Integrated flood control optimization scheduling method based on multi-level hydraulic engineering

Abstract

The invention discloses an integrated flood control optimization scheduling method based on multi-level hydraulic engineering, and belongs to the technical field of hydraulic engineering. The method comprises the following steps: acquiring real-time multi-source monitoring data, and carrying out preprocessing and quality control on the data; constructing a hydraulic engineering group dynamic model based on a graph neural network, and capturing dynamic characteristics and state evolution of a hydraulic engineering group through a message passing mechanism; constructing a three-layer deep reinforcement learning framework; the scheduling strategy subjected to multi-layer optimization is interpreted into a gate opening degree and action time sequence instruction, and multiple safety verification is implemented to ensure the feasibility of actual control; the global flood control effect, the local scheduling performance and the control action rationality are quantified respectively, and a final reward signal is formed through dynamic weight adaptive combination; and visualizing the key decision factor by using the attention weight, displaying a decision reasoning process through a causal graph, and automatically generating a scheduling scheme analysis report.

Description

Technical Field

[0001] The present application relates to the technical field of water conservancy projects, and more specifically, to an integrated flood control optimization scheduling method based on multi-level water conservancy projects. Background Art

[0002] Traditional flood control scheduling methods often rely on empirical rules and a single data source. This approach is prone to problems such as low scheduling efficiency, delayed response, and insufficient scheduling optimization when faced with complex and changing hydrological and meteorological conditions. Especially in multi-reservoir systems, achieving a balance between global and local scheduling, improving water resource utilization efficiency, and ensuring flood control safety has become a pressing challenge.

[0003] In recent years, with the rapid development of artificial intelligence (AI) technologies, particularly graph neural networks and deep reinforcement learning (DL), an increasing number of researchers and engineers have begun exploring their applications in water conservancy scheduling systems. Graph neural networks can effectively capture the complex dynamic relationships between reservoirs within a hydropower project cluster, while DL can automatically optimize scheduling strategies through interactive learning with the environment, improving the system's adaptability and decision-making efficiency. While research has made some progress in single-reservoir scheduling, the coordinated scheduling of multi-level hydropower projects still faces challenges such as the coordination between global scheduling and local optimization, insufficient interpretability of the decision-making process, and inefficient information transfer between different levels.

[0004] To sum up, how to achieve coordinated flood control optimization scheduling based on multi-level water conservancy project groups through advanced intelligent algorithms and improve the intelligence, precision and explainability of scheduling has become a technical problem that needs to be solved urgently. Summary of the Invention

[0005] In order to overcome a series of defects in the prior art, the purpose of this application is to provide a flood control optimization scheduling method based on multi-level integrated water conservancy project to address the above problems, including the following steps:

[0006] Step 1: Acquire real-time multi-source monitoring data and perform data preprocessing and quality control;

[0007] Step 2: Build a dynamic model of the water conservancy project group based on the graph neural network, and capture the dynamic characteristics and state evolution of the water conservancy project group through the message passing mechanism;

[0008] Step 3: Build a three-layer deep reinforcement learning framework. The top layer uses the Transformer architecture to process global scheduling information, the middle layer uses a graph attention network to optimize regional scheduling strategies, and the bottom layer uses a deep Q-network to generate control parameters as the micro-decision basis for scheduling strategies.

[0009] Step 4: The multi-layer optimized scheduling strategy is interpreted into gate opening and action sequence instructions, and multiple safety checks are performed to ensure the feasibility of actual control.

[0010] Step 5: quantify the global flood control effect, local scheduling performance, and control action rationality respectively, and form the final reward signal through dynamic weight adaptive combination;

[0011] Step 6: Use attention weights to visualize key decision factors, display the decision reasoning process through a causal diagram, and automatically generate a scheduling plan analysis report.

[0012] Furthermore, step 1 includes the following steps:

[0013] Receive various monitoring data in real time through a variety of monitoring equipment, including: water level data collected by automatic reservoir water level monitoring instruments, inflow and outflow data obtained by flow monitoring equipment, rainfall data measured by automatic rain gauges, and weather forecast information provided by meteorological departments;

[0014] Synchronize and integrate data from different sources to ensure that each monitoring data has consistent timestamps and spatial correlations, thereby ensuring data consistency;

[0015] Convert data in different formats into a unified data format to reduce the additional complexity caused by format incompatibility;

[0016] Perform preliminary screening and cleaning of the acquired data, including removing missing values and processing abnormal values;

[0017] For water level, flow and rainfall data, smoothing algorithms and filtering techniques are applied to eliminate high-frequency noise interference and improve data stability.

[0018] Furthermore, step 2 includes the following steps:

[0019] Construct the basic structure of the dynamic model of the water conservancy project group, decomposing the water conservancy project group into hydraulic units, power units, and control units. Among them, the hydraulic unit describes the water flow movement characteristics, the power unit represents the energy conversion process, and the control unit reflects the project scheduling behavior;

[0020] The dynamic model is formally described based on a graph structure. Hydraulic units are mapped as nodes in the graph. Node attributes include water level, flow rate, and reservoir capacity variables, as well as the characteristics of regulating reservoir capacity and discharge capacity. At the same time, the connection relationship between hydraulic units is represented as edges in the graph. Edge attributes include hydraulic transmission delay and energy loss parameters, thus constructing a complete graph network structure.

[0021] Based on the principles of conservation of mass and energy, a mathematical model of the node's dynamic characteristics is designed, and a set of differential equations describing the state evolution is established to accurately reflect the physical characteristics of the water conservancy project;

[0022] Dynamically transmit state information between nodes through graph neural networks, and design an adaptive message aggregation function so that each node can integrate upstream and downstream influences to achieve node status updates;

[0023] The numerical integration method is used to solve the differential equations, and the solution efficiency is improved through the adaptive step size control algorithm.

[0024] Furthermore, the dynamic model of the water conservancy project group is expressed as follows: Among them, K p , K i , K d They are proportional, integral, and differential gain coefficients, respectively, used for PID control; V hyd (t) is the water level of the current hydraulic unit; V target is the target water level to be achieved; Q in (t) is the flow rate of water into the reservoir, which is related to the upstream water flow and precipitation factors; Q out (t) represents the amount of water flowing out of the reservoir, which is controlled by reservoir operation, flood discharge, and power generation factors; is the time derivative; represents the cumulative effect from time 0 to t; V hyd (τ) represents the amount of water in the reservoir at time point τ, which is used for integral calculation and error analysis; dτ is a small increment of the historical time variable τ. By integrating τ, the reservoir water volume V from time 0 to the current time t is calculated. hyd (τ) and target water volume V target The cumulative error between t and dt is the small increment of the time variable t.

[0025] Furthermore, the node status update formula is: in, For node v i The state at time step T+1; σ is the activation function; b is the bias term; W is the weight matrix, which adjusts the fusion method of node state and message aggregation results; Δh ij is the dynamic change difference between upstream and downstream nodes; β ij is the adjustment coefficient of upstream and downstream influence between nodes; α ij For node v i and neighbor node v j Adaptive message passing weights between For node v i The state at time step T; For node v i The set of neighbor nodes of For node v jThe state at time step T.

[0026] Furthermore, step 3 includes the following steps:

[0027] A top-level Transformer architecture is built to encode the historical operation data, current status, and forecast information of the entire basin's reservoir group into a sequence input. The multi-head self-attention mechanism captures long-term spatiotemporal dependencies and outputs a global operation strategy.

[0028] Design a graph attention network in the middle layer, and learn the dynamic influence relationship between nodes by setting a multi-layer graph attention mechanism to optimize the coordinated scheduling strategy of the reservoir group in the region;

[0029] Build the underlying deep Q network, design the mapping relationship between state space and action space for each reservoir, and output specific control instructions;

[0030] Design a hierarchical training mechanism and use hierarchical reinforcement learning methods to build a hierarchical learning framework;

[0031] An information transmission channel is designed to achieve two-way interaction between layers. The decisions of the upper layer guide the behavior of the lower layer by layer decomposition, and the execution feedback of the lower layer helps the upper layer optimize the decision by layer-by-layer aggregation, forming a complete hierarchical learning closed loop.

[0032] Furthermore, step 4 includes the following steps:

[0033] According to the hydraulic structure characteristic curve and hydraulic relationship, the scheduling strategy is converted into specific gate opening instructions; at the same time, based on the current water conditions and weather forecast information, the execution sequence of each control action is reasonably arranged;

[0034] Conduct compliance checks on generated gate control instructions, including gate opening limits, opening and closing rate constraints, and flow gradient verification, to screen out potential violations and make corrections.

[0035] Input gate control instructions into the dynamic model of the water conservancy project group for rolling simulation, predict the response after the instruction is executed, evaluate the dynamic impact of the control instruction, and make preventive adjustments to possible risk points;

[0036] Design alternative control schemes and emergency response procedures for possible extreme situations and equipment failures, so that rapid switching can be carried out when an abnormality is detected;

[0037] Implement a step-by-step execution strategy, break down large-scale regulatory instructions into multiple small steps and implement them gradually. Evaluate the effects after each step is executed, and only execute the next step after confirming safety, thereby reducing the risks brought about by large-scale regulation.

[0038] Furthermore, step 5 includes the following steps:

[0039] A hierarchical scoring mechanism is established based on the risk of exceeding water levels in important control sections of the basin, the flow rate of key river sections, the flood peak reduction rate and the matching degree of flood storage and discharge time sequence;

[0040] A performance calculation method based on fuzzy comprehensive evaluation is established for each region's water supply guarantee rate, power generation efficiency, and ecological base flow guarantee targets. This method dynamically evaluates the overall effectiveness of regional scheduling by tracking the completion of various scheduling indicators in real time, and adaptively adjusts the weight of each indicator based on the current water situation and task importance.

[0041] Evaluate the execution cost of control instructions from the perspectives of equipment life, energy consumption level, and scheduling stability;

[0042] In the deep reinforcement learning framework, a credit allocation algorithm is used to reasonably distribute the rewards of long-term control effects to each decision moment, so as to correctly evaluate the contribution of each control action to the final goal.

[0043] By analyzing the control effects under different working conditions, the weight ratios of the three levels of global flood control, local scheduling and control rationality are automatically adjusted to achieve a dynamic balance in reward evaluation.

[0044] Furthermore, the overall efficiency of regional scheduling is expressed as: C(r) = w s (r)·μ s (x s (r))+w G (r)·μ G (x G (r))+w E (r)·μ E (w E (r)), where G(r) is the dispatch efficiency evaluation at time r, which represents the comprehensive performance of the entire regional dispatch and reflects the completion of water supply, power generation and ecological goals; w s (r) is the weight of the water supply target at time r; w G (r) is the weight of the power generation efficiency target at time r; w E (r) is the weight of the ecological base flow protection target at time r; x s (r) represents the real-time evaluation data of the water supply guarantee rate at time r; μ s (x s (r)) is the membership function of the water supply target, which indicates the completion of the water supply guarantee rate target; x G (r) represents the real-time evaluation data of power generation efficiency at time r; μ G (x G (r)) is the membership function of the power generation benefit target, which indicates the completion of the power generation target; x E(r) represents the real-time assessment data of ecological base flow protection at time r; μ E (x E (r)) is the membership function of the ecological base flow protection target, which indicates the completion of the ecological target.

[0045] Furthermore, step 6 includes the following steps:

[0046] Extract weight matrices at multiple time scales from the Transformer layer and graph attention layer, and convert them into dynamic heatmap displays;

[0047] The complete reasoning chain from environmental status, engineering constraints to control strategy is formalized as a directed acyclic graph, with node size representing factor importance, line thickness representing influence intensity, and arrow direction indicating causal relationship, enabling intuitive and interpretable display of the decision-making process.

[0048] Use a variety of visual methods to demonstrate the degree of achievement of flood control safety, water supply guarantee rate and power generation efficiency indicators;

[0049] Automatically generate structured reports including water situation analysis, plan description, expected results and risk warnings based on key data and decision-making basis of the scheduling process;

[0050] Establish interactive exploration capabilities to support decision makers in conducting in-depth analysis of specific scheduling nodes.

[0051] Compared with the prior art, this application has the following beneficial effects:

[0052] This application obtains real-time multi-source monitoring data for preprocessing and quality control, builds a dynamic model of the water conservancy project group based on graph neural network and captures dynamic characteristics through message passing mechanism, optimizes the scheduling strategy by combining the three-layer deep reinforcement learning framework, interprets it into gate opening and action timing instructions, and quantifies the flood control effect and dynamically adjusts the weight, ultimately realizing a global flood control optimization scheduling plan for the water conservancy project group. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flow chart of a method for optimizing flood control and scheduling based on integrated multi-level water conservancy projects disclosed in an embodiment of the present application. DETAILED DESCRIPTION

[0054] To make the objectives, technical solutions, and advantages of the present invention more apparent, the technical solutions in the embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Throughout the drawings, identical or similar reference numerals represent identical or similar elements or elements having identical or similar functions. The described embodiments are only some, not all, of the embodiments of the present invention.

[0055] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0056] The embodiments and directional terms described below with reference to the accompanying drawings are exemplary and intended to be used to explain the present invention, but should not be construed as limiting the present invention.

[0057] like Figure 1 As shown, a flood control optimization scheduling method based on multi-level integrated water conservancy project includes the following steps:

[0058] Step 1: Acquire real-time multi-source monitoring data and perform data preprocessing and quality control;

[0059] Step 2: Build a dynamic model of the water conservancy project group based on the graph neural network, and capture the dynamic characteristics and state evolution of the water conservancy project group through the message passing mechanism;

[0060] Step 3: Build a three-layer deep reinforcement learning framework. The top layer uses the Transformer architecture to process global scheduling information, the middle layer uses a graph attention network to optimize regional scheduling strategies, and the bottom layer uses a deep Q-network to generate control parameters as the micro-decision basis for scheduling strategies.

[0061] Step 4: The multi-layer optimized scheduling strategy is interpreted into gate opening and action sequence instructions, and multiple safety checks are performed to ensure the feasibility of actual control.

[0062] Step 5: quantify the global flood control effect, local scheduling performance, and control action rationality respectively, and form the final reward signal through dynamic weight adaptive combination;

[0063] Step 6: Use attention weights to visualize key decision factors, display the decision reasoning process through a causal diagram, and automatically generate a scheduling plan analysis report.

[0064] Step 1 involves acquiring data from various sources, including various monitoring data such as water levels, flow rates, precipitation, and meteorological data. This data is often diverse, time-sensitive, and uncertain, making preprocessing and quality control crucial. Preprocessing involves data cleaning, denoising, and interpolation to ensure that the input data meets high-quality standards. Quality control verifies the rationality of the data, identifies outliers and errors, and corrects any discrepancies. This step improves data quality, ensuring that subsequent model training and decision-making processes are based on authentic, high-quality information, effectively enhancing the accuracy of flood control operations.

[0065] In step 2, a graph neural network (GNN) models a cluster of water conservancy projects using a graph structure, effectively capturing the spatiotemporal correlations between various water conservancy facilities. Through a message-passing mechanism, GNNs can propagate information within the cluster, capturing the dynamic characteristics and state evolution of each facility. Specifically, graph neural networks can reveal the interconnectedness of different water conservancy facilities and how they adjust in a changing environment when simulating water flow distribution and facility scheduling decisions. This technology allows the overall status of a cluster of water conservancy projects to be tracked and predicted in real time, enabling the scheduling system to better respond to complex hydrological and meteorological changes and improving the intelligence of flood control scheduling.

[0066] In step 3, deep reinforcement learning (DRL), a self-learning technology, can optimize decisions based on feedback in a dynamically changing environment. Step 3 optimizes the scheduling strategy of water conservancy projects by constructing a three-layer deep reinforcement learning framework. The top layer uses the Transformer architecture, which focuses on processing global scheduling information and can effectively integrate various scheduling information across a large area to ensure global coordination. The middle layer uses the Graph Attention Network (GAT) to optimize regional scheduling strategies, improving the accuracy of local scheduling by weighting the attention paid to different regions. The bottom layer uses the Deep Q Network (DQN) to generate specific control parameters as the micro-decision-making basis for the scheduling strategy. The technical effect of this three-layer architecture is that it hierarchically optimizes the scheduling strategy, which can not only handle macro scheduling problems, but also finely resolve the scheduling details of local areas, thereby improving the overall scheduling efficiency and effectiveness.

[0067] Step 4 translates the scheduling strategy generated by the deep reinforcement learning framework into specific gate opening and action sequence instructions for actual execution. This process not only requires consideration of the optimized scheduling plan but also requires multiple safety checks to ensure the feasibility and safety of actual operation. Safety check mechanisms can detect and prevent operations that do not meet safety standards, thereby avoiding potential dangers and failures. The technical benefit of this step is to ensure the implementation of the scheduling plan through precise instruction execution, maximize flood control effectiveness, and ensure safety in actual control.

[0068] To ensure the effectiveness and accuracy of scheduling decisions, step 5 forms a comprehensive evaluation index by quantifying the global flood control effect, local scheduling performance, and the rationality of control actions. By quantifying the flood control effect, the system can evaluate the flood control effect of each scheduling decision on a global scale; by quantifying the local scheduling performance, the scheduling efficiency of the local area is optimized; at the same time, the quantification of the rationality of the control action is to evaluate the rationality of each specific operation. These quantitative results will be adaptively combined into the final reward signal based on dynamic weights, driving the system to continuously self-optimize and improve. The technical effect of this step is to achieve a comprehensive evaluation of the scheduling effect, and promote decision optimization through quantitative feedback, so that the system has good self-learning and self-adaptive capabilities.

[0069] Step 6 uses the attention mechanism to visualize key decision factors, helping decision makers understand the model's rationale. By displaying the decision-making reasoning process through a causal diagram, it clearly demonstrates how various factors influence the scheduling results, providing a transparent and reliable basis for decision-making. Furthermore, it automatically generates analysis reports for scheduling plans, helping decision makers better evaluate and adjust flood control scheduling strategies. The technical benefits of this step are to improve interpretability and transparency, enhance user confidence in decision-making, and provide strong support for subsequent policy optimization.

[0070] Through the collaborative work of these six steps, this integrated flood control optimization and scheduling method based on multi-level water conservancy projects achieves comprehensive automation and intelligent management from data collection, modeling, decision optimization, to execution and evaluation. Advanced technical means are used at each stage to ensure efficient, accurate, and safe flood control scheduling. In particular, the combined application of reinforcement learning and graph neural networks enables rapid response and optimized decision-making in dynamically changing environments, demonstrating strong adaptability and real-time performance. Overall, the implementation of this method will greatly enhance the flood control capabilities of water conservancy projects, reduce human intervention, ensure that water conservancy projects can perform optimally under complex and changing hydrological and meteorological conditions, and protect people's lives and property.

[0071] Furthermore, step 1 includes the following steps:

[0072] Receive various monitoring data in real time through a variety of monitoring equipment, including: water level data collected by automatic reservoir water level monitoring instruments, inflow and outflow data obtained by flow monitoring equipment, rainfall data measured by automatic rain gauges, and weather forecast information provided by meteorological departments;

[0073] Synchronize and integrate data from different sources to ensure that each monitoring data has consistent timestamps and spatial correlations, thereby ensuring data consistency;

[0074] Convert data in different formats into a unified data format to reduce the additional complexity caused by format incompatibility;

[0075] Perform preliminary screening and cleaning of the acquired data, including removing missing values and processing abnormal values;

[0076] For water level, flow and rainfall data, smoothing algorithms and filtering techniques are applied to eliminate high-frequency noise interference and improve data stability.

[0077] The core of Step 1 lies in being able to obtain multi-dimensional, real-time monitoring data from various sources and process this data to meet the requirements of subsequent analysis and decision-making. First, the real-time monitoring data for water conservancy projects comes from a wide range of sources, covering a variety of information such as water level, flow, rainfall, and weather forecasts. Reservoir water level monitors, flow monitoring equipment, rain gauges, and weather forecast data from meteorological departments each have their own unique collection methods and characteristics. Through these devices, reservoir water level data, inflow and outflow data, precipitation information, and upcoming weather conditions can be obtained in real time. This diverse data provides a rich source of information for flood control operations, reflecting the current hydrological and meteorological conditions and ensuring that scheduling decisions are based on accurate real-time data.

[0078] Data integration is a key step in step 1. Data generated by different monitoring devices usually have different timestamps and spatial correlations. How to synchronize and integrate this data into a unified time series is the key to ensuring data reliability. First, time alignment technology is needed to ensure that all types of data correspond to the same time point to avoid data distortion due to time differences. Spatial correlation mainly refers to the unified organization of data from different locations according to geographical location to ensure that the data can be correctly matched, which is crucial for accurately reflecting the hydrological changes in the watershed. The integrated data needs to ensure consistency to avoid deviations between different data sources, thereby providing an accurate basis for subsequent analysis.

[0079] After ensuring data synchronization and consistency, a unified data format conversion is required. Data generated by different monitoring devices may differ in format. For example, water level data uses CSV format, flow data uses JSON format, and meteorological data is returned using an API interface. To reduce the additional complexity caused by format incompatibility, all data needs to be converted to a unified standard format, typically a numerical matrix or time series format, to facilitate subsequent data processing and model training. In this process, data format conversion helps improve versatility and compatibility while reducing errors or mismatches caused by inconsistent formats.

[0080] Data cleaning and screening are crucial steps in ensuring data quality. During long-term operation, monitoring equipment can experience sensor failure, data loss, or environmental interference, leading to missing or outliers in the collected data. To eliminate these interferences, data must first undergo preliminary screening and cleaning to remove unexpected and anomalous data. Missing values are typically addressed through interpolation, filling in the gaps and ensuring smooth and continuous data. Outliers are identified through statistical analysis or model-based detection methods and corrected or removed based on the actual situation. This step is crucial for improving data accuracy and reliability, directly impacting the precision of subsequent modeling and decision-making.

[0081] Finally, for water level, flow, and rainfall data, smoothing algorithms and filtering techniques are applied to eliminate high-frequency noise interference. Since monitoring equipment is often affected by environmental noise, the data may contain high-frequency fluctuations, which will interfere with the stability and accuracy of the data. By applying smoothing algorithms (such as moving averages) and filtering techniques (such as Kalman filtering or low-pass filtering), these irrelevant fluctuations can be effectively removed, making the data smoother and more stable. The smoothed data can better reflect the long-term trends in the hydrological process without being disturbed by short-term noise, thereby improving the effectiveness of flood control scheduling.

[0082] The data collection and preprocessing in Step 1 laid a solid foundation for the entire flood control optimization process. Real-time acquisition and integration of multi-source data ensured data consistency and reliability; format conversion and data cleaning improved data usability; and smoothing algorithms and filtering techniques removed noise interference and enhanced data stability. The application of these technical approaches ensured that subsequent model training and decision-making could be based on high quality, ultimately providing accurate and real-time support for flood control scheduling and facilitating efficient and scientific management of water conservancy projects in complex environments.

[0083] Furthermore, step 2 includes the following steps:

[0084] Construct the basic structure of the dynamic model of the water conservancy project group, decomposing the water conservancy project group into hydraulic units, power units, and control units. Among them, the hydraulic unit describes the water flow movement characteristics, the power unit represents the energy conversion process, and the control unit reflects the project scheduling behavior;

[0085] The dynamic model is formally described based on a graph structure. Hydraulic units are mapped as nodes in the graph. Node attributes include water level, flow rate, and reservoir capacity variables, as well as the characteristics of regulating reservoir capacity and discharge capacity. At the same time, the connection relationship between hydraulic units is represented as edges in the graph. Edge attributes include hydraulic transmission delay and energy loss parameters, thus constructing a complete graph network structure.

[0086] Based on the principles of conservation of mass and energy, a mathematical model of the node's dynamic characteristics is designed, and a set of differential equations describing the state evolution is established to accurately reflect the physical characteristics of the water conservancy project;

[0087] Dynamically transmit state information between nodes through graph neural networks, and design an adaptive message aggregation function so that each node can integrate upstream and downstream influences to achieve node status updates;

[0088] The numerical integration method is used to solve the differential equations, and the solution efficiency is improved through the adaptive step size control algorithm.

[0089] In summary, Step 2 fully utilizes the graph structure to express the complex relationships between hydraulic units by constructing a dynamic model of a water conservancy project cluster based on a graph neural network. It then accurately describes the physical characteristics of the water conservancy project through a system of differential equations and numerical methods. In this process, the graph neural network not only enables efficient information transmission between nodes but also improves the model's flexibility and solution efficiency through an adaptive message aggregation mechanism and step-size control algorithm. Overall, the design of Step 2 enables the dynamic model of the water conservancy project cluster to reflect the dynamic state of the project group in real time under variable hydrological and meteorological conditions, providing a precise simulation foundation for subsequent scheduling optimization and decision-making.

[0090] Furthermore, the dynamic model of the water conservancy project group is expressed as follows: Among them, K p , K i , K d They are proportional, integral, and differential gain coefficients, respectively, used for PID control; V hyd (t) is the water level of the current hydraulic unit; V target is the target water level to be achieved; Q in (t) is the flow rate of water into the reservoir, which is related to the upstream water flow and precipitation factors; Q out (t) represents the amount of water flowing out of the reservoir, which is controlled by reservoir operation, flood discharge, and power generation factors; is the time derivative; represents the cumulative effect from time 0 to t; V hyd (τ) represents the amount of water in the reservoir at time point τ, which is used for integral calculation and error analysis; dτ is a small increment of the historical time variable τ. By integrating τ, the reservoir water volume V from time 0 to the current time t is calculated. hyd (τ) and target water volume V target The cumulative error between t and dt is the small increment of the time variable t.

[0091] In summary, by incorporating mathematical formulas and PID control, the dynamic model of the hydraulic engineering complex can accurately simulate changes in reservoir water levels and regulate water flow based on real-time data. The settings for proportional, integral, and differential gain coefficients enable the model to not only respond to immediate changes but also effectively correct long-term errors, improving stability and responsiveness. This dynamic model enables more accurate and efficient reservoir operation, ensuring optimal flood control performance under complex hydrological and meteorological conditions.

[0092] Furthermore, the node status update formula is: in, For node v i The state at time step T+1; σ is the activation function; b is the bias term; W is the weight matrix, which adjusts the fusion method of node state and message aggregation results; Δh ij is the dynamic change difference between upstream and downstream nodes; β ij is the adjustment coefficient of upstream and downstream influence between nodes; α ij For node v i and neighbor node v j Adaptive message passing weights between For node v i The state at time step T; For node v i The set of neighbor nodes of For node v j The state at time step T.

[0093] In summary, the node state update formula, through carefully designed parameters and mechanisms, enables state updates and optimized scheduling for each reservoir in a water conservancy project cluster. By weightedly aggregating information from neighboring nodes, combined with adaptive message passing weights and nonlinear activation functions, it dynamically adjusts reservoir operation strategies and achieves efficient flood control scheduling. This formula fully leverages the advantages of graph neural networks, capturing the complex relationships between reservoirs and making precise scheduling decisions based on real-time data, thereby improving the scheduling efficiency and flood control effectiveness of the water conservancy project cluster.

[0094] Furthermore, step 3 includes the following steps:

[0095] A top-level Transformer architecture is built to encode the historical operation data, current status, and forecast information of the entire basin's reservoir group into a sequence input. The multi-head self-attention mechanism captures long-term spatiotemporal dependencies and outputs a global operation strategy.

[0096] Design a graph attention network in the middle layer, and learn the dynamic influence relationship between nodes by setting a multi-layer graph attention mechanism to optimize the coordinated scheduling strategy of the reservoir group in the region;

[0097] Build the underlying deep Q network, design the mapping relationship between state space and action space for each reservoir, and output specific control instructions;

[0098] Design a hierarchical training mechanism and use hierarchical reinforcement learning methods to build a hierarchical learning framework;

[0099] An information transmission channel is designed to achieve two-way interaction between layers. The decisions of the upper layer guide the behavior of the lower layer by layer decomposition, and the execution feedback of the lower layer helps the upper layer optimize the decision by layer-by-layer aggregation, forming a complete hierarchical learning closed loop.

[0100] In summary, step 3 decomposes the water conservancy project scheduling problem into three levels: global scheduling, regional scheduling, and local control, by building a three-layer deep reinforcement learning framework. The top-level Transformer architecture captures global spatiotemporal dependencies through a multi-head self-attention mechanism. The middle-level graph attention network optimizes the regional coordinated scheduling of the reservoir cluster, and the bottom-level deep Q network generates specific control instructions. A layered training mechanism and information transmission channels ensure coordination and interaction between each layer, forming a closed-loop optimization process. This architecture enables precise and efficient reservoir scheduling under complex hydrological and meteorological conditions, optimizing flood control effectiveness.

[0101] Furthermore, step 4 includes the following steps:

[0102] According to the hydraulic structure characteristic curve and hydraulic relationship, the scheduling strategy is converted into specific gate opening instructions; at the same time, based on the current water conditions and weather forecast information, the execution sequence of each control action is reasonably arranged;

[0103] Conduct compliance checks on generated gate control instructions, including gate opening limits, opening and closing rate constraints, and flow gradient verification, to screen out potential violations and make corrections.

[0104] Input gate control instructions into the dynamic model of the water conservancy project group for rolling simulation, predict the response after the instruction is executed, evaluate the dynamic impact of the control instruction, and make preventive adjustments to possible risk points;

[0105] Design alternative control schemes and emergency response procedures for possible extreme situations and equipment failures, so that rapid switching can be carried out when an abnormality is detected;

[0106] Implement a step-by-step execution strategy, break down large-scale regulatory instructions into multiple small steps and implement them gradually. Evaluate the effects after each step is executed, and only execute the next step after confirming safety, thereby reducing the risks brought about by large-scale regulation.

[0107] In summary, step 4, through a series of refined operations, ensures the safe and effective execution of the scheduling strategy in actual control. After converting the scheduling strategy into gate control instructions, compliance checks, rolling simulations, and risk assessments further ensure the feasibility and safety of the control instructions. At the same time, alternative plans and emergency response procedures are designed for possible extreme situations and equipment failures. By implementing the strategy in a step-by-step manner, the potential risks associated with large-scale control operations are reduced. The entire process fully considers the safety, reliability, and flexibility of the scheduling operations, effectively ensuring the smooth achievement of the scheduling objectives of the water conservancy project group.

[0108] Furthermore, step 5 includes the following steps:

[0109] A hierarchical scoring mechanism is established based on the risk of exceeding water levels in important control sections of the basin, the flow rate of key river sections, the flood peak reduction rate and the matching degree of flood storage and discharge time sequence;

[0110] A performance calculation method based on fuzzy comprehensive evaluation is established for each region's water supply guarantee rate, power generation efficiency, and ecological base flow guarantee targets. This method dynamically evaluates the overall effectiveness of regional scheduling by tracking the completion of various scheduling indicators in real time, and adaptively adjusts the weight of each indicator based on the current water situation and task importance.

[0111] Evaluate the execution cost of control instructions from the perspectives of equipment life, energy consumption level, and scheduling stability;

[0112] In the deep reinforcement learning framework, a credit allocation algorithm is used to reasonably distribute the rewards of long-term control effects to each decision moment, so as to correctly evaluate the contribution of each control action to the final goal.

[0113] By analyzing the control effects under different working conditions, the weight ratios of the three levels of global flood control, local scheduling and control rationality are automatically adjusted to achieve a dynamic balance in reward evaluation.

[0114] In summary, step 5 ensures the continuous optimization and efficient execution of the scheduling strategy through a series of dynamic evaluation and adjustment mechanisms. By establishing a hierarchical scoring mechanism, fuzzy comprehensive evaluation, and execution cost assessment, scheduling effectiveness can be monitored in real time, and the weights of various indicators can be adaptively adjusted based on factors such as water conditions and task importance. Furthermore, a credit allocation algorithm rationally evaluates the contribution of each control action, ultimately achieving a dynamic balance of rewards. These methods work together to ensure the stable and efficient operation of water conservancy scheduling in a complex and changing environment, maximizing its multiple objectives of flood control, water supply, power generation, and ecological protection.

[0115] Furthermore, the overall efficiency of regional scheduling is expressed as: C(r) = w s (r)·μ s(x s (r))+w G (r)·μ G (x G (r))+w E (r)·μ E (w E (r)), where G(r) is the dispatch efficiency evaluation at time r, which represents the comprehensive performance of the entire regional dispatch and reflects the completion of water supply, power generation and ecological goals; w s (r) is the weight of the water supply target at time r; w G (r) is the weight of the power generation efficiency target at time r; w E (r) is the weight of the ecological base flow protection target at time r; x s (r) represents the real-time evaluation data of the water supply guarantee rate at time r; μ s (x s (r)) is the membership function of the water supply target, which indicates the completion of the water supply guarantee rate target; x G (r) represents the real-time evaluation data of power generation efficiency at time r; μ G (x G (r)) is the membership function of the power generation benefit target, which indicates the completion of the power generation target; x E (r) represents the real-time assessment data of ecological base flow protection at time r; μ E (x E (r)) is the membership function of the ecological base flow protection target, which indicates the completion of the ecological target.

[0116] In summary, the overall effectiveness of regional scheduling is comprehensively evaluated through a weighted membership function to assess the achievement of the three objectives of water supply, power generation efficiency, and ecological baseflow protection. The weights of each objective are dynamically adjusted based on real-time water conditions and task priorities. This flexible and comprehensive evaluation mechanism provides real-time feedback on the effectiveness of scheduling strategies and allows for adaptive adjustments based on varying needs, ensuring a balance between multiple objectives such as water supply, power generation, and ecological protection while ensuring flood control. This evaluation method can effectively improve the flexibility, accuracy, and stability of regional scheduling.

[0117] Furthermore, step 6 includes the following steps:

[0118] Extract weight matrices at multiple time scales from the Transformer layer and graph attention layer, and convert them into dynamic heatmap displays;

[0119] The complete reasoning chain from environmental status, engineering constraints to control strategy is formalized as a directed acyclic graph, with node size representing factor importance, line thickness representing influence intensity, and arrow direction indicating causal relationship, enabling intuitive and interpretable display of the decision-making process.

[0120] Use a variety of visual methods to demonstrate the degree of achievement of flood control safety, water supply guarantee rate and power generation efficiency indicators;

[0121] Automatically generate structured reports including water situation analysis, plan description, expected results and risk warnings based on key data and decision-making basis of the scheduling process;

[0122] Establish interactive exploration capabilities to support decision makers in conducting in-depth analysis of specific scheduling nodes.

[0123] In summary, step 6 improves the transparency and explainability of the decision-making process through visualization and automatic report generation. Through heat maps, directed acyclic graphs, multi-dimensional visualizations, and structured reports, decision-makers can more intuitively understand the operational effectiveness of scheduling strategies and the underlying decision-making logic. This not only provides decision-makers with powerful data support but also enhances the traceability and explainability of the decision-making process, allowing complex water conservancy scheduling issues to be clearly presented and facilitated for timely adjustments and optimization. Interactive exploration further enhances flexibility and decision-making depth, providing more precise support for addressing scheduling challenges in different scenarios.

[0124] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art will appreciate that modifications may be made to the technical solutions described in the above embodiments, or that some of the technical features may be replaced with equivalents; such modifications or replacements do not deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A flood control optimization scheduling method based on multi-level integrated water conservancy project, characterized in that: The following steps are involved: Step 1: Acquire real-time multi-source monitoring data and perform data preprocessing and quality control; Step 2: Build a dynamic model of the water conservancy project group based on the graph neural network, and capture the dynamic characteristics and state evolution of the water conservancy project group through the message passing mechanism; Step 3: Build a three-layer deep reinforcement learning framework. The top layer uses the Transformer architecture to process global scheduling information, the middle layer uses a graph attention network to optimize regional scheduling strategies, and the bottom layer uses a deep Q-network to generate control parameters as the micro-decision basis for scheduling strategies. Step 4: The multi-layer optimized scheduling strategy is interpreted into gate opening and action sequence instructions, and multiple safety checks are performed to ensure the feasibility of actual control. Step 5: quantify the global flood control effect, local scheduling performance, and control action rationality respectively, and form the final reward signal through dynamic weight adaptive combination; Step 6: Use attention weights to visualize key decision factors, display the decision reasoning process through a causal diagram, and automatically generate a scheduling plan analysis report.

2. The method for optimizing flood control and scheduling based on the integration of multi-level water conservancy projects according to claim 1 is characterized in that: Step 1 includes the following steps: Receive various monitoring data in real time through a variety of monitoring equipment, including: water level data collected by automatic reservoir water level monitoring instruments, inflow and outflow data obtained by flow monitoring equipment, rainfall data measured by automatic rain gauges, and weather forecast information provided by meteorological departments; Synchronize and integrate data from different sources to ensure that each monitoring data has consistent timestamps and spatial correlations, thereby ensuring data consistency; Convert data in different formats into a unified data format to reduce the additional complexity caused by format incompatibility; Perform preliminary screening and cleaning of the acquired data, including removing missing values and processing abnormal values; For water level, flow and rainfall data, smoothing algorithms and filtering techniques are applied to eliminate high-frequency noise interference and improve data stability.

3. The method for optimizing flood control and scheduling based on the integration of multi-level water conservancy projects according to claim 1 is characterized in that: Step 2 includes the following steps: Construct the basic structure of the dynamic model of the water conservancy project group, decomposing the water conservancy project group into hydraulic units, power units, and control units. Among them, the hydraulic unit describes the water flow movement characteristics, the power unit represents the energy conversion process, and the control unit reflects the project scheduling behavior; The dynamic model is formally described based on a graph structure. Hydraulic units are mapped as nodes in the graph. Node attributes include water level, flow rate, and reservoir capacity variables, as well as the characteristics of regulating reservoir capacity and discharge capacity. At the same time, the connection relationship between hydraulic units is represented as edges in the graph. Edge attributes include hydraulic transmission delay and energy loss parameters, thus constructing a complete graph network structure. Based on the principles of conservation of mass and energy, a mathematical model of the node's dynamic characteristics is designed, and a set of differential equations describing the state evolution is established to accurately reflect the physical characteristics of the water conservancy project; Dynamically transmit state information between nodes through graph neural networks, and design an adaptive message aggregation function so that each node can integrate upstream and downstream influences to achieve node status updates; The numerical integration method is used to solve the differential equations, and the solution efficiency is improved through the adaptive step size control algorithm.

4. The method for optimizing flood control and scheduling based on the integration of multi-level water conservancy projects according to claim 3 is characterized in that: The dynamic model of the water conservancy project group is expressed as follows: Among them, K p , K i , K d They are proportional, integral, and differential gain coefficients, respectively, used for PID control; V hyd (t) is the water level of the current hydraulic unit; V target is the target water level to be achieved; Q in (t) is the flow rate of water into the reservoir, which is related to the upstream water flow and precipitation factors; Q out (t) represents the amount of water flowing out of the reservoir, which is controlled by reservoir operation, flood discharge, and power generation factors; is the time derivative; represents the cumulative effect from time 0 to t; V hyd (τ) represents the amount of water in the reservoir at time point τ, which is used for integral calculation and error analysis; dτ is a small increment of the historical time variable τ. By integrating τ, the reservoir water volume V from time 0 to the current time t is calculated. hyd (τ) and target water volume V target The cumulative error between t and dt is the small increment of the time variable t.

5. The method for optimizing flood control and scheduling based on multi-level integrated water conservancy project according to claim 3 is characterized in that: The node status update formula is: in, For node v i The state at time step T+1; σ is the activation function; b is the bias term; W is the weight matrix, which adjusts the fusion method of node state and message aggregation results; Δh ij is the dynamic change difference between upstream and downstream nodes; β ij is the adjustment coefficient of upstream and downstream influence between nodes; α ij For node v i and neighbor node v j Adaptive message passing weights between For node v i The state at time step T; For node v i The set of neighbor nodes of For node v j The state at time step T.

6. The method for optimizing flood control and scheduling based on the integration of multi-level water conservancy projects according to claim 1 is characterized in that: Step 3 includes the following steps: A top-level Transformer architecture is built to encode the historical operation data, current status, and forecast information of the entire basin's reservoir group into a sequence input. The multi-head self-attention mechanism captures long-term spatiotemporal dependencies and outputs a global operation strategy. Design a graph attention network in the middle layer, and learn the dynamic influence relationship between nodes by setting a multi-layer graph attention mechanism to optimize the coordinated scheduling strategy of the reservoir group in the region; Build the underlying deep Q network, design the mapping relationship between state space and action space for each reservoir, and output specific control instructions; Design a hierarchical training mechanism and use hierarchical reinforcement learning methods to build a hierarchical learning framework; An information transmission channel is designed to achieve two-way interaction between layers. The decisions of the upper layer guide the behavior of the lower layer by layer decomposition, and the execution feedback of the lower layer helps the upper layer optimize the decision by layer-by-layer aggregation, forming a complete hierarchical learning closed loop.

7. The method for optimizing flood control and scheduling based on the integration of multi-level water conservancy projects according to claim 1 is characterized in that: Step 4 includes the following steps: According to the hydraulic structure characteristic curve and hydraulic relationship, the scheduling strategy is converted into specific gate opening instructions; at the same time, based on the current water conditions and weather forecast information, the execution sequence of each control action is reasonably arranged; Conduct compliance checks on generated gate control instructions, including gate opening limits, opening and closing rate constraints, and flow gradient verification, to screen out potential violations and make corrections. Input gate control instructions into the dynamic model of the water conservancy project group for rolling simulation, predict the response after the instruction is executed, evaluate the dynamic impact of the control instruction, and make preventive adjustments to possible risk points; Design alternative control schemes and emergency response procedures for possible extreme situations and equipment failures, so that rapid switching can be carried out when an abnormality is detected; Implement a step-by-step execution strategy, break down large-scale regulatory instructions into multiple small steps and implement them gradually. Evaluate the effects after each step is executed, and only execute the next step after confirming safety, thereby reducing the risks brought about by large-scale regulation.

8. The method for optimizing flood control and scheduling based on the integration of multi-level water conservancy projects according to claim 1 is characterized in that: Step 5 includes the following steps: A hierarchical scoring mechanism is established based on the risk of exceeding water levels in important control sections of the basin, the flow rate of key river sections, the flood peak reduction rate and the matching degree of flood storage and discharge time sequence; A performance calculation method based on fuzzy comprehensive evaluation is established for each region's water supply guarantee rate, power generation efficiency, and ecological base flow guarantee targets. This method dynamically evaluates the overall effectiveness of regional scheduling by tracking the completion of various scheduling indicators in real time, and adaptively adjusts the weight of each indicator based on the current water situation and task importance. Evaluate the execution cost of control instructions from the perspectives of equipment life, energy consumption level, and scheduling stability; In the deep reinforcement learning framework, a credit allocation algorithm is used to reasonably distribute the rewards of long-term control effects to each decision moment, so as to correctly evaluate the contribution of each control action to the final goal. By analyzing the control effects under different working conditions, the weight ratios of the three levels of global flood control, local scheduling and control rationality are automatically adjusted to achieve a dynamic balance in reward evaluation.

9. The method for optimizing flood control and scheduling based on multi-level integrated water conservancy project according to claim 8, characterized in that: The overall efficiency of regional scheduling is expressed as: C(r) = w s (r)·μ s (x s (r))+w G (r)·μ G (x G (r))+w e (r)·μ E (x E (r)), where C(r) is the dispatch efficiency evaluation at time r, which represents the comprehensive performance of the entire regional dispatch and reflects the completion of water supply, power generation and ecological goals; w s (r) is the weight of the water supply target at time r; w G (r) is the weight of the power generation efficiency target at time r; w E (r) is the weight of the ecological base flow protection target at time r; x s (r) represents the real-time evaluation data of the water supply guarantee rate at time r; μ s (x s (r)) is the membership function of the water supply target, which indicates the completion of the water supply guarantee rate target; x G (r) represents the real-time evaluation data of power generation efficiency at time r; μ G (x G (r)) is the membership function of the power generation benefit target, which indicates the completion of the power generation target; x E (r) represents the real-time assessment data of ecological base flow protection at time r; μ E (x E (r)) is the membership function of the ecological base flow protection target, which indicates the completion of the ecological target.

10. The method for optimizing flood control and scheduling based on multi-level integrated water conservancy project according to claim 1, characterized in that: Step 6 includes the following steps: Extract weight matrices at multiple time scales from the Transformer layer and graph attention layer, and convert them into dynamic heatmap displays; The complete reasoning chain from environmental status, engineering constraints to control strategy is formalized as a directed acyclic graph, with node size representing factor importance, line thickness representing influence intensity, and arrow direction indicating causal relationship, enabling intuitive and interpretable display of the decision-making process. Use a variety of visual methods to demonstrate the degree of achievement of flood control safety, water supply guarantee rate and power generation efficiency indicators; Automatically generate structured reports including water situation analysis, plan description, expected results and risk warnings based on key data and decision-making basis of the scheduling process; Establish interactive exploration capabilities to support decision makers in conducting in-depth analysis of specific scheduling nodes.

Images