which the current real-time data fall, which is defined by the multiple closest preset scenarios; when the real-time data exactly match a preset scenario, the module directly outputs the preset strategy weight corresponding to the scenario; when the real-time data fall between two preset scenarios, the module generates and outputs a set of dynamic weight coefficients applicable to the current scenario by performing linear interpolation calculation on the preset strategy weights corresponding to the two preset scenarios.

5. The flat river network dike area flood control and waterlogging elimination multi-objective particle swarm optimization scheduling system according to claim 1, characterized in that, The particle swarm optimization scheduling module generates an initial candidate scheduling scheme as the initial population of the particle swarm based on the topological structure information of the plain river network and preset heuristic rules, comprising: obtaining topological structure information of a plain river network represented by nodes and edges and containing upstream and downstream relationships of water conservancy facilities; applying preset heuristic rules to perform path traversal and priority sorting on the topological structure; the heuristic rules include a spatial priority rule for determining the spatial execution order of scheduling operations according to the upstream and downstream relationships in the topological structure; the heuristic rules also include a time priority rule for determining the time sequence of scheduling operations according to the division of the main stream and the branch stream in the topological structure; based on the determined spatial execution order and time sequence, a set of candidate scheduling schemes with random perturbations in specific operation parameters and conforming to the heuristic rules as a whole are generated as the initial population.

6. The flat river network dike area flood control and waterlogging elimination multi-objective particle swarm optimization scheduling system according to claim 1, characterized in that, In the iterative optimization process, the feasible region constraint of the optimization algorithm is dynamically adjusted according to the real-time received external working condition information, the candidate scheduling scheme is input into the neural network proxy model, and the predicted scheduling state is output, comprising: receiving external working condition information including real-time fault alarm information of water conservancy facilities, and dynamically adjusting the value range of the corresponding decision variable in the optimization algorithm as the feasible region constraint; checking the candidate scheduling scheme newly generated by the particle swarm algorithm, and performing boundary repair processing on the scheme that exceeds the feasible region constraint; taking the neural network proxy model as the fitness function evaluation engine of the particle swarm algorithm, taking the candidate scheduling scheme generated by the particle swarm algorithm in each iteration as the input vector, and transmitting it to the input layer of the neural network proxy model; performing a forward propagation calculation by the neural network proxy model to replace calling the coupled hydrodynamic model to perform a complete and time-consuming physical simulation; and outputting a multi-dimensional physical state prediction scheduling state vector representing the highest water level of the key section and the waterlogging submerged area in the region under the candidate scheduling scheme from the output layer of the neural network proxy model.

7. The flat river network dike area flood control and waterlogging relief multi-objective particle swarm optimization scheduling system according to claim 1, characterized in that, Based on the predicted scheduling state and the weight coefficient, a fitness value of each candidate scheduling scheme is calculated, and an optimal scheduling scheme is output through iteration optimization of the particle swarm algorithm, including: the weight coefficient is weighted and summed with the predicted scheduling state vector to calculate a scalar fitness value of each candidate scheduling scheme; each candidate scheduling scheme is taken as a particle, and the speed and position of the particle in the solution space are updated according to the scheme corresponding to the optimal fitness value in the historical iteration of the particle itself and the scheme corresponding to the optimal fitness value in the historical iteration of the whole particle population, to generate a new candidate scheduling scheme; after a preset termination condition is met, the scheme corresponding to the optimal fitness value in the historical iteration of the whole particle population is output as the optimal scheduling scheme.

Images