Regional flood control dispatching system and method based on flood model and reinforcement learning

Abstract

The invention discloses a regional flood control scheduling system and method based on a flood model and reinforcement learning, and relates to the technical field of flood control scheduling of water conservancy projects, and the system comprises a data integration module, a hydrodynamic simulation module, an intelligent decision module and an evaluation module. The data integration module realizes fusion of multi-source data, the hydrodynamic simulation module adopts HiPIMS to solve a two-dimensional shallow water equation and realizes hydrodynamic simulation through CUDA acceleration, and the intelligent decision module realizes a multi-objective optimization strategy through reinforcement learning training. And the evaluation module realizes real-time monitoring, visual output and emergency response feedback by constructing an intelligent management platform. Through the architecture of data layer-simulation layer-decision layer-application layer, the purposes of reducing the submerged area, improving the storage capacity utilization rate and reducing the peak flood rate are achieved, and the regional flood control toughness is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of flood control scheduling technology in water conservancy engineering, and in particular relates to a regional flood control scheduling system and method based on flood model and reinforcement learning. Background Technology

[0002] Regional flood control scheduling is a core technical means to ensure the safety of river basins. In particular, against the backdrop of frequent extreme rainfall events caused by climate change, traditional scheduling models are facing severe challenges.

[0003] Current regional flood control scheduling mainly relies on experience-driven static rules, which faces three major technical bottlenecks that urgently need to be overcome. At the data integration level, existing systems mostly use single relational databases to store structured data, making it difficult to integrate heterogeneous information. Monitoring data from small reservoirs is scattered across various sub-district systems, hydrological and meteorological data updates are delayed, and there is a lack of linkage between engineering parameters and real-time monitoring data, resulting in insufficient accuracy in multi-source data fusion. Traditional SQL databases have inherent limitations in handling dependencies in emergency response plans; linkage rules must be manually entered one by one, leading to low response efficiency. At the hydrodynamic simulation level, mainstream models use fixed grid technology, which results in significant errors in complex terrain areas. Traditional models have large errors in simulating receding water times, and flood process simulations consume a large amount of CPU computing resources, failing to meet real-time scheduling requirements. Although some studies have introduced GPU acceleration, its dynamic adaptability is insufficient. At the decision-making strategy level, static rules are difficult to handle multi-objective conflicts. In the scheduling of pond and dam groups, when flood control discharge and irrigation water storage demands conflict, traditional methods require manual weighing, resulting in decision delays exceeding one hour.

[0004] Through research, existing patents such as CN115272345A employ static scheduling rules triggered by preset thresholds. Their core flaw lies in their inability to handle dynamic conflicts between multiple objectives; the system can only execute a single flood control objective, leading to downstream irrigation water shortages, agricultural losses, and an inability to adapt to hydrological characteristic variations caused by climate change. CN116875432B achieves parallel computation of the hydrodynamic model using a single GPU, compressing the 36-hour flood simulation time to 18 hours. However, when dealing with complex polder areas, the single GPU's memory bandwidth bottleneck results in lower acceleration efficiency than theoretically expected, and it lacks a load balancing mechanism between nodes. CN117344894A uses a fixed 50-meter grid resolution to construct the hydrodynamic model, resulting in excessive density in urban areas and wasted computational resources, while in mountainous areas, the sparse grid leads to delayed identification of the inundation range. CN118917058A can only simulate the evolution of a single flood disaster, ignoring the chain effect of rainstorm-flood-geological disasters, and only supports hydrological monitoring data input, unable to integrate multi-hazard early warning information such as geological radar.

[0005] To address the aforementioned issues, this invention proposes a three-layer architecture of "data layer - simulation layer - decision layer". It achieves real-time fusion of multi-source data through hybrid storage of SQL and Graph, improves simulation accuracy by adopting the HiPIMS model and dynamic adaptive grid technology, and generates dynamic scheduling strategies by combining deep reinforcement learning. The aim is to build a high-precision, high-efficiency, and intelligent regional flood control scheduling system. Summary of the Invention

[0006] To achieve regional flood control scheduling to ensure watershed safety and to build a high-precision, high-efficiency, and intelligent regional flood control scheduling system, this invention provides a regional flood control scheduling system based on flood models and reinforcement learning. The system comprises a regional flood control scheduling system based on flood models and reinforcement learning, and a regional flood control scheduling method based on flood models and reinforcement learning.

[0007] The aforementioned regional flood control scheduling system based on flood models and reinforcement learning includes a data integration module, a hydrodynamic simulation module, an intelligent decision-making module, and an evaluation module.

[0008] The data integration module is used to integrate multi-source data and includes a hybrid storage architecture of SQL and Graph.

[0009] Furthermore, the SQL is used to store structured scheduling parameters.

[0010] Furthermore, the Graph is used to store contingency plan dependencies.

[0011] Furthermore, the SQL and Graph hybrid storage architecture uses Kafka to achieve real-time data synchronization at 15-minute intervals.

[0012] The hydrodynamic simulation module uses HiPIMS to solve two-dimensional shallow water equations and accelerates the process with CUDA to achieve hydrodynamic simulation.

[0013] Furthermore, the hydrodynamic simulation module adopts dynamic adaptive grid technology, with an initial resolution of 10 meters in flat areas, which is automatically refined to 1-2 meters when the flow velocity is >1.5m / s or the water depth is >2m.

[0014] Furthermore, the hydrodynamic simulation module supports parallel computing with NVIDIA A100 GPUs, enabling a 36-hour flood simulation to be compressed to 4 hours.

[0015] The intelligent decision-making module implements multi-objective optimization strategies through reinforcement learning training.

[0016] Furthermore, the intelligent decision-making module has a reinforcement learning unit reward function configured with a flooded area weight of 0.4, a reservoir capacity balance weight of 0.3, and an ecological flow weight of 0.3.

[0017] Furthermore, the intelligent decision-making module employs the PPO algorithm to achieve two-stage optimization: offline training and online fine-tuning (updating the learning rate η=0.001 every hour).

[0018] The assessment module enables real-time monitoring, visual output, and emergency response feedback by building an intelligent management platform.

[0019] The aforementioned regional flood control scheduling method based on flood models and reinforcement learning includes the following steps:

[0020] The S100 data layer achieves multi-source data fusion through a data integration module;

[0021] The S200 simulation layer uses the hydrodynamic simulation module to solve the two-dimensional shallow water equations using HiPIMS and accelerates the process with CUDA.

[0022] The S300 decision layer generates multi-objective optimization strategies through the intelligent decision-making module and reinforcement learning training.

[0023] The S400 application layer enables real-time monitoring, visual output, and emergency response feedback through the intelligent management platform of the evaluation module.

[0024] S100 includes the following steps:

[0025] The S101 data acquisition system collects 1-meter resolution LiDAR terrain data and IoT sensor network data, and stores them in a categorized manner using a hybrid SQL and Graph storage architecture.

[0026] S102 Preprocessing: Cleaning and standardizing the collected data;

[0027] S103 data fusion uses the Kalman filter algorithm to perform spatiotemporal fusion of multi-source data, achieving a fusion accuracy of 98%.

[0028] The data acquisition in S101 is performed at a sampling interval of 1 minute.

[0029] In the S102 preprocessing, data cleaning is used to remove outliers.

[0030] In the S102 preprocessing, standardization is the process of standardizing the data.

[0031] S300 includes the following steps:

[0032] The S301 state space definition includes eight key variables: current reservoir water level, inflow, outflow, flood control limit water level, rainfall, downstream water level, reservoir capacity utilization rate, and ecological flow.

[0033] S302 features dual-stage training: 5000 offline training rounds (based on nearly 10 years of historical data) + hourly online fine-tuning (learning rate defined as 0.001).

[0034] The S400 includes the following steps:

[0035] S401 provides real-time monitoring and collects water level (accuracy ±0.01m) and flow velocity (range 0-10m³ / s) data.

[0036] S402 provides visual output, generating dynamic process curves and PDF scheduling reports;

[0037] S403 Emergency Response: When the water level exceeds the warning value by 0.5m, a pre-discharge command is automatically triggered.

[0038] In summary, this application includes the following beneficial technical effects:

[0039] 1. This invention effectively reduces the flood inundation area and optimizes reservoir scheduling strategies through precise flood simulation and intelligent scheduling. It improves water resource utilization efficiency while ensuring flood control safety. Through scientific flood discharge scheduling, it effectively reduces downstream flood peak flow and alleviates flood control pressure.

[0040] 2. This invention significantly shortens simulation time and reduces computing resource input through GPU acceleration technology. Precise flood warning and scheduling can effectively reduce direct economic losses caused by floods, and optimized reservoir management enhances the comprehensive utilization value of water resources.

[0041] 3. This invention significantly enhances the resilience of regional flood control systems by constructing a complete architecture of "data layer - simulation layer - decision layer - application layer", incorporates ecological flow into the optimization objective, takes ecological protection into account in flood control scheduling, and automatically triggers pre-discharge instructions when the water level exceeds the warning value by 0.5m, greatly shortening the emergency response time;

[0042] 4. This invention solves the three major technical bottlenecks in traditional flood control scheduling: difficulty in data fusion, low simulation efficiency, and rigid decision-making. It provides an intelligent solution for flood control scheduling in water conservancy projects, promotes technological upgrading in the industry, and can effectively cope with extreme rainfall events caused by climate change, thereby improving the adaptability of the flood control system. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of a regional flood control scheduling system and method based on flood model and reinforcement learning provided by the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] First Embodiment

[0046] This embodiment is applied to a typical mountainous watershed, which includes two large reservoirs, eight medium-sized reservoirs, and 36 small reservoirs, covering an area of ​​approximately 1,800 square kilometers. Due to the complex terrain, steep river slopes, and rapid flood convergence, traditional flood control methods are insufficient to cope with sudden floods.

[0047] The system deployment utilizes two NVIDIA A100 GPU servers equipped with 80GB of dedicated memory, coupled with a network of 90 IoT sensors, primarily deployed in narrow sections of the river and at reservoir spillways. Data acquisition employs 0.5-meter resolution LiDAR scanning to acquire high-precision topographic data, with sensor sampling intervals set to 30 seconds to accommodate rapidly changing hydrological conditions. In the hybrid storage architecture, an SQL database stores structured scheduling parameters, while the Neo4j graph database primarily stores inter-reservoir spillway linkage rules and the correlation between river flood discharge capacity.

[0048] The hydrodynamic simulation layer employs dynamic adaptive mesh technology, with an initial resolution of 5 meters, automatically refined to 0.5-1 meters when the flow velocity exceeds 2 m / s. The model uses HiPIMS to solve the two-dimensional shallow water equations; enabling CUDA acceleration reduces the 24-hour flood simulation time from 24 hours using traditional methods to 3 hours. Validation using nearly 10 years of historical flash flood data shows that the simulation results, compared with measured data, have a water level error controlled within ±0.03 m.

[0049] The intelligent decision-making layer's state space includes eight key variables: current reservoir water level, inflow, outflow, flood control limit water level, rainfall, downstream water level, reservoir capacity utilization, and river flood discharge capacity. The reinforcement learning unit reward function is configured with a weight of 0.5 for inundated area, 0.25 for reservoir capacity balance, and 0.25 for ecological flow, emphasizing the priority of flash flood disaster prevention. The PPO algorithm was used for 8000 rounds of offline training, with the learning rate η=0.001 updated every 30 minutes during the online fine-tuning phase to adapt to rapidly changing flash flood conditions.

[0050] Practical application results show that the system successfully responded to a sudden flash flood in the summer of 2024, reducing the inundated area by about 40%, preventing two downstream villages from being flooded, and achieving a flood peak reduction rate of 30%. The emergency response time was shortened from the traditional 45 minutes to 3 minutes, effectively protecting the lives and property of residents in mountainous areas.

[0051] Second Embodiment

[0052] This embodiment is applied to a plain polder area, which includes 5 medium-sized reservoirs and 120 small ponds and dams, with a drainage area of ​​approximately 3,200 square kilometers. Due to the flat terrain and poor drainage, prolonged waterlogging is likely to occur, and traditional flood control methods are difficult to balance the needs of drainage and water storage.

[0053] The system deployment utilizes four NVIDIA A100 GPU servers, coupled with a network of 150 IoT sensors, primarily deployed at the drainage gates and pumping stations within the polder area. Data acquisition employs 1-meter resolution LiDAR scanning to acquire topographical data of the polder area, with sensor sampling intervals set to 1 minute. In the hybrid storage architecture, an SQL database stores structured data such as polder water levels and pumping station flow rates, while a Neo4j graph database stores polder drainage linkage rules and pumping station scheduling dependencies.

[0054] The hydrodynamic simulation layer employs dynamic adaptive mesh technology, with an initial resolution of 15 meters, automatically refining to 2-3 meters when the water depth exceeds 1.5 meters. The model uses HiPIMS to solve the two-dimensional shallow water equations; enabling CUDA acceleration reduces the 48-hour flood simulation time from 48 hours using traditional methods to 6 hours. Validation using nearly 10 years of historical flooding data shows that the simulation results, compared with measured data, have a water level error controlled within ±0.04m.

[0055] The intelligent decision-making layer's state space includes eight key variables: current water level in the polder area, pump station flow rate, gate opening, rainfall, external river water level, reservoir capacity utilization rate, ecological flow, and drainage capacity. The reinforcement learning unit reward function is configured with a weight of 0.4 for the flooded area, 0.3 for the reservoir capacity balance, and 0.3 for the ecological flow, balancing drainage and water storage needs. The PPO algorithm was used for 6000 rounds of offline training, with the learning rate η=0.001 updated every hour during the online fine-tuning phase.

[0056] Practical application results show that the system reduces the flooded area by approximately 35%, improves drainage efficiency by 20%, and increases irrigation water supply capacity by 18%. The ecological flow guarantee rate reaches 96%, improving the ecological environment of the polder area. Emergency response time has been reduced from the traditional 1.5 hours to 8 minutes, effectively alleviating flooding problems in the plain polder area.

[0057] Third Embodiment

[0058] This embodiment is applied to cross-basin joint flood control scheduling, involving three adjacent basins, including four large reservoirs, 16 medium-sized reservoirs, and 72 small reservoirs, with a total basin area of ​​approximately 5,600 square kilometers. Due to the hydraulic connections between the basins, traditional single-basin scheduling methods are difficult to achieve overall optimization.

[0059] The system deployment utilizes six NVIDIA A100 GPU servers to build a distributed computing architecture, coupled with a network of 200 IoT sensors, covering key hydrological nodes across three watersheds. Data acquisition employs 1-meter resolution LiDAR scanning to acquire cross-watershed topographic data, with sensor sampling intervals set to 1 minute. The hybrid storage architecture utilizes a distributed SQL database and a distributed Neo4j graph database to achieve cross-watershed data sharing and linkage rule management.

[0060] The hydrodynamic simulation layer employs a cross-basin dynamic adaptive grid technique, with an initial resolution of 10 meters. When the flow velocity exceeds 1.5 m / s or the water depth exceeds 2 meters, the resolution is automatically refined to 1-2 meters. The model uses HiPIMS to solve the two-dimensional shallow water equations. Enabling multi-GPU parallel computing reduces the 72-hour cross-basin flood simulation time from the traditional 72 hours to 8 hours. Validation using nearly 10 years of historical cross-basin flood data shows that the simulation results are compared with measured data, and the water level error is controlled within ±0.05 m.

[0061] The intelligent decision-making layer's state space includes nine key variables: cross-basin reservoir water level, inflow, outflow, flood control limit water level, rainfall, downstream water level, reservoir capacity utilization, ecological flow, and inter-basin hydraulic connections. The reinforcement learning unit reward function is configured with a weight of 0.35 for inundation area, 0.3 for reservoir capacity balance, 0.25 for ecological flow, and 0.1 for cross-basin scheduling coordination, emphasizing the collaborative optimization of cross-basin joint scheduling. The PPO algorithm was used for 10,000 rounds of offline training, with the learning rate η=0.001 updated every hour during the online fine-tuning phase.

[0062] Practical application results show that the system reduces the cross-basin inundation area by approximately 28%, achieves a flood peak reduction rate of 22%, and increases the cross-basin water resource utilization rate by 22%. The ecological flow guarantee rate reaches 94%, improving the cross-basin ecological environment. Emergency response time is reduced from the traditional 2 hours to 10 minutes, achieving overall optimization of cross-basin flood control scheduling.

Claims

1. A regional flood control scheduling system based on flood models and reinforcement learning, characterized in that... include: The system includes a data integration module, a hydrodynamic simulation module, an intelligent decision-making module, and an evaluation module. The data integration module is used to fuse data from multiple sources; The hydrodynamic simulation module uses HiPIMS to solve two-dimensional shallow water equations and accelerates the process with CUDA to achieve hydrodynamic simulation. The intelligent decision-making module implements multi-objective optimization strategies through reinforcement learning training. The assessment module enables real-time monitoring, visual output, and emergency response feedback by building an intelligent management platform.

2. A regional flood control scheduling system based on flood model and reinforcement learning according to claim 1, characterized in that... The data integration module includes a hybrid storage architecture of SQL and Graph, with SQL storing structured scheduling parameters and Graph storing contingency plan dependencies, and real-time data synchronization at 15-minute intervals achieved through Kafka.

3. A regional flood control scheduling system based on flood model and reinforcement learning according to claim 1, characterized in that... The hydrodynamic simulation module uses dynamic adaptive grid technology. The initial resolution in flat areas is 10 meters, and it is automatically refined to 1-2 meters when the flow velocity is greater than 1.5 m / s or the water depth is greater than 2 m. It supports parallel computing with NVIDIA A100 GPU, which can compress a 36-hour flood simulation to 4 hours.

4. A regional flood control scheduling system based on flood model and reinforcement learning according to claim 1, characterized in that... The reward function of the reinforcement learning unit of the intelligent decision-making module is configured as follows: inundation area weight 0.4, reservoir capacity balance weight 0.3, and ecological flow weight 0.

3. The PPO algorithm is used to achieve two-stage optimization of offline training and online fine-tuning (learning rate η=0.001 updated every hour).

5. A regional flood control scheduling method based on flood models and reinforcement learning, characterized in that... Includes the following steps: The S100 data layer achieves multi-source data fusion through a data integration module; The S200 simulation layer uses the hydrodynamic simulation module to solve the two-dimensional shallow water equations using HiPIMS and accelerates the process with CUDA. The S300 decision layer generates multi-objective optimization strategies through the intelligent decision-making module and reinforcement learning training. The S400 application layer enables real-time monitoring, visual output, and emergency response feedback through the intelligent management platform of the evaluation module.

6. A regional flood control scheduling method based on flood model and reinforcement learning according to claim 5, characterized in that... S100 includes the following steps: S101 Data Acquisition: Acquires 1-meter resolution LiDAR terrain data and IoT sensor network data (water level / rain gauge, 1-minute sampling interval), and classifies and stores the data using the SQL and Graph hybrid storage architecture described in claim 2. S102 Preprocessing: Cleaning (removing outliers) and standardizing (unifying to the WGS84 coordinate system) the collected data. S103 data fusion uses the Kalman filter algorithm to perform spatiotemporal fusion of multi-source data, achieving a fusion accuracy of 98%.

7. A regional flood control scheduling method based on flood model and reinforcement learning as described in claim 5, characterized in that... S300 includes the following steps: The S301 state space definition includes eight key variables: current reservoir water level, inflow, outflow, flood control limit water level, rainfall, downstream water level, reservoir capacity utilization rate, and ecological flow. S302 features dual-stage training: 5000 offline training rounds (based on nearly 10 years of historical data) + hourly online fine-tuning (learning rate defined as 0.001).

8. A regional flood control scheduling method based on flood model and reinforcement learning as described in claim 5, characterized in that... The S400 includes the following steps: S401 provides real-time monitoring and collects water level (accuracy ±0.01m) and flow velocity (range 0-10m³ / s) data. S402 provides visual output, generating dynamic process curves and PDF scheduling reports; S403 Emergency Response: When the water level exceeds the warning value by 0.5m, a pre-discharge command is automatically triggered.

Images