A method and system for constructing a digital twin of a water conservancy and hydropower project

Abstract

The application relates to the technical field of water conservancy and hydropower engineering intellectualization, in particular to a construction method and system of a water conservancy and hydropower engineering digital twin. The method comprises the following steps: constructing a three-dimensional model of a water conservancy and hydropower engineering, and dividing the water conservancy and hydropower engineering into multiple substructures; collecting space-time data in the water conservancy and hydropower engineering, wherein the space-time data comprises hydrological data and meteorological data; establishing a space-time graph structure according to the multiple substructures of the water conservancy and hydropower engineering and the space-time data; constructing a digital twin of the water conservancy and hydropower engineering based on a space-time graph attention network, a physically constrained neural radiation field and a deep reinforcement learning model; and inputting the space-time graph structure into the digital twin, and outputting scheduling instructions of the water conservancy and hydropower engineering. The digital twin established by the application realizes accurate simulation of the scheduling strategy of the water conservancy and hydropower engineering by combining the space-time graph attention network, the physically constrained neural radiation field and the deep reinforcement learning model, so that corresponding scheduling decisions can be intelligently generated.

Description

Technical Field

[0001] This invention relates to the field of intelligent technology for water conservancy and hydropower engineering, specifically to a method and system for constructing a digital twin of a water conservancy and hydropower project. Background Technology

[0002] In water conservancy and hydropower projects, the dynamic simulation of hydrological processes, such as flood evolution and sediment transport, is crucial. These processes are strongly correlated with the spatiotemporal dynamics of the engineering structure. However, traditional models often rely on static rules or single data sources, making it difficult to capture the complex spatiotemporal characteristics of multi-factor coupling. This leads to prediction lags or error accumulation, failing to meet the needs of real-time flood control scheduling. Furthermore, the core of water conservancy and hydropower projects lies in the physical laws of hydraulics and structural mechanics, but traditional digital twin models often employ purely data-driven methods, easily falling into the "black box model" trap, ignoring physical mechanisms and causing simulation results to deviate from reality. Therefore, how to deeply integrate physical laws with data-driven models has become key to improving the credibility of digital twins. In addition, the scheduling of water conservancy and hydropower projects needs to consider multiple objectives such as flood control, power generation, and ecological water use, and faces the dynamic uncertainty of sudden water events.

[0003] Traditional methods mostly rely on human experience or static rule bases, making it difficult to quickly generate optimal scheduling strategies, leading to decision-making delays and impacting overall benefits. Existing digital twin models are mostly customized for specific projects, lacking sufficient spatiotemporal feature extraction capabilities, making it difficult to capture the dynamic correlation between hydrological processes and engineering structures, and unable to be reused across watersheds and projects, resulting in high costs of repeated development. Therefore, building a portable and scalable digital twin architecture has become a core challenge in promoting the large-scale application of smart water conservancy. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method and system for constructing a digital twin of a water conservancy and hydropower project. The method involves constructing a three-dimensional model of the water conservancy and hydropower project and dividing it into multiple substructures; collecting spatiotemporal data from the project, including hydrological and meteorological data; establishing a spatiotemporal graph structure based on the multiple substructures and the spatiotemporal data; constructing a digital twin of the project based on a spatiotemporal graph attention network, a physically constrained neural radiation field, and a deep reinforcement learning model; and inputting the spatiotemporal graph structure into the digital twin to output scheduling instructions for the project. The digital twin established by this invention, by combining a spatiotemporal graph attention network, a physically constrained neural radiation field, and a deep reinforcement learning model, achieves accurate simulation of the scheduling strategy of the water conservancy and hydropower project, thereby intelligently generating corresponding scheduling decisions.

[0005] This invention adopts the following technical solution: a method for constructing a digital twin of a water conservancy and hydropower project, comprising: Construct a three-dimensional model of the water conservancy and hydropower project and divide the water conservancy and hydropower project into multiple substructures; Collect spatiotemporal data from water conservancy and hydropower projects, including hydrological and meteorological data; A spatiotemporal graph structure is established based on multiple substructures and spatiotemporal data of water conservancy and hydropower projects; A digital twin of a water conservancy and hydropower project is constructed based on a spatiotemporal graph attention network, a physically constrained neural radiation field, and a deep reinforcement learning model. The spatiotemporal graph structure is input into the digital twin, and the dispatching instructions for water conservancy and hydropower projects are output.

[0006] Furthermore, the spatiotemporal graph structure is input into the digital twin, and the dispatch instructions for water conservancy and hydropower projects are output, including: The spatiotemporal graph structure is used to extract features to obtain spatiotemporal feature data; The spatiotemporal feature data is input into a physically constrained neural radiation field to obtain the physical simulation results of water conservancy and hydropower project scheduling; A decision model is constructed based on deep reinforcement learning, and the physical simulation results are input into the decision model to obtain the scheduling instructions for water conservancy and hydropower projects.

[0007] Furthermore, a spatiotemporal graph attention network is used to extract features from the spatiotemporal graph structure, specifically as follows: The spatiotemporal graph structure is standardized, and each node in the spatiotemporal graph structure is temporally partitioned. The spatiotemporal graph attention network includes a spatial attention layer, a temporal attention layer, and a fusion layer; The spatial attention weight of each node in the spatiotemporal graph structure is calculated through a spatial attention layer; The spatial attention weights of each node at different time sequences are input into the temporal attention layer, and the features of each node are weighted and fused to obtain the weighted fused features. The spatiotemporal feature data is obtained by dimensionality reduction of the weighted fusion features through the fusion layer.

[0008] Furthermore, the spatiotemporal feature data is input into a physically constrained neural radiation field to obtain the physical simulation results of water conservancy and hydropower project scheduling, specifically: A neural radiation field structure with fused residual connections is constructed based on a multilayer perceptron; and a Navier-Stokes equation residual correction term is introduced into the residual connections. Construct a total loss function for the neural radiation field that includes data loss terms and physical loss terms; The spatiotemporal feature data is input into the physically constrained neural radiation field for training until the total loss function of the neural radiation field converges, and the physical simulation results of the water conservancy and hydropower project scheduling are output.

[0009] Furthermore, a decision-making model is constructed based on deep reinforcement learning, and the physical simulation results are input into the decision-making model to obtain the scheduling instructions for water conservancy and hydropower projects, specifically including: Configure the state space, action space, and reward function; The physical simulation results are transformed into state vectors based on the state space. The scheduling operations for water conservancy and hydropower projects are defined according to the action space. Design a multi-objective weighted reward function based on the state vector and the scheduling operation of the water conservancy and hydropower project; A deep reinforcement learning model based on the PPO algorithm is established based on the multi-objective weighted reward function. By inputting the physical simulation results into the deep reinforcement learning model, the system outputs scheduling instructions for water conservancy and hydropower projects.

[0010] Furthermore, the method for establishing a spatiotemporal graph structure based on multiple substructures and spatiotemporal data of water conservancy and hydropower projects is as follows: Each substructure of a water conservancy and hydropower project is used as a node in a spatiotemporal graph structure. Obtain the relationships between nodes and establish edges between nodes in the spatiotemporal graph structure; The spatiotemporal data of each substructure of a water conservancy and hydropower project are obtained as the corresponding node value of that substructure in the spatiotemporal graph structure.

[0011] Furthermore, the relationships between the nodes include: physical connection relationships, spatial topological relationships, and functional dependencies.

[0012] The present invention further proposes a system for constructing a digital twin of a water conservancy and hydropower project, to execute any of the above-mentioned methods for constructing a digital twin of a water conservancy and hydropower project. The system includes: a three-dimensional model substructure partitioning module, a spatiotemporal data acquisition module, a spatiotemporal graph structure establishment module, and a digital twin generation module. The three-dimensional model substructure partitioning module is used to construct a three-dimensional model of a water conservancy and hydropower project and divide the water conservancy and hydropower project into multiple substructures. The spatiotemporal data acquisition module is used to collect spatiotemporal data in water conservancy and hydropower projects, including hydrological data and meteorological data. The spatiotemporal graph structure establishment module is used to establish a spatiotemporal graph structure based on multiple substructures and spatiotemporal data of a water conservancy and hydropower project. The digital twin generation module is used to construct a digital twin of a water conservancy and hydropower project based on a spatiotemporal graph attention network, a physically constrained neural radiation field, and a deep reinforcement learning model; the spatiotemporal graph structure is input into the digital twin, and the scheduling instructions for the water conservancy and hydropower project are output.

[0013] The beneficial effects of this invention are as follows: The spatiotemporal graph attention network used in this invention, through hierarchical connection of "space → time", ultimately outputs "spatiotemporal feature data" that not only contains the physical connection characteristics of engineering substructures, but also reflects the dynamic evolution of data over time, providing accurate input for subsequent physical simulation and decision-making models. Furthermore, in the physically constrained neural radiation field, the Navier-Stokes equations are forcibly embedded through residual connections, avoiding the "physical paradox" of purely data-driven models. This not only solves the training problem of deep networks, but also serves as a "forced channel for physical laws", ensuring that the physical simulation results output by the neural radiation field not only match the measured data, but also strictly satisfy the core physical equations of water conservancy and hydropower engineering. This design upgrades the physical simulation of digital twins from "empirical prediction" to "interpretable physical deduction", providing a reliable basis for engineering scheduling decisions. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of a method for constructing a digital twin of a water conservancy and hydropower project according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a system for constructing a digital twin of a water conservancy and hydropower project, according to an embodiment of the present invention. Detailed Implementation

[0016] 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.

[0017] A schematic flowchart of a method for constructing a digital twin of a water conservancy and hydropower project according to an embodiment of the present invention is shown below. Figure 1 As shown, it includes: Construct a three-dimensional model of the water conservancy and hydropower project and divide the water conservancy and hydropower project into multiple substructures; In this embodiment of the invention, the three-dimensional model of the water conservancy and hydropower project is obtained based on its BIM model or laser point cloud data. After obtaining the three-dimensional model, the three-dimensional model is divided into multiple substructures according to the different types of engineering components in the water conservancy and hydropower project, such as dams, gates, and spillways. Each substructure has its corresponding physical feature data, such as geometric dimensions or material properties.

[0018] In a specific embodiment of the present invention, each substructure has its corresponding physical characteristic data, which can be specifically represented as follows: For the dam substructure, its physical characteristic data includes the geometric dimensions of the dam body, such as dam height and dam crest length, and material properties such as concrete grade or permeability coefficient; in this embodiment of the present invention, it can also include stress-strain relationship, that is, the relationship between stress and corresponding strain generated by the material under stress; for the gate substructure, the required physical characteristic data includes the shape of the gate, such as rectangular or trapezoidal, as well as roughness coefficient and connection method with the reservoir, etc., which can be selected and obtained according to the actual situation. It should be noted that when dividing the substructure, it is necessary to ensure the independence and modeling requirements of each substructure so as to ensure that the substructures can interact through standardized interfaces.

[0019] In another specific embodiment of the present invention, the physical feature data corresponding to each substructure includes geometric parameters, material properties, hydraulic parameters, and topological relationships. For geometric parameters, these can be length, width, height, radius of curvature, etc., which are the basic parameters for constructing a three-dimensional model. For material properties, these can be density, elastic modulus, Poisson's ratio, coefficient of thermal expansion, etc. For hydraulic parameters, these can be flow coefficient, roughness, permeability coefficient, etc. For topological relationships, these can be spatial relationships or logical dependencies between substructures. Spatial relationships refer to adjacent or inclusive relationships in spatial location, while logical dependencies refer to a certain order between the inputs or outputs of two substructures, such as the direction of water flow.

[0020] Collect spatiotemporal data from water conservancy and hydropower projects, including hydrological and meteorological data; In this embodiment of the invention, water level gauges, piezometers, stress gauges, and other equipment are deployed in the corresponding water conservancy and hydropower engineering parts of each substructure to collect real-time monitoring data such as deformation, seepage pressure, and stress strain; flow meters and ADCP (Acoustic Doppler Current Profiler) are deployed in river cross-sections and reservoir areas to obtain hydrological parameters such as water flow velocity and flow rate; spatial information such as the water area and vegetation cover of the water conservancy and hydropower engineering is obtained using optical satellites (such as Sentinel-2); remote sensing images of the water conservancy and hydropower engineering area are collected using synthetic aperture radar (SAR) satellites to monitor the deformation of the terrain and changes in meteorological data of the water conservancy and hydropower engineering site; and point cloud data of the water conservancy and hydropower engineering can be obtained by periodically inspecting the site using drones equipped with multispectral cameras and lidar (LiDAR) to monitor local details such as surface cracks and seepage points.

[0021] In one specific embodiment of the present invention, the collected spatiotemporal data is further preprocessed. Specifically, outliers (such as sudden increases in piezometer readings) are detected in the sensor data using a sliding window statistical method, and noise is corrected or removed using Kalman filtering or isolated forest algorithms. Radiometric and geometric corrections are performed on the remote sensing images to eliminate atmospheric interference and terrain distortion. At the same time, the timestamps of multi-source data are aligned, for example, by interpolating and synchronizing the satellite transit time with the sensor sampling time. The point cloud data collected by the UAV is unified to the same spatial reference system through coordinate transformation.

[0022] A spatiotemporal graph structure is established based on multiple substructures and spatiotemporal data of water conservancy and hydropower projects; In this embodiment of the invention, each substructure of the water conservancy and hydropower project is taken as a node of the spatiotemporal graph structure; the association between nodes is obtained to establish the edges between nodes in the spatiotemporal graph structure; wherein, the association between nodes includes: physical connection relationship, spatial topological relationship and functional dependency relationship; the spatiotemporal data of each substructure of the water conservancy and hydropower project is obtained as the node value corresponding to the substructure in the spatiotemporal graph structure.

[0023] In a specific embodiment of the present invention, the substructures corresponding to each engineering component after the water conservancy and hydropower project is divided are selected as nodes of the spatiotemporal graph structure, such as reservoirs and gates. At the same time, the spatiotemporal data corresponding to each node is bound. For example, the node corresponding to the reservoir is bound to water level gauge data and piezometer data, which are the corresponding hydrological data in the spatiotemporal data; the node corresponding to the gate is bound to opening sensor data and flow meter data, which are the hydrological data and monitoring data in the spatiotemporal data. The edges between nodes are described in this embodiment of the present invention according to the spatial topological relationship. For example, the reservoir node and the downstream gate node are connected by the water flow path as the edge between the nodes, which can represent the direction of water flow; the dam node and its adjacent mountain node can be connected by geological coupling as the edge between the nodes, which represents the constraint relationship of the foundation.

[0024] A digital twin of a water conservancy and hydropower project is constructed based on a spatiotemporal graph attention network, a physically constrained neural radiation field, and a deep reinforcement learning model; the spatiotemporal graph structure is input into the digital twin, and the scheduling instructions of the water conservancy and hydropower project are output.

[0025] In this embodiment of the invention, the constructed digital twin uses a spatiotemporal graph attention network to extract features from the spatiotemporal graph structure to obtain spatiotemporal feature data; the spatiotemporal feature data is input into a physically constrained neural radiation field to obtain physical simulation results of water conservancy and hydropower project scheduling; a decision model is constructed based on deep reinforcement learning, and the physical simulation results are input into the decision model to obtain water conservancy and hydropower project scheduling instructions.

[0026] In this embodiment of the invention, the Spatiotemporal Graph Attention Network (STGAT) achieves global context awareness by processing dependencies in the temporal and spatial dimensions in parallel. In the spatial dimension, it calculates the attention weights between nodes to identify key spatial associations, such as the strong association between reservoir nodes and downstream river nodes in a flood evolution scenario. In the temporal dimension, it captures temporal dependency patterns, such as the seasonal periodic characteristics of sediment transport processes. Through end-to-end training, STGAT dynamically adjusts the weights of nodes and edges. Node weights reflect their contribution to the current task, such as the increase in the weight of a reservoir node during a rainstorm event. Edge weights quantify the strength of interactions between nodes, such as the weight of a water flow path edge increasing with the flow rate. Taking the flood evolution process as an example, STGAT's spatial attention captures the impact of the rising water level of the reservoir node on the flow rate of the downstream gate node, while the temporal attention mechanism learns the impact of the inflow rate over the past 24 hours on the current gate opening decision. This capture of spatiotemporal dependencies enables STGAT to generate data representations that integrate spatiotemporal features. This data representation is a high-order feature abstraction containing spatiotemporal context information. For example, the feature representation of a reservoir node may include the current water level and the rate of change of water level over the past 3 hours. At the same time, dynamic weights are implicitly encoded in the feature vectors, reflecting the real-time change pattern of the engineering status. Finally, the output spatiotemporal feature data provides key inputs for physical constraint modeling of the Physically Constrained Neural Radiation Field (PC-NeRF) and decision optimization of the decision-making model based on deep reinforcement learning (DRL-Scheduler).

[0027] In a specific embodiment of the present invention, the method for feature extraction of the spatiotemporal graph structure using a spatiotemporal graph attention network is as follows: The spatiotemporal graph structure is standardized, and each node in the spatiotemporal graph structure is temporally partitioned; the spatiotemporal graph attention network includes a spatial attention layer, a temporal attention layer, and a fusion layer; The spatial attention weight of each node in the spatiotemporal graph structure is calculated through a spatial attention layer. The spatial attention layer adopts a graph attention network with a multi-head attention mechanism. In this embodiment of the invention, an eight-head attention mechanism can be adopted, with each head attention mechanism having a dimension of 64.

[0028] The spatial attention weights of each node at different time sequences are input into the temporal attention layer. The temporal attention layer uses a Transformer encoder to perform weighted fusion of the features of each node to obtain weighted fused features. The fusion layer performs dimensionality reduction on the weighted fusion features through a gating mechanism to obtain spatiotemporal feature data.

[0029] In this embodiment of the invention, the method for inputting spatiotemporal feature data into a physically constrained neural radiation field to obtain the physical simulation results of water conservancy and hydropower project scheduling is as follows: a neural radiation field structure with fused residual connections is constructed based on a multilayer perceptron; a Navier-Stokes equation residual correction term is introduced into the residual connection; a total loss function of the neural radiation field containing data loss term and physical loss term is constructed; the spatiotemporal feature data is input into the physically constrained neural radiation field for training until the total loss function of the neural radiation field converges, and the physical simulation results of water conservancy and hydropower project scheduling are output.

[0030] In one specific embodiment of the present invention, a neural radiation field structure with fused residual connections is constructed based on a multilayer perceptron. In this embodiment, the multilayer perceptron has eight layers. By inputting information such as three-dimensional coordinates, viewpoint direction, and physical parameters contained in the spatiotemporal feature data into the multilayer perceptron, the corresponding physical quantities are output. At the same time, a residual block is inserted between the fourth layer and the output layer of the multilayer perceptron, which is a residual connection. The residual connection adopts the residual correction term of the Navier-Stokes equation. By solving the residual correction term of the equation, the neural radiation field structure with fused residual connections constructed by the multilayer perceptron can output physical simulation results that satisfy the conservation of momentum.

[0031] In one specific embodiment of the present invention, a total loss function for the neural radiation field is constructed, which includes a data loss term and a physical loss term. The data loss term refers to minimizing the mean square error between the predicted value and the measured data, and the physical loss term is the Navier-Stokes equation residual correction term introduced through residual connection as a residual penalty. In this embodiment of the present invention, the discretized Navier-Stokes equation residual is used as the core part of the fluid dynamics residual loss. The physical loss term is constructed by combining physical equations such as Darcy's law. The network parameters of the neural radiation field are optimized through backpropagation to ensure that the physical simulation results conform to physical laws and achieve deep integration of physical equations and neural radiation field.

[0032] In this embodiment of the invention, after constructing the physically constrained neural radiation field, it is trained using historical spatiotemporal characteristic data of water conservancy and hydropower projects. The parameters of the physically constrained neural radiation field are optimized through backpropagation until the total loss function of the neural radiation field converges, resulting in a well-trained physically constrained neural radiation field. At this point, the spatiotemporal characteristic data of the physically constrained neural radiation field are input, and the physical simulation results of the water conservancy and hydropower project are output, such as the dynamic changes of the water flow velocity field within a set time period after the gate is opened; and the correlation between the gate opening degree and the downstream water level.

[0033] In this embodiment of the invention, the method for constructing a decision model based on deep reinforcement learning and inputting physical simulation results into the decision model to obtain water conservancy and hydropower project scheduling instructions specifically includes: Set up a state space, action space, and reward function; transform the physical simulation results into state vectors based on the state space; define the scheduling operations for water conservancy and hydropower projects based on the action space; design a multi-objective weighted reward function based on the state vectors and the scheduling operations for water conservancy and hydropower projects; establish a deep reinforcement learning model based on the PPO algorithm based on the multi-objective weighted reward function; and output water conservancy and hydropower project scheduling instructions by inputting the physical simulation results into the deep reinforcement learning model.

[0034] In one specific embodiment of the present invention, a Markov Decision Process (MDP) is used to define the scheduling objective of a water conservancy and hydropower project. A Markov Decision Process is a mathematical framework for making sequential decisions in an uncertain environment, consisting of a state space, an action space, and a reward function. In this embodiment of the present invention, the key indicators in the physical simulation results are transformed into state vectors containing parameters specific to water conservancy projects through the state space. In addition to the conventional indicators such as water level, flow rate, and equipment status, this embodiment of the present invention also provides indicators such as equipment aging coefficient, safety margin of key structures, ecological flow deficit, and flood evolution stage as inputs to the state space.

[0035] According to the motion space definition of the scheduling operation of water conservancy and hydropower projects, in this embodiment of the invention, the motion space is refined into continuous control quantities. Optional scheduling operations include: adjusting the gate opening with a step size of 0.1m, and selecting frequency conversion control with a frequency range of 30-50Hz for pump station start-up and shutdown.

[0036] The reward function is designed using a multi-objective weighted approach. In this embodiment of the invention, the multi-objective weighting is taken as an example using the time scale, and the reward function can be expressed as follows: ; in, For immediate reward, its value equals flood control safety plus power generation efficiency. Flood control safety is obtained through a negative exponential function of the water level relative to the warning value, and power generation efficiency is obtained through the ratio of actual power generation to theoretical maximum power generation. As a mid-term reward, its value is equal to Ecological flow deviation Equipment wear and tear and It is a weighting coefficient used to adjust the impact of ecological flow deviation and equipment wear on the medium-term reward. The medium-term reward focuses on ecological protection and the sustainable use of equipment. For long-term rewards, they are usually obtained through a comprehensive benefit assessment of the simulated scheduling results of water conservancy and hydropower projects; , , These are the weighting coefficients, and Their values ​​determine the relative importance of rewards at each time scale within the total reward, and can be adjusted according to actual needs and project priorities. For example, when ecological protection is given more emphasis, the rewards can be appropriately increased. The value of this value will be increased, and the proportion of mid-term ecological flow deviation in the total reward will be increased.

[0037] In one specific embodiment of the present invention, a deep reinforcement learning model based on the PPO algorithm is established according to a multi-objective weighted reward function. The PPO algorithm separates the decision-making and evaluation functions through a dual-network architecture, which includes a policy network and a value network. The policy network takes reservoir water level, flow rate, and unit status as inputs and outputs a continuous action distribution, such as the gate opening adjustment amount. It also uses a Gaussian distribution to parameterize the action space to adapt to the needs of fine control. The value network evaluates the long-term value of the state to assist in policy updates. Its pruning mechanism limits the magnitude of a single policy update, avoiding training oscillations caused by excessively large step sizes in traditional policy gradient algorithms. For example, when the probability ratio of the new and old policies exceeds a preset range, the PPO algorithm controls the update magnitude within a safe threshold by pruning terms to ensure training stability.

[0038] In this embodiment of the invention, a digital twin environment for water conservancy and hydropower projects is further constructed based on a physically constrained neural radiation field to train a deep reinforcement learning model based on the PPO algorithm. When generating the digital twin environment, the physically constrained neural radiation field model, combined with historical monitoring data and real-time sensor data, generates simulation scenarios. Specifically, this is achieved through the following methods: Historical data-driven scenario generation: Monte Carlo simulation sets are constructed using collected historical spatiotemporal data to generate diverse training scenarios including extreme conditions such as flood evolution and equipment failure; Real-time data injection mechanism: Real-time data such as current water level, flow rate, and equipment status are injected into the digital twin environment as initial conditions to ensure that the simulation scenario is synchronized with the physical system; Random perturbation enhancement: Random noise that conforms to physical laws is superimposed on historical data to enhance the robustness of the agent; Multi-scale simulation: Refined sub-models are constructed for key equipment (such as gates) to generate equipment-level control command samples. Training samples are collected by executing random strategies in the digital twin environment, and after labeling state transitions and reward values, they are stored in an experience replay buffer for iterative optimization by the PPO algorithm. Thus, the scheduling decisions for water conservancy and hydropower projects can be output based on the trained deep reinforcement learning model.

[0039] In another embodiment of the present invention, a system for constructing a digital twin of a water conservancy and hydropower project is also proposed, and the system structure diagram is shown below. Figure 2 As shown, the system includes: a 3D model substructure partitioning module, a spatiotemporal data acquisition module, a spatiotemporal graph structure establishment module, and a digital twin generation module; In a specific embodiment of the present invention: a three-dimensional model substructure partitioning module is used to construct a three-dimensional model of a water conservancy and hydropower project and divide the water conservancy and hydropower project into multiple substructures; a spatiotemporal data acquisition module is used to collect spatiotemporal data in the water conservancy and hydropower project, wherein the spatiotemporal data includes hydrological data and meteorological data; a spatiotemporal graph structure establishment module is used to establish a spatiotemporal graph structure based on the multiple substructures of the water conservancy and hydropower project and the spatiotemporal data; a digital twin generation module is used to construct a digital twin of the water conservancy and hydropower project based on a spatiotemporal graph attention network, a physically constrained neural radiation field, and a deep reinforcement learning model; the spatiotemporal graph structure is input into the digital twin, and the scheduling instructions of the water conservancy and hydropower project are output.

[0040] In one specific embodiment of the present invention, the spatiotemporal graph attention network, physically constrained neural radiation field and deep reinforcement learning model in the digital twin generation module are encapsulated as independent modules, and a standardized API interface is defined to realize the interaction between data flow and control flow, that is, the process from spatiotemporal features to physical simulation and then to decision output.

[0041] Specifically, the spatiotemporal graph attention network module encapsulates spatiotemporal attention calculation logic and exposes spatiotemporal feature generation interfaces. The input is raw monitoring data (water level, flow rate, etc.), and the output is spatiotemporal feature data that integrates spatiotemporal dependencies.

[0042] The physical constraint neural radiation field encapsulates the three-dimensional field simulation and physical constraint solution logic, provides a physical simulation interface, receives spatiotemporal characteristics as input, and returns physical simulation results such as water flow field and stress field that conform to physical laws.

[0043] The deep reinforcement learning model encapsulates the PPO algorithm and scheduling strategy generation logic, provides a strategy reasoning interface, takes the current engineering status and simulation results as input, and outputs decision commands such as gate opening and unit output.

[0044] This invention further constructs a 3D digital twin visualization platform based on WebGL, which has the following functions: Engineering status monitoring: The reservoir water level distribution and flow velocity vector field are displayed using a heat map, and abnormal status alarms are achieved through color mapping (such as red representing water levels exceeding the warning level).

[0045] The simulation results are displayed dynamically: the evolution of floods is simulated using a particle system, and the timeline playback and key frame comparison are supported. For example, the water level change curves of the "current strategy" and the "historical best strategy" can be compared.

[0046] Interactive decision-making instructions: Provides slider controls for parameters such as gate opening and unit output, which users can manually adjust and observe the changes in simulation results in real time, enabling human-machine collaborative decision-making.

[0047] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a digital twin of a water conservancy and hydropower project, characterized in that, include: Construct a three-dimensional model of the water conservancy and hydropower project and divide the water conservancy and hydropower project into multiple substructures; Collect spatiotemporal data from water conservancy and hydropower projects, including hydrological and meteorological data; A spatiotemporal graph structure is established based on multiple substructures and spatiotemporal data of water conservancy and hydropower projects; A digital twin of a water conservancy and hydropower project is constructed based on a spatiotemporal graph attention network, a physically constrained neural radiation field, and a deep reinforcement learning model. The spatiotemporal graph structure is input into the digital twin, and the dispatching instructions for water conservancy and hydropower projects are output. The specific steps of inputting the spatiotemporal diagram structure into the digital twin and outputting scheduling instructions for water conservancy and hydropower projects include: The spatiotemporal graph structure is used to extract features to obtain spatiotemporal feature data; The spatiotemporal feature data is input into a physically constrained neural radiation field to obtain the physical simulation results of water conservancy and hydropower project scheduling; A decision model is constructed based on deep reinforcement learning, and the physical simulation results are input into the decision model to obtain the scheduling instructions for water conservancy and hydropower projects.

2. The method for constructing a digital twin of a water conservancy and hydropower project according to claim 1, characterized in that: Feature extraction of the spatiotemporal graph structure is performed using a spatiotemporal graph attention network, specifically as follows: The spatiotemporal graph structure is standardized, and each node in the spatiotemporal graph structure is temporally partitioned. The spatiotemporal graph attention network includes a spatial attention layer, a temporal attention layer, and a fusion layer; The spatial attention weight of each node in the spatiotemporal graph structure is calculated through a spatial attention layer; The spatial attention weights of each node at different time sequences are input into the temporal attention layer, and the features of each node are weighted and fused to obtain the weighted fused features. The spatiotemporal feature data is obtained by dimensionality reduction of the weighted fusion features through the fusion layer.

3. The method for constructing a digital twin of a water conservancy and hydropower project according to claim 1, characterized in that: The spatiotemporal feature data is input into a physically constrained neural radiation field to obtain the physical simulation results of water conservancy and hydropower project scheduling, specifically: A neural radiation field structure with fused residual connections is constructed based on a multilayer perceptron; and a Navier-Stokes equation residual correction term is introduced into the residual connections. Construct a total loss function for the neural radiation field that includes data loss terms and physical loss terms; The spatiotemporal feature data is input into the physically constrained neural radiation field for training until the total loss function of the neural radiation field converges, and the physical simulation results of the water conservancy and hydropower project scheduling are output.

4. The method for constructing a digital twin of a water conservancy and hydropower project according to claim 1, characterized in that: A decision model is constructed based on deep reinforcement learning. The physical simulation results are input into the decision model to obtain the scheduling instructions for water conservancy and hydropower projects, specifically including: Configure the state space, action space, and reward function; The physical simulation results are transformed into state vectors based on the state space. The scheduling operations for water conservancy and hydropower projects are defined according to the action space. Design a multi-objective weighted reward function based on the state vector and the scheduling operation of the water conservancy and hydropower project; A deep reinforcement learning model based on the PPO algorithm is established based on the multi-objective weighted reward function. By inputting the physical simulation results into the deep reinforcement learning model, the system outputs scheduling instructions for water conservancy and hydropower projects.

5. The method for constructing a digital twin of a water conservancy and hydropower project according to claim 1, characterized in that: The method for establishing a spatiotemporal diagram structure based on multiple substructures and spatiotemporal data of a water conservancy and hydropower project is as follows: Each substructure of a water conservancy and hydropower project is used as a node in a spatiotemporal graph structure. Obtain the relationships between nodes and establish edges between nodes in the spatiotemporal graph structure; The spatiotemporal data of each substructure of a water conservancy and hydropower project are obtained as the corresponding node value of that substructure in the spatiotemporal graph structure.

6. The method for constructing a digital twin of a water conservancy and hydropower project according to claim 5, characterized in that: The relationships between the nodes include: physical connection relationships, spatial topological relationships, and functional dependencies.

7. A system for constructing a digital twin of a water conservancy and hydropower project, comprising executing a method for constructing a digital twin of a water conservancy and hydropower project as described in any one of claims 1-6, characterized in that, include: The module includes a 3D model substructure partitioning module, a spatiotemporal data acquisition module, a spatiotemporal graph structure establishment module, and a digital twin generation module. The three-dimensional model substructure partitioning module is used to construct a three-dimensional model of a water conservancy and hydropower project and divide the water conservancy and hydropower project into multiple substructures. The spatiotemporal data acquisition module is used to collect spatiotemporal data in water conservancy and hydropower projects, including hydrological data and meteorological data. The spatiotemporal graph structure establishment module is used to establish a spatiotemporal graph structure based on multiple substructures and spatiotemporal data of a water conservancy and hydropower project. The digital twin generation module is used to construct a digital twin of a water conservancy and hydropower project based on a spatiotemporal graph attention network, a physically constrained neural radiation field, and a deep reinforcement learning model; the spatiotemporal graph structure is input into the digital twin, and the scheduling instructions for the water conservancy and hydropower project are output.

Images