A water conservancy project digital operation management method and system

Abstract

The application discloses a kind of digital operation management method and system for hydraulic engineering, it is related to water conservancy management technical field, to the water intake of hydraulic engineering carries out three-dimensional flow field calculation, obtains three-dimensional vortex core equivalent surface, and constructs topological feature vector, vortex core morphological phase diagram is generated by state projection;Wall surface fluctuating pressure and local cross-section flow velocity are collected in real time at water intake, and reservoir state data is obtained, and online monitoring data is fused and generated;Online monitoring data is converted into dynamic space-time diagram data, dynamic space-time diagram data is input into space-time diagram operator manifold mapping network, and real-time dynamic coordinate is output;The coordinate moving track of real-time dynamic coordinate is recorded in vortex core morphological phase diagram, and future moving track is obtained by using time series prediction algorithm prediction, future moving track is judged, and touch remaining time and risk early warning signal are generated;Risk early warning signal is received, and model predictive control algorithm is used to obtain anti-vortex intervention strategy and execute regulating unit action.

Description

Technical Field

[0001] This invention relates to the field of water conservancy management technology, and in particular to a digital operation management method and system for water conservancy projects. Background Technology

[0002] When the water inlet of a hydropower station or large pumping station is operating under unfavorable conditions such as low water level and high flow rate, vertical shaft suction vortices are easily formed on the free surface. Once these vortices are connected, they will continuously draw air and floating objects into the pressure pipeline, causing the unit to vibrate violently, blade cavitation, and efficiency to drop sharply. In severe cases, it may even cause structural damage to the unit and limit the output of the entire plant, which is a major hidden danger to the safe and stable operation of hydraulic machinery.

[0003] Currently, existing vortex monitoring mainly relies on visual observation by operators above the intake, video monitoring with cameras, or experience-based judgment based on indirect signals such as unit vibration and noise. Visual and video methods are completely ineffective in low light, when water surface fluctuations are severe, or when floating debris obstructs the intake. Vibration and noise signals have significant hysteresis, often only becoming apparent after the intake vortex has fully developed and caused damage. While high-frequency signals such as wall pulsating pressure can detect flow anomalies earlier, there is currently a lack of effective methods to establish a physical correlation between these signals and the reservoir's operational boundary conditions and the vortex's three-dimensional structure. The generation, development, stretching, and collapse of a vortex is a highly nonlinear three-dimensional dynamic process. Traditional methods use single-point velocity or pressure threshold alarms, which can only determine the presence of a vortex but cannot quantitatively characterize its geometry, spatial structure, rotational intensity, and evolutionary stages. They also cannot answer the crucial question of how far the vortex is from danger. Common anti-vortex measures in engineering include installing fixed vortex-suppressing beams or baffles at the intake, or urgently reducing unit output and closing gates upon detecting an intake vortex. The former increases permanent head loss, sacrificing power generation efficiency; the latter is a passive emergency response, with a delayed response and impact on the power grid. Current technology lacks a refined closed-loop control method that can drive away vortices through minimal active intervention before they become a disaster. Although there have been attempts to apply deep learning to flow field identification in recent years, pure black-box models are heavily dependent on training data. Under extreme conditions that have never occurred before, the prediction results may seriously violate fundamental physical laws such as mass and energy conservation, making it difficult to pass engineering safety reviews and be deployed in actual machine control loops. Summary of the Invention

[0004] The technical problem solved by this invention is that existing technologies are difficult to quantitatively characterize and visualize the three-dimensional vortex state at the inlet, fail to integrate easily obtainable boundary measurement data such as high-frequency pulsating pressure on the wall and local cross-sectional flow velocity with the reservoir operation status, and cannot quantitatively predict the remaining time for the vortex to touch the danger zone, and automatically generate graded response warning signals according to the degree of urgency.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] Firstly, a digital operation and management method for water conservancy projects includes the following steps:

[0007] Step S1: Perform three-dimensional flow field calculation on the water intake of the water conservancy project to obtain the three-dimensional vortex core isosurface, construct the topological feature vector, and generate the vortex core morphology phase diagram through state projection;

[0008] Step S2: Real-time collection of wall pulsating pressure and local cross-sectional flow velocity at the inlet, acquisition of reservoir status data, and fusion to generate online monitoring data;

[0009] Step S3: Convert the online monitoring data into dynamic spatiotemporal graph data, input the dynamic spatiotemporal graph data into the spatiotemporal graph operator manifold mapping network, and output real-time dynamic coordinates;

[0010] Step S4: Record the coordinate movement trajectory of the real-time dynamic coordinates in the vortex core morphology phase diagram, use the time series prediction algorithm to predict the future movement trajectory, judge the future movement trajectory, and generate the remaining contact time and risk warning signal.

[0011] Step S5: Receive risk warning signals, use model predictive control algorithms to obtain anti-vortex intervention strategies, and execute regulating unit actions.

[0012] Preferably, step S1 includes the following sub-steps:

[0013] Step S11: Establish a three-dimensional geometric mesh model of the water intake of the water conservancy project, perform unsteady three-dimensional flow field calculations on the water intake, and obtain global flow field data. The global flow field data includes the three-dimensional velocity field and pressure field time series of each grid point in space.

[0014] Step S12: Based on the three-dimensional velocity field, calculate the velocity gradient tensor of each grid point in the flow field domain, and decompose the velocity gradient tensor to obtain the symmetric strain rate tensor and the antisymmetric rotation rate tensor. Use the Q criterion to calculate the second matrix invariant Q value of each grid point in the flow field. Use the isosurface extraction algorithm to extract the spatially connected region where the central pressure of the flow field is lower than the local environmental pressure, and obtain the three-dimensional vortex core isosurface.

[0015] Step S13: Spatial geometry and fluid dynamics features are quantized on the three-dimensional vortex core isosurface to calculate feature parameters, including vortex core volume, principal axis aspect ratio, average curvature and relative helicity. The feature parameters are then dimensionless and normalized, and vector fusion is performed in time order to construct a topological feature vector.

[0016] Step S14: A nonlinear manifold dimensionality reduction algorithm is used to reduce the dimensionality of the topological feature vectors and project them into a three-dimensional feature space through state projection. A density clustering algorithm is used to divide the scatter set in the feature space into boundaries. Based on the hydrodynamic hazard level corresponding to the scatter clustering area, state regions are divided in the feature space. The state regions include flow safety zones, warning zones, and danger zones, and a vortex core morphology phase diagram is generated.

[0017] Preferably, step S2 includes the following sub-steps:

[0018] Step S21: Deploy a high-frequency sensor array in the critical flow change zone of the water conservancy project intake to collect the wall pulsating pressure and local cross-sectional flow velocity in the critical flow change zone in real time. The critical flow change zone includes the tailrace zone of the trash rack pier, the shear layer of the water diversion pier, and the sudden expansion part of the gate slot of the intake.

[0019] Step S22: Obtain reservoir status data, which includes the current reservoir water level, the opening degree of the intake gate, and the real-time power generation load of the generating unit;

[0020] Step S23: Preprocess and spatiotemporally correlate the wall pulsating pressure, local cross-sectional flow velocity, and reservoir status data. Based on a unified timestamp, resample and synchronize the wall pulsating pressure, local cross-sectional flow velocity, and reservoir status data, and remove abnormal noise signals. Then, stitch and fuse the resampled, synchronized, and noise-removed wall pulsating pressure, local cross-sectional flow velocity, and reservoir status data to generate online monitoring data.

[0021] Preferably, step S3 includes the following sub-steps:

[0022] Step S31: Using the physical location of each sensor in the high-frequency sensing array as graph nodes, the wall pulsating pressure and local cross-sectional flow velocity collected at the corresponding locations are used as node attribute features. The directed edge weights between graph nodes are dynamically assigned based on the fluid topological connectivity and spatial flow velocity gradient, thereby constructing dynamic spatiotemporal graph data.

[0023] Step S32: Input the dynamic spatiotemporal graph data into the spatiotemporal graph operator manifold mapping network, use the spatiotemporal graph convolution encoder to perform graph convolution operation on the dynamic spatiotemporal graph data, aggregate the hydrodynamic spatial correlation features between adjacent graph nodes, and extract the high-frequency pulsation time series features of the flow field.

[0024] Step S33: The spatiotemporal graph operator manifold mapping network transmits the extracted spatial correlation features and temporal features to the neural operator integration layer. The neural operator integration layer uses the Navier-Stokes evolution operator to approximate the local flow field in the continuous time domain to capture nonlinear dynamic features.

[0025] Step S34: By comparing the manifold projection layer, the high-dimensional feature vector of the nonlinear dynamic characteristics is nonlinearly reduced in dimension, and the real-time dynamic coordinates of the current flow state at the inlet are mapped and output.

[0026] Preferably, the spatiotemporal graph operator manifold mapping network employs a joint loss function during the offline training phase, and the processing logic of the joint loss function is as follows:

[0027] Calculate the data-driven mean squared error loss term:

[0028] Extract the real-time dynamic coordinates output by the spatiotemporal graph operator manifold mapping network, and obtain the real phase graph coordinates corresponding to the labels in the offline state. Calculate the mean square error between the real-time dynamic coordinates and the real phase graph coordinates as the first loss term.

[0029] Calculate the physical residual terms based on the mass continuity equation:

[0030] The sum of the partial derivatives of the local velocity field approximated by the neural operator integral layer in the three spatial dimensions is calculated using automatic differentiation technology to obtain the velocity field divergence. Based on the mass conservation principle of incompressible fluids, the absolute value of the velocity field divergence deviating from zero is used as the local mass continuity residual, which constitutes the second loss term.

[0031] Calculate the physical residuals based on Bernoulli's energy conservation equation:

[0032] Randomly sample flow field nodes along the streamline direction, calculate the total head difference between any two nodes on the same streamline, the total head includes position head, pressure head and velocity head, calculate the absolute value of the deviation between the total head difference and the theoretical friction head loss, and use it as the energy conservation residual to constitute the third loss term;

[0033] The first, second, and third loss terms are multiplied by preset dynamic penalty weight coefficients and then weighted and superimposed to obtain the final joint loss function. The joint loss function is minimized by the backpropagation algorithm, and the parameters of the spatiotemporal graph operator manifold mapping network are updated.

[0034] Preferably, step S4 includes the following sub-steps:

[0035] Step S41: In the vortex core morphology phase diagram, the real-time dynamic coordinates are continuously recorded according to the time sequence, and the real-time dynamic coordinates are sequentially connected to generate the coordinate movement trajectory of the inlet flow state in the phase diagram feature space.

[0036] Step S42: Use a time series prediction algorithm to model the coordinate movement trajectory in time series and extrapolate the evolution trend, predict the development position of the flow state coordinate point in the vortex core morphology phase diagram within the future time window, and obtain the future movement trajectory.

[0037] Step S43, determine the future movement trajectory:

[0038] Calculate the movement rate of the future trajectory pointing towards the danger zone in the feature space. When the movement rate exceeds the set safety threshold, it is determined to be a high-risk formation stage.

[0039] Step S44: When it is determined to be a high-risk formation stage, calculate the remaining time for the flow coordinates to touch the boundary of the danger zone along the current direction based on the spatial distance and movement speed of the future movement trajectory, and divide the urgency level according to the numerical range of the remaining time, and generate a risk warning signal.

[0040] Preferably, the processing logic for the risk warning signal is as follows:

[0041] A first time threshold and a second time threshold are preset, wherein the first time threshold... The second time threshold, based on the comparison between the calculated remaining touch time and the time threshold, classifies the risk warning signal into three emergency levels and executes corresponding digital control actions:

[0042] When the remaining time is touched When the second time threshold is reached, it is determined that the flow state has a tendency to evolve in a deteriorating direction. The future movement trajectory is flashed on the digital operation management screen. The sampling frequency of the high-frequency sensor array and the calculation frequency of the time-series prediction algorithm are increased simultaneously. Without triggering the unit control action of the entity, a first-level risk warning signal is generated.

[0043] When the first time threshold is less than the remaining time before the touch event When the second time threshold is reached, a secondary risk warning signal is generated, indicating that the vortex evolution has entered the warning critical period. The system activates the model predictive control algorithm in the background in advance to perform rolling optimization and solve the optimal anti-vortex intervention strategy. After calculating the optimal anti-vortex intervention strategy, the adjustment target is pushed to the operator's terminal, and control commands are issued after manual confirmation.

[0044] When the remaining time is touched When the first threshold is reached, a level three risk warning signal is generated, indicating that a large-scale through-draft vortex is about to be fully formed. This triggers an emergency overriding mechanism, bypassing the manual confirmation process and automatically executing the regulating unit's actions.

[0045] Preferably, step S5 includes the following sub-steps:

[0046] Step S51: Receive risk warning signal, activate digital adaptive anti-vortex control system, and input the current online monitoring data and real-time dynamic coordinates on the vortex core morphology phase diagram as the initial system state into the model predictive control algorithm;

[0047] Step S52: Within the set future prediction time domain, the model predictive control algorithm constructs a control cost function constrained by hydrodynamic evolution, with the optimization objective of breaking the eddy evolution chain and minimizing the overall head loss cost to the unit. The optimal anti-eddy intervention strategy is obtained through rolling optimization solution.

[0048] Step S53: Based on the anti-vortex intervention strategy, control commands are sent to the water conservancy project digital execution system to execute the regulating unit actions. The regulating unit actions include fine-tuning the active load of the local or adjacent units and adjusting the asymmetric opening of the intake gate.

[0049] Step S54: By executing the regulating unit action, the flow field boundary conditions at the inlet are actively changed, the secondary circulation effect is stimulated, and the corresponding coordinates of the inlet flow state on the vortex core morphology phase diagram are forced to degenerate to the flow state safety zone.

[0050] Preferably, step S52 specifically includes:

[0051] Using the spatiotemporal graph operator manifold mapping network as the state prediction model, a multi-objective constrained control cost function J is constructed within the set future prediction time domain. The control cost function J consists of weighted penalty terms, which include flow state safety penalty terms, power generation economic loss penalty terms, and equipment action smoothing penalty terms.

[0052] The processing logic for flow state safety penalty items is as follows:

[0053] Calculate the Euclidean distance between the predicted coordinate point of the inlet flow regime in the vortex core morphology phase diagram and the center coordinate point of the flow regime safety zone within the future prediction time domain. The farther the predicted coordinate point is from the center coordinate point of the flow regime safety zone, the larger the value of the corresponding flow regime safety penalty term. When the predicted coordinate point approaches or enters the danger zone, the flow regime safety penalty term is dynamically adjusted to force the predicted coordinate point to retreat back to the flow regime safety zone.

[0054] The logic for handling the penalty for economic losses in power generation is as follows:

[0055] Calculate the difference between the active load adjustment of the unit and the target value of the original grid dispatch command after the implementation of the anti-vortex intervention strategy, as well as the converted value of the local head loss at the inlet caused by the adjustment of the asymmetrical opening of the inlet gate;

[0056] The processing logic for the equipment motion smoothing penalty is as follows:

[0057] Calculate the sum of squares of the rate of change of the intake gate opening and the rate of change of the unit guide vane opening between the current control time and the previous control time;

[0058] The control cost function is obtained by linearly superimposing the flow safety penalty term, the power generation economic loss penalty term, and the equipment action smoothing penalty term after multiplying them by their corresponding adaptive weight coefficients.

[0059] A sequential quadratic programming optimization solver is used to perform rolling time-domain optimization on the control cost function to solve for the future control sequence.

[0060] Extract the control increment of the first control cycle in the future control sequence and output it as the current optimal anti-vortex intervention strategy.

[0061] Secondly, a digital operation and management system for water conservancy projects includes a vortex core phase diagram module, an online monitoring module, a flow state coordinate mapping module, a phase space trajectory prediction module, and a control intervention module;

[0062] The vortex core phase diagram module is used to perform three-dimensional flow field calculations on the water intake of a water conservancy project, obtain three-dimensional vortex core isosurfaces, construct topological feature vectors, and generate vortex core morphology phase diagrams through state projection.

[0063] The online monitoring module is used to collect wall pulsating pressure and local cross-sectional flow velocity at the inlet in real time, and to acquire reservoir status data, which are then fused to generate online monitoring data.

[0064] The fluid coordinate mapping module is used to convert online monitoring data into dynamic spatiotemporal graph data, input the dynamic spatiotemporal graph data into the spatiotemporal graph operator manifold mapping network, and output real-time dynamic coordinates;

[0065] The phase space trajectory prediction module is used to record the coordinate movement trajectory of real-time dynamic coordinates in the vortex core morphology phase diagram, predict the future movement trajectory using a time-series prediction algorithm, judge the future movement trajectory, and generate the remaining contact time and risk warning signal.

[0066] The control intervention module is used to receive risk warning signals, use model predictive control algorithms to obtain anti-vortex intervention strategies, and execute the actions of the regulating unit.

[0067] The beneficial effects of this invention are as follows: By constructing a vortex core morphology phase diagram, the complex three-dimensional unsteady vortex dynamics process is compressed into a low-dimensional feature space and presented intuitively in the form of safe zones, warning zones, and danger zones. This frees operation and management personnel from dependence on experience and visual observation, enabling the first quantitative monitoring of vortex health status. The invention explicitly encodes the flow physics correlation between sensors using a dynamic spatiotemporal diagram, and through a manifold mapping network with embedded neural operators, it achieves real-time inversion from high-frequency sensor data at the wall to underwater three-dimensional vortex phase diagram coordinates, solving the problem of the difficulty in directly perceiving underwater flow states. By introducing mass continuity residuals and energy conservation residuals as joint loss functions in offline network training, the spatiotemporal prediction model's output does not violate basic physical laws even under unseen extreme conditions, greatly enhancing the model's reliability and deployability. The invention uses a time-series prediction model to extrapolate vortex evolution trends in phase space and quantitatively calculates the remaining time to reach the danger zone, enabling the system to predict when a vortex will cause a disaster, providing a clear quantitative window for control intervention. By employing model predictive control methods, with flow safety as a constraint and minimizing power generation head loss and operational costs as optimization objectives, vortices can be effectively driven away through limited interventions such as fine-tuning load and asymmetrical gate opening. This avoids the permanent head loss caused by traditional fixed facilities and the impact of passive emergency operations on the power grid, significantly improving overall power generation efficiency while ensuring safety. Attached Figure Description

[0068] Figure 1 A flowchart illustrating the steps of a digital operation and management method for water conservancy projects, as provided in one embodiment of the present invention;

[0069] Figure 2 This is a basic flowchart of a digital operation and management system for water conservancy projects provided in one embodiment of the present invention. Detailed Implementation

[0070] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0071] Example 1, referring to Figure 1 This paper provides a digital operation and management method for water conservancy projects, which includes the following steps:

[0072] Step S1: Perform three-dimensional flow field calculation on the water intake of the water conservancy project to obtain the three-dimensional vortex core isosurface, construct the topological feature vector, and generate the vortex core morphology phase diagram through state projection;

[0073] Step S2: Real-time collection of wall pulsating pressure and local cross-sectional flow velocity at the inlet, acquisition of reservoir status data, and fusion to generate online monitoring data;

[0074] Step S3: Convert the online monitoring data into dynamic spatiotemporal graph data, input the dynamic spatiotemporal graph data into the spatiotemporal graph operator manifold mapping network, and output real-time dynamic coordinates;

[0075] Step S4: Record the coordinate movement trajectory of the real-time dynamic coordinates in the vortex core morphology phase diagram, use the time series prediction algorithm to predict the future movement trajectory, judge the future movement trajectory, and generate the remaining contact time and risk warning signal.

[0076] Step S5: Receive risk warning signals, use model predictive control algorithms to obtain anti-vortex intervention strategies, and execute regulating unit actions.

[0077] The method in this embodiment focuses on the management and control of hydraulic vortex disasters at the intake of a hydropower station. First, a phase diagram of the vortex core morphology is generated through offline CFD calculation. During online operation, high- and low-frequency multimodal sensing data are fused, and graph neural networks and neural operators are used to reduce the dimensionality of the data and map it to real-time coordinate points on the phase diagram. Then, trajectory extrapolation and prediction of the remaining time to touch the danger zone are performed in the phase space. Finally, model predictive control is used to solve for the optimal strategy and actively change the unit's operating boundary to destroy the vortex.

[0078] This method advances the anti-vortex management of water conservancy projects by tens of seconds to several minutes, captures the trend of vortex deterioration, and responds by cutting off the vortex evolution chain, thus solving the problem of cavitation and severe vibration of the unit caused by vortices at the high head and high flow rate inlet.

[0079] Step S1 includes the following sub-steps:

[0080] Step S11: Establish a three-dimensional geometric mesh model of the water intake of the water conservancy project, perform unsteady three-dimensional flow field calculations on the water intake, and obtain global flow field data, which includes the three-dimensional velocity field and pressure field time series of each grid point in space.

[0081] Step S12: Based on the three-dimensional velocity field, calculate the velocity gradient tensor of each grid point in the flow field domain, and decompose the velocity gradient tensor to obtain the symmetric strain rate tensor and the antisymmetric rotation rate tensor. Use the Q criterion to calculate the second matrix invariant Q value of each grid point in the flow field. Use the isosurface extraction algorithm to extract the spatially connected region where the central pressure of the flow field is lower than the local environmental pressure, and obtain the three-dimensional vortex core isosurface.

[0082] Step S13: Spatial geometry and fluid dynamics features are quantized on the three-dimensional vortex core isosurface to calculate the feature parameters, including vortex core volume, principal axis aspect ratio, average curvature and relative helicity. The feature parameters are then dimensionless and normalized, and vector fusion is performed in time order to construct a topological feature vector.

[0083] Step S14: A nonlinear manifold dimensionality reduction algorithm is used to reduce the dimensionality of the topological feature vectors and project them into a three-dimensional feature space through state projection. A density clustering algorithm is used to divide the scatter set in the feature space into boundaries. Based on the hydrodynamic hazard level corresponding to the scatter clustering area, state regions are divided in the feature space. The state regions include the flow safety zone, warning zone, and danger zone, and a vortex core morphology phase diagram is generated.

[0084] In this embodiment, a fine mesh is constructed at the inlet for large eddy simulation. Instead of using the low-pressure method, which is susceptible to shear layer interference, the Q-criterion is employed when extracting the eddy core, defined as follows:

[0085] ;

[0086] in, The antisymmetric gyrability tensor represents the rotational intensity of the fluid. The strain rate tensor is symmetric, representing the deformation or shear strength of the fluid. The extracted isosurfaces are used to peel away the rotationally dominant core. The volume, principal axis aspect ratio, curvature, and relative helicity of the vortex core are calculated and concatenated into a topological feature vector. Finally, a nonlinear manifold dimensionality reduction algorithm (such as UMAP) is used to reduce it to three-dimensional space, and a density clustering algorithm is used to automatically delineate the boundaries.

[0087] Two thousand three-dimensional flow field snapshots under multiple operating conditions of a certain water conservancy project were extracted, and eight-dimensional topological features were extracted. The dimensionality-reduced three-dimensional scatter points naturally clustered: cluster A (small depressions) is the safe flow zone, cluster B (elongated wall-attached vortex bands) is the warning zone, and cluster C (large curvature through vortex) is the danger zone. The aim is to transform the expert experience-driven judgment of flow condition into mathematical vectors and boundaries. In this embodiment, the infinitely diverse three-dimensional flow field morphology is compressed into an intuitive phase diagram dashboard, which greatly reduces the threshold for on-site personnel to understand vortex states. The prior library generated based on CFD covers extreme flood conditions, solving the data bottleneck of the extreme scarcity of high-risk samples in actual engineering.

[0088] Step S2 includes the following sub-steps:

[0089] Step S21: Deploy a high-frequency sensor array in the critical flow change zone of the water conservancy project intake to collect the wall pulsating pressure and local cross-sectional flow velocity in the critical flow change zone in real time. The critical flow change zone includes the tailrace zone of the trash rack pier, the shear layer of the water diversion pier, and the sudden expansion part of the gate slot.

[0090] Step S22: Obtain reservoir status data, which includes the current reservoir water level, the opening degree of the intake gate, and the real-time power generation load of the generating unit.

[0091] Step S23: Preprocess and spatiotemporally correlate the wall pulsating pressure, local cross-sectional flow velocity, and reservoir status data. Based on a unified timestamp, resample and synchronize the wall pulsating pressure, local cross-sectional flow velocity, and reservoir status data, and remove abnormal noise signals. Then, stitch and fuse the resampled, synchronized, and noise-removed wall pulsating pressure, local cross-sectional flow velocity, and reservoir status data to generate online monitoring data.

[0092] In this embodiment, a high-frequency sensor array is deployed in the wake zone of the inlet trash rack, the shear layer of the diversion pier, and the abrupt expansion of the gate slot. It includes 18 combined sensor nodes. Each node is composed of a miniature pulsating pressure sensor and a three-dimensional acoustic Doppler velocity probe (sampling rate 100Hz), for a total of N=18 sensor nodes. The frequency response of the miniature pulsating pressure sensor is >1kHz, the range is 0~200kPa, and the sampling rate of the three-dimensional acoustic Doppler velocity probe is 100Hz.

[0093] The reservoir status data is obtained from the power plant's SCADA system, including the current reservoir water level with an accuracy of 0.01m, the opening degree of the intake gate with an accuracy of 0.1°, and the real-time power generation load of the generating units with an accuracy of 0.1MW.

[0094] Using a unified GPS / NTP timestamp as a reference, all pulsating pressure and flow velocity data were resampled to 100Hz, and reservoir status data were synchronized to the same time grid through linear interpolation. An adaptive Kalman filter was used to remove impulsive noise from the pressure signal and bubble interference from the flow velocity signal. Finally, all channels were stitched together to form an online monitoring data vector. ,in =4 corresponds to the 1D pulsating pressure and the 3D spatial velocity components, which serve as the input for the subsequent spatiotemporal graph network;

[0095] This step precisely aligns local high-frequency pulsations with global low-frequency states in time and space, constructing a complete input tensor that can capture early instability features of vortices and reflect evolutionary boundary conditions. The number and deployment location of sensor nodes are determined by the CFD analysis in step S1, ensuring the physical targeting of information acquisition. The data time synchronization accuracy after fusion using this method is better than 1ms, effectively eliminating abnormal noise and providing a clean, dimension-fixed data source for subsequent dynamic spatiotemporal graph networks.

[0096] Step S3 includes the following sub-steps:

[0097] Step S31: Using the physical location of each sensor in the high-frequency sensing array as graph nodes, the wall pulsating pressure and local cross-sectional flow velocity collected at the corresponding locations are used as node attribute features. The directed edge weights between graph nodes are dynamically assigned based on the fluid topological connectivity and spatial flow velocity gradient, thereby constructing dynamic spatiotemporal graph data.

[0098] Step S32: Input the dynamic spatiotemporal graph data into the spatiotemporal graph operator manifold mapping network, use the spatiotemporal graph convolution encoder to perform graph convolution operation on the dynamic spatiotemporal graph data, aggregate the hydrodynamic spatial correlation features between adjacent graph nodes, and extract the high-frequency pulsation time series features of the flow field.

[0099] Step S33: The spatiotemporal graph operator manifold mapping network transmits the extracted spatial correlation features and temporal features to the neural operator integration layer. The neural operator integration layer uses the Navier-Stokes evolution operator to approximate the local flow field in the continuous time domain to capture nonlinear dynamic features.

[0100] Step S34: By comparing the manifold projection layer, the high-dimensional feature vector of the nonlinear dynamic characteristics is nonlinearly reduced in dimension, and the real-time dynamic coordinates of the current flow state at the inlet are mapped and output.

[0101] This embodiment uses a spatiotemporal graph operator manifold mapping network to map online monitoring data into dynamic coordinates in a vortex core morphology phase diagram in real time; the dynamic spatiotemporal graph data construction is used to define 18 sensors as a set of graph nodes. Node attribute matrix ,in =4. Adjacency matrix edge weight The following formula is used for dynamic calculation:

[0102] ;

[0103] in The square of the geometric distance between the two nodes. For the length dimension parameter, in this embodiment, it is taken as 1.2 times the average spacing between sensors. The indicator function is preset based on the CFD streamline connectivity; connectivity in the downstream direction is 1. The velocity gradient modulus between nodes along the streamline direction is represented by a weight matrix that is dynamically updated according to the flow field characteristics at each time step, forming dynamic spatiotemporal graph data.

[0104] Feature extraction is performed using a spatiotemporal graph convolutional network. The update formula for the spatiotemporal graph convolution at layer l is:

[0105] ;

[0106] in, This represents the hidden state feature matrix of a node at layer l. The updated hidden state feature matrix of the node at layer l+1. Given an adjacency matrix with self-loops, For degree matrix, For learnable weight matrix, The ReLU function is used. After each graph convolution, a one-dimensional dilated convolution is applied in the time dimension with a dilation coefficient of 2 to capture the long-range temporal dependence of the flow field fluctuations. The network input is graph data with a time window of T=50 time points, i.e., 100Hz data for 0.5 seconds. The input tensor dimension is [Batch, 18, 50, 4]. The number of hidden layer channels is uniformly set to 64, and Batch represents the batch size.

[0107] The spatiotemporal latent variables output by the encoder The input Fourier neural operator layer (FNO) approximates the local NS equation evolution operator in the continuous-time domain. Its mapping formula is:

[0108] ;

[0109] in and For the Fast Fourier Transform and its inverse transform, R is the value before frequency domain truncation and preservation. A learnable complex weight tensor for low-frequency modes, where W is a linear transformation matrix, enables the network to learn the frequency domain features of flow field state transitions with fewer parameters.

[0110] The high-dimensional features output by the neural operator are mapped to a three-dimensional vector through a fully connected projection layer. The projection layer uses contrastive learning pre-training to align its output coordinates directly with the vortex core morphology phase diagram space generated in step S1. During online inference, the network can output a real-time dynamic coordinate for each new time window of data received, with a processing time of less than 10 milliseconds.

[0111] This embodiment explicitly encodes the physical topology between sensors using a graph structure, embeds the frequency domain evolution law within neural operators, and uses manifold projection to align with the physical phase diagram, significantly reducing the model's dependence on massive fault samples.

[0112] The spatiotemporal graph operator manifold mapping network employs a joint loss function during the offline training phase. The processing logic of the joint loss function is as follows:

[0113] Calculate the data-driven mean squared error loss term:

[0114] Extract the real-time dynamic coordinates output by the spatiotemporal graph operator manifold mapping network, and obtain the real phase graph coordinates corresponding to the labels in the offline state. Calculate the mean square error between the real-time dynamic coordinates and the real phase graph coordinates as the first loss term.

[0115] Calculate the physical residual terms based on the mass continuity equation:

[0116] The sum of partial derivatives of the local velocity field obtained by the neural operator integral layer approximation in three spatial dimensions is calculated using automatic differentiation technology to obtain the velocity field divergence. Based on the mass conservation principle of incompressible fluids, the absolute value of the velocity field divergence deviating from zero is used as the local mass continuity residual, which constitutes the second loss term.

[0117] Calculate the physical residuals based on Bernoulli's energy conservation equation:

[0118] Randomly sample flow field nodes along the streamline direction, calculate the total head difference between any two nodes on the same streamline, the total head includes position head, pressure head and velocity head, calculate the absolute value of the deviation between the total head difference and the theoretical friction head loss, and use it as the energy conservation residual to constitute the third loss term;

[0119] The first, second, and third loss terms are multiplied by preset dynamic penalty weight coefficients and then weighted and superimposed to obtain the final joint loss function. The joint loss function is minimized by the backpropagation algorithm, and the parameters of the spatiotemporal graph operator manifold mapping network are updated.

[0120] This embodiment provides the specific composition of the joint loss function and training parameters used in the offline training phase to ensure that the model both fits the data and follows the laws of fluid physics.

[0121] The joint loss function formula defines the total loss. It consists of three weighted factors:

[0122] ;

[0123] The first item is data-driven mean squared error loss. :

[0124] ;

[0125] in Let B be the coordinates of the i-th true phase diagram obtained through CFD simulation and dimensionality reduction projection under the same operating conditions, and let B be the batch size. The coordinates of the i-th predicted phase diagram;

[0126] The second item is the physical residual of quality continuity. :

[0127] ;

[0128] in, This represents the total number of nodes within the grid. This is the index number of the current network node. Let x be the velocity of the fluid in the x-axis direction. Let be the velocity of the fluid in the y-axis direction. The velocity of the fluid in the z-axis direction;

[0129] The divergence of the local velocity field recovered from the intermediate layer of the network is calculated using PyTorch's automatic differentiation function. The average of the squared divergence at all grid points is then calculated to constrain incompressibility.

[0130] For the spatiotemporal graph operator manifold mapping network in the offline training phase, a physical decoding branch is added after the neural operator integration layer. This branch maps the continuous latent variables of the neural operator integration layer to the regression-normalized three-dimensional physical velocity components of each grid node through a multi-layer fully connected network. , , This physical decoding branch only participates in loss calculation during offline training and is not activated during online inference;

[0131] The third item is the physical residual of energy conservation. :

[0132] ;

[0133] Randomly sample points (A, B) along the predicted streamline and calculate the total head difference and theoretical friction head loss. The absolute value of the deviation, f is the friction factor, taken as 0.02~0.03, and D is the hydraulic diameter of the flow channel. The streamline length between two points The average flow velocity, It is the acceleration due to gravity. This represents the total number of sampling streamlines. and These represent the geometric vertical heights of points A and B in three-dimensional space, respectively, in meters (m), reflecting the gravitational potential energy of the fluid. and These represent the local hydrostatic pressures at points A and B, respectively, in Pa. and Let represent the squares of the combined velocity at points A and B, respectively, and be the sum of the squares of the three-dimensional velocity components. For fluid density;

[0134] The training consists of 800 rounds in total, with the first 200 rounds using only... Training, i.e. =1.0, = =0, ensuring the network converges quickly to the data distribution, starting from round 201. The change linearly increases from 0 to 0.1. The value increases linearly from 0 to 0.05, gradually introducing physical constraints.

[0135] The AdamW optimizer was used, with a weight decay factor. The initial learning rate was 1×10⁻³, which was reduced to 1×10⁻³ using a cosine annealing strategy. Batch size 64;

[0136] By adopting an incremental training strategy of data first and physics later, the problem of unstable training caused by excessive physics loss in the early stage is avoided. The introduction of mass and energy conservation residuals enables the network to output predictions that conform to basic physical laws even in sparsely covered areas of the training set, which greatly enhances the extrapolation reliability of the model under rare conditions.

[0137] Step S4 includes the following sub-steps:

[0138] Step S41: In the vortex core morphology phase diagram, the real-time dynamic coordinates are continuously recorded according to the time series, and the real-time dynamic coordinates are connected in sequence to generate the coordinate movement trajectory of the inlet flow state in the phase diagram feature space.

[0139] Step S42: Use a time series prediction algorithm to model the coordinate movement trajectory in time series and extrapolate the evolution trend, predict the development position of the flow state coordinate point in the vortex core morphology phase diagram within the future time window, and obtain the future movement trajectory.

[0140] Step S43, determine the future movement trajectory:

[0141] Calculate the movement rate of the future trajectory pointing towards the danger zone in the feature space. When the movement rate exceeds the set safety threshold, it is determined to be a high-risk formation stage.

[0142] Step S44: When it is determined to be a high-risk formation stage, calculate the remaining time for the flow coordinates to touch the boundary of the danger zone along the current direction based on the spatial distance and movement speed of the future movement trajectory, and divide the urgency level according to the numerical range of the remaining time, and generate a risk warning signal.

[0143] This embodiment quantifies the risk of vortex disasters by recording and predicting the flow state coordinate trajectory in the phase diagram.

[0144] In the vortex core morphology phase diagram, the real-time dynamic coordinates of the output are continuously recorded at 0.5-second intervals. And connect them with smooth curves to form a trajectory;

[0145] Trajectory prediction based on Transformer employs a Transformer architecture with a multi-head self-attention mechanism for trajectory extrapolation. Input sequence length... (Historical 10 seconds, sampling interval 0.5 seconds), output prediction length (The next 5 seconds). Self-attention is calculated as follows:

[0146] ;

[0147] in, It is obtained by linear transformation of the input coordinate sequence. For querying the matrix, The key matrix, For value matrices, This is the transpose of the key matrix. The scaling factor is used to configure the model as hidden layer dimensions. The system has 8 attention heads, 4 layers each for the encoder and decoder, 300 training epochs, and an initial learning rate of 5×10⁻⁶. The loss function is MSE;

[0148] Considering the nonlinear acceleration characteristics of actual flow evolution, and in order to more safely capture the vortex deterioration trend, this embodiment will target the movement rate towards the danger zone. Defined as the maximum instantaneous velocity of movement in the normal direction towards the boundary of the danger zone within the predicted time domain, the calculation formula is modified as follows:

[0149] ;

[0150] in, Let k be the normal vector of the hyperplane at the boundary of the danger zone, where k is the index of the future prediction steps. The sequence length of the predicted trajectory is given by Δt = 0.5, where Δt = 0.5 is the time interval for recording the trajectory. and These are the predicted phase diagram coordinate vectors for the k-th and (k-1)-th future steps;

[0151] Signed distance from current coordinates to the boundary of the danger zone for:

[0152] ;

[0153] in, For the bias term of the boundary hyperplane, Let be the normal weight vector of the hyperplane, representing the tilt direction of the danger zone boundary in three-dimensional space. This is the real-time dynamic coordinate vector at the current time t.

[0154] If d > 0, it means the flow is currently still in the safe zone, but is approaching the boundary. A value of 0 indicates that the line has been crossed or the danger zone has been completely entered. The remaining touch time (TTC) is then calculated as follows:

[0155] ;

[0156] when When the value is >0, it is determined that the flow is approaching the danger zone, the TTC is calculated, and the early warning judgment process is initiated.

[0157] like If the value is ≤0, the flow is considered to be stable and no warning is triggered.

[0158] To prevent division by zero, a corresponding risk warning signal is generated based on the TTC value;

[0159] Trend prediction in phase space transforms complex flow evolution into a simple geometric distance and velocity problem, avoiding the uncertainty of directly predicting high-dimensional flow fields. The quantitative indicators of TTC provide an objective basis for subsequent graded responses.

[0160] The processing logic for risk warning signals is as follows:

[0161] A first time threshold and a second time threshold are preset, wherein the first time threshold... The second time threshold, based on the comparison between the calculated remaining touch time and the time threshold, classifies the risk warning signal into three emergency levels and executes corresponding digital control actions:

[0162] When the remaining time is touched When the second time threshold is reached, it is determined that the flow state has a tendency to evolve in a deteriorating direction. The future movement trajectory is flashed on the digital operation management screen. The sampling frequency of the high-frequency sensor array and the calculation frequency of the time-series prediction algorithm are increased simultaneously. Without triggering the unit control action of the entity, a first-level risk warning signal is generated.

[0163] When the first time threshold is less than the remaining time before the touch event When the second time threshold is reached, a secondary risk warning signal is generated, indicating that the vortex evolution has entered the warning critical period. The system activates the model predictive control algorithm in the background in advance to perform rolling optimization and solve the optimal anti-vortex intervention strategy. After calculating the optimal anti-vortex intervention strategy, the adjustment target is pushed to the operator's terminal, and control commands are issued after manual confirmation.

[0164] When the remaining time is touched When the first threshold is reached, a level three risk warning signal is generated, indicating that a large-scale through-draft vortex is about to be fully formed. This triggers an emergency overriding mechanism, bypassing the manual confirmation process and automatically executing the regulating unit's actions.

[0165] This embodiment details the three-level classification of risk warning signals and the corresponding digital control actions, based on the comparison between the TTC calculated in step S4 and the preset time threshold.

[0166] Preset first time threshold Seconds, second time threshold Second.

[0167] Level 1 Warning (TTC) The system determines that the flow condition is deteriorating but has not yet entered an emergency state. On the digital operation management screen, the future movement trajectory is displayed as a bright dashed line flashing. The sampling rate of the high-frequency sensor array is automatically increased from 100Hz to 200Hz, and the trajectory prediction calculation frequency of S42 is increased from 2Hz to 5Hz. No physical unit control actions are triggered, only prompting operators to pay attention.

[0168] Level II warning (T1) <TTC≤ (The vortex evolution has entered a critical warning period.) The system backend immediately activates the model predictive control algorithm in step S5 to perform rolling optimization and solution; after calculating the optimal anti-vortex intervention strategy, the specific adjustment plan is pushed to the operator's terminal in the form of a pop-up window, and the execution command is issued after manual confirmation.

[0169] Level 3 Warning (TTC≤T1): A large-scale through-draft intake vortex is about to form, indicating an extremely critical situation. The system triggers an emergency overriding mechanism, bypassing manual confirmation and directly invoking the S5 fully automatic control unit to execute the regulating actions, while simultaneously issuing audible and visual alarms.

[0170] By retaining human decision-making power in the secondary early warning phase with sufficient buffer time, and entrusting millisecond-level emergency response to the system autonomously, a reasonable allocation of human and machine responsibilities is achieved. This tiered response significantly reduces unnecessary unit disturbances while ensuring a safety baseline in extreme situations.

[0171] Step S5 includes the following sub-steps:

[0172] Step S51: Receive risk warning signal, activate digital adaptive anti-vortex control system, and input the current online monitoring data and real-time dynamic coordinates on the vortex core morphology phase diagram as the initial system state into the model predictive control algorithm;

[0173] Step S52: Within the set future prediction time domain, the model predictive control algorithm constructs a control cost function constrained by hydrodynamic evolution, with the optimization objective of breaking the eddy evolution chain and minimizing the overall head loss cost to the unit. The optimal anti-eddy intervention strategy is obtained through rolling optimization solution.

[0174] Step S53: Based on the anti-vortex intervention strategy, control commands are sent to the water conservancy project digital execution system to execute the regulating unit actions. The regulating unit actions include fine-tuning the active load of the local or adjacent units and adjusting the asymmetric opening of the intake gate.

[0175] Step S54: By executing the regulating unit action, the flow field boundary conditions at the inlet are actively changed, the secondary circulation effect is stimulated, and the corresponding coordinates of the inlet flow state on the vortex core morphology phase diagram are forced to degenerate to the flow state safety zone.

[0176] This embodiment demonstrates how to use model predictive control (MPC) to actively intervene and eliminate intake vortices.

[0177] Upon receiving a risk warning signal, the digital adaptive anti-vortex control system is activated. The current online monitoring data from S2 and the real-time dynamic coordinate z(t) output from S3 are used as the initial state and input into the MPC controller.

[0178] Within the set prediction time domain, MPC aims to interrupt the eddy evolution chain and minimize the power generation head loss by continuously solving for the optimal anti-eddy intervention strategy.

[0179] The adjustment increments within the first control cycle solved by MPC are sent to the actuators, specifically including: fine-tuning the active load of the local unit and adjacent units (achieved through the speed governor, with the power adjustment rate limited to ±1% of rated power / second) and adjusting the asymmetrical opening of the inlet gate (through the hydraulic hoist, with a single action increment not exceeding 0.5° / s).

[0180] The aforementioned adjustments actively altered the flow field boundary conditions at the inlet, deliberately introducing a lateral velocity gradient to stimulate secondary circulation, thereby disrupting the vertical axis structure of the intake vortex and forcing the phase diagram coordinates to degenerate to the safe flow region. The system continuously monitors the process after execution, and if the coordinates return to the warning zone, MPC is triggered again.

[0181] By taking proactive measures with small-scale, multi-unit coordination, the risk of vortex hazards can be completely eliminated with minimal power generation loss, thus avoiding the huge head loss caused by limiting gate opening for a long time in the past to prevent vortexes.

[0182] Step S52 specifically includes:

[0183] Using the spatiotemporal graph operator manifold mapping network as the state prediction model, a multi-objective constrained control cost function J is constructed within the set future prediction time domain. The control cost function J consists of weighted penalty terms, including flow state safety penalty terms, power generation economic loss penalty terms, and equipment action smoothing penalty terms.

[0184] The processing logic for flow state safety penalty items is as follows:

[0185] Calculate the Euclidean distance between the predicted coordinate point of the inlet flow regime in the vortex core morphology phase diagram and the center coordinate point of the flow regime safety zone within the future prediction time domain. The farther the predicted coordinate point is from the center coordinate point of the flow regime safety zone, the larger the value of the corresponding flow regime safety penalty term. When the predicted coordinate point approaches or enters the danger zone, the flow regime safety penalty term is dynamically adjusted to force the predicted coordinate point to retreat back to the flow regime safety zone.

[0186] The logic for handling the penalty for economic losses in power generation is as follows:

[0187] Calculate the difference between the active load adjustment of the unit and the target value of the original grid dispatch command after the implementation of the anti-vortex intervention strategy, as well as the converted value of the local head loss at the inlet caused by the adjustment of the asymmetrical opening of the inlet gate;

[0188] The processing logic for the equipment motion smoothing penalty is as follows:

[0189] Calculate the sum of squares of the rate of change of the intake gate opening and the rate of change of the unit guide vane opening between the current control time and the previous control time;

[0190] The control cost function is obtained by linearly superimposing the flow safety penalty term, the power generation economic loss penalty term, and the equipment action smoothing penalty term after multiplying them by their corresponding adaptive weight coefficients.

[0191] A sequential quadratic programming optimization solver is used to perform rolling time-domain optimization on the control cost function to solve for the future control sequence.

[0192] Extract the control increment of the first control cycle in the future control sequence and output it as the current optimal anti-vortex intervention strategy.

[0193] This embodiment describes the MPC control cost function, constraints, solution algorithm, and engineering parameters used by S52 to ensure the physical feasibility and economic optimality of anti-vortex control.

[0194] The control cost function formula MPC in the prediction time domain and control time domain Internally solve the following nonlinear programming problem:

[0195] ;

[0196] The future k-th step phase diagram coordinates predicted by the network. The coordinates of the center of the safe flow region in the phase diagram are: This refers to the deviation between the active power load regulation of the generating units and the original dispatch instructions from the power grid. Let Q be the control increment rate for the asymmetric opening of the gate, R be the positive definite diagonal weighted matrix, and S be the penalized state deviation, R be the penalized control quantity itself, and S be the rate of change of the penalized control quantity. , and These are adaptive weighting coefficients.

[0197] Unit output constraints: ;

[0198] Gate opening and speed constraints are and , This is the minimum permissible limit for the active power output of the generator unit. This refers to the maximum permissible limit of the unit's active power output. This is the minimum permissible angle for the opening of the intake gate. This represents the maximum permissible angle of the intake gate opening.

[0199] State prediction constraints ;

[0200] in For a spatiotemporal graph operator manifold mapping network, a sequential quadratic programming (SQP) algorithm based on the trust region is employed. At time t, the following steps are performed:

[0201] Using the current coordinate z(t) as the initial state, the SQP solver calculates the future... The optimal control increment sequence for each control cycle is obtained, and only the first control increment is extracted and sent to the PLC for execution.

[0202] The time window advances to t+1, and the above process is repeated;

[0203] Control step size Seconds, prediction time domain Step, control time domain step;

[0204] During normal adjustment, To balance safety and economy;

[0205] When a Level III critical danger warning is triggered, the safety weight index is amplified and set. At this point, safety weights far outweigh economic weights, forcibly prioritizing safety, and constraining the maximum active power offset of a single unit to ±3%. , Rated power, ±3% This means that each power adjustment cannot exceed 3% of the rated power;

[0206] This embodiment transforms eddy current control into an optimization problem with multiple constraints, enabling the system to automatically find the optimal strategy that minimizes power generation costs while ensuring flow safety under any hazard level. The introduction of physical constraints eliminates unsafe actions such as overshoot and oscillation. Compared with traditional fixed-rule control, the MPC strategy can reduce the total power generation loss during the eddy current intervention process by 40% to 60%, while reducing the number of gate actions by more than 50%, significantly improving engineering economy and equipment lifespan.

[0207] Example 2, refer to Figure 2 This paper provides a digital operation and management system for water conservancy projects, including a vortex core phase diagram module, an online monitoring module, a flow state coordinate mapping module, a phase space trajectory prediction module, and a control intervention module.

[0208] The vortex core phase diagram module is used to perform three-dimensional flow field calculations at the intake of water conservancy projects, obtain three-dimensional vortex core isosurfaces, construct topological feature vectors, and generate vortex core morphology phase diagrams through state projection.

[0209] The online monitoring module is used to collect wall pulsating pressure and local cross-sectional flow velocity in real time at the inlet, and to acquire reservoir status data, which is then fused to generate online monitoring data.

[0210] The manifold coordinate mapping module is used to convert online monitoring data into dynamic spatiotemporal graph data, input the dynamic spatiotemporal graph data into the spatiotemporal graph operator manifold mapping network, and output real-time dynamic coordinates;

[0211] The phase space trajectory prediction module is used to record the coordinate movement trajectory of real-time dynamic coordinates in the vortex core morphology phase diagram, use a time-series prediction algorithm to predict the future movement trajectory, judge the future movement trajectory, and generate the remaining contact time and risk warning signal.

[0212] The control intervention module is used to receive risk warning signals, use model predictive control algorithms to obtain anti-vortex intervention strategies, and execute the actions of the regulating unit.

[0213] In this embodiment, a vortex core morphology phase diagram is generated through offline CFD calculation. During online operation, high- and low-frequency multimodal sensing data are fused, and graph neural networks and neural operators are used to reduce the dimensionality of the data and map it to real-time coordinate points on the phase diagram. Then, trajectory extrapolation and prediction of the remaining time to touch the danger zone are performed in the phase space. Finally, model predictive control is used to solve the optimal strategy, actively change the unit's operating boundary to destroy the vortex, and solve the problem of cavitation and severe vibration of the unit caused by the high-head, high-flow-rate inlet vortex.

[0214] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0215] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for digital operation and management of water conservancy projects, characterized in that, Includes the following steps: Step S1: Perform three-dimensional flow field calculation on the water intake of the water conservancy project to obtain the three-dimensional vortex core isosurface, construct the topological feature vector, and generate the vortex core morphology phase diagram through state projection; Step S2: Real-time collection of wall pulsating pressure and local cross-sectional flow velocity at the inlet, acquisition of reservoir status data, and fusion to generate online monitoring data; Step S3: Convert the online monitoring data into dynamic spatiotemporal graph data, input the dynamic spatiotemporal graph data into the spatiotemporal graph operator manifold mapping network, and output real-time dynamic coordinates; Step S4: Record the coordinate movement trajectory of the real-time dynamic coordinates in the vortex core morphology phase diagram, use the time series prediction algorithm to predict the future movement trajectory, judge the future movement trajectory, and generate the remaining contact time and risk warning signal. Step S5: Receive risk warning signals, use model predictive control algorithms to obtain anti-vortex intervention strategies, and execute regulating unit actions.

2. The digital operation and management method for water conservancy projects as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Step S11: Establish a three-dimensional geometric mesh model of the water intake of the water conservancy project, perform unsteady three-dimensional flow field calculations on the water intake, and obtain global flow field data. The global flow field data includes the three-dimensional velocity field and pressure field time series of each grid point in space. Step S12: Based on the three-dimensional velocity field, calculate the velocity gradient tensor of each grid point in the flow field domain, and decompose the velocity gradient tensor to obtain the symmetric strain rate tensor and the antisymmetric rotation rate tensor. Use the Q criterion to calculate the second matrix invariant Q value of each grid point in the flow field. Use the isosurface extraction algorithm to extract the spatially connected region where the central pressure of the flow field is lower than the local environmental pressure, and obtain the three-dimensional vortex core isosurface. Step S13: Spatial geometry and fluid dynamics features are quantized on the three-dimensional vortex core isosurface to calculate feature parameters, including vortex core volume, principal axis aspect ratio, average curvature and relative helicity. The feature parameters are then dimensionless and normalized, and vector fusion is performed in time order to construct a topological feature vector. Step S14: A nonlinear manifold dimensionality reduction algorithm is used to reduce the dimensionality of the topological feature vectors and project them into a three-dimensional feature space through state projection. A density clustering algorithm is used to divide the scatter set in the feature space into boundaries. Based on the hydrodynamic hazard level corresponding to the scatter clustering area, state regions are divided in the feature space. The state regions include flow safety zones, warning zones, and danger zones, and a vortex core morphology phase diagram is generated.

3. The digital operation and management method for water conservancy projects as described in claim 2, characterized in that, Step S2 includes the following sub-steps: Step S21: Deploy a high-frequency sensor array in the critical flow change zone of the water conservancy project intake to collect the wall pulsating pressure and local cross-sectional flow velocity in the critical flow change zone in real time. The critical flow change zone includes the tailrace zone of the trash rack pier, the shear layer of the water diversion pier, and the sudden expansion part of the gate slot of the intake. Step S22: Obtain reservoir status data, which includes the current reservoir water level, the opening degree of the intake gate, and the real-time power generation load of the generating unit; Step S23: Preprocess and spatiotemporally correlate the wall pulsating pressure, local cross-sectional flow velocity, and reservoir status data. Based on a unified timestamp, resample and synchronize the wall pulsating pressure, local cross-sectional flow velocity, and reservoir status data, and remove abnormal noise signals. Then, stitch and fuse the resampled, synchronized, and noise-removed wall pulsating pressure, local cross-sectional flow velocity, and reservoir status data to generate online monitoring data.

4. The digital operation and management method for water conservancy projects as described in claim 3, characterized in that, Step S3 includes the following sub-steps: Step S31: Using the physical location of each sensor in the high-frequency sensing array as graph nodes, the wall pulsating pressure and local cross-sectional flow velocity collected at the corresponding locations are used as node attribute features. The directed edge weights between graph nodes are dynamically assigned based on the fluid topological connectivity and spatial flow velocity gradient, thereby constructing dynamic spatiotemporal graph data. Step S32: Input the dynamic spatiotemporal graph data into the spatiotemporal graph operator manifold mapping network, use the spatiotemporal graph convolution encoder to perform graph convolution operation on the dynamic spatiotemporal graph data, aggregate the hydrodynamic spatial correlation features between adjacent graph nodes, and extract the high-frequency pulsation time series features of the flow field. Step S33: The spatiotemporal graph operator manifold mapping network transmits the extracted spatial correlation features and temporal features to the neural operator integration layer. The neural operator integration layer uses the Navier-Stokes evolution operator to approximate the local flow field in the continuous time domain to capture nonlinear dynamic features. Step S34: By comparing the manifold projection layer, the high-dimensional feature vector of the nonlinear dynamic characteristics is nonlinearly reduced in dimension, and the real-time dynamic coordinates of the current flow state at the inlet are mapped and output.

5. The digital operation and management method for water conservancy projects as described in claim 4, characterized in that, The spatiotemporal graph operator manifold mapping network employs a joint loss function during the offline training phase. The processing logic of the joint loss function is as follows: Calculate the data-driven mean squared error loss term: Extract the real-time dynamic coordinates output by the spatiotemporal graph operator manifold mapping network, and obtain the real phase graph coordinates corresponding to the labels in the offline state. Calculate the mean square error between the real-time dynamic coordinates and the real phase graph coordinates as the first loss term. Calculate the physical residual terms based on the mass continuity equation: The sum of the partial derivatives of the local velocity field approximated by the neural operator integral layer in the three spatial dimensions is calculated using automatic differentiation technology to obtain the velocity field divergence. Based on the mass conservation principle of incompressible fluids, the absolute value of the velocity field divergence deviating from zero is used as the local mass continuity residual, which constitutes the second loss term. Calculate the physical residuals based on Bernoulli's energy conservation equation: Randomly sample flow field nodes along the streamline direction, calculate the total head difference between any two nodes on the same streamline, the total head includes position head, pressure head and velocity head, calculate the absolute value of the deviation between the total head difference and the theoretical friction head loss, and use it as the energy conservation residual to constitute the third loss term; The first, second, and third loss terms are multiplied by preset dynamic penalty weight coefficients and then weighted and superimposed to obtain the final joint loss function. The joint loss function is minimized by the backpropagation algorithm, and the parameters of the spatiotemporal graph operator manifold mapping network are updated.

6. The digital operation and management method for water conservancy projects as described in claim 5, characterized in that, Step S4 includes the following sub-steps: Step S41: In the vortex core morphology phase diagram, the real-time dynamic coordinates are continuously recorded according to the time sequence, and the real-time dynamic coordinates are sequentially connected to generate the coordinate movement trajectory of the inlet flow state in the phase diagram feature space. Step S42: Use a time series prediction algorithm to model the coordinate movement trajectory in time series and extrapolate the evolution trend, predict the development position of the flow state coordinate point in the vortex core morphology phase diagram within the future time window, and obtain the future movement trajectory. Step S43, determine the future movement trajectory: Calculate the movement rate of the future trajectory pointing towards the danger zone in the feature space. When the movement rate exceeds the set safety threshold, it is determined to be a high-risk formation stage. Step S44: When it is determined to be a high-risk formation stage, calculate the remaining time for the flow coordinates to touch the boundary of the danger zone along the current direction based on the spatial distance and movement speed of the future movement trajectory, and divide the urgency level according to the numerical range of the remaining time, and generate a risk warning signal.

7. The digital operation and management method for water conservancy projects as described in claim 6, characterized in that, The processing logic for the risk warning signal is as follows: A first time threshold and a second time threshold are preset, wherein the first time threshold... The second time threshold, based on the comparison between the calculated remaining touch time and the time threshold, classifies the risk warning signal into three emergency levels and executes corresponding digital control actions: When the remaining time is touched When the second time threshold is reached, it is determined that the flow state has a tendency to evolve in a deteriorating direction. The future movement trajectory is flashed on the digital operation management screen. The sampling frequency of the high-frequency sensor array and the calculation frequency of the time-series prediction algorithm are increased simultaneously. Without triggering the unit control action of the entity, a first-level risk warning signal is generated. When the first time threshold Touch the remaining time When the second time threshold is reached, a secondary risk warning signal is generated, indicating that the vortex evolution has entered the warning critical period. The system activates the model predictive control algorithm in the background in advance to perform rolling optimization and solve the optimal anti-vortex intervention strategy. After calculating the optimal anti-vortex intervention strategy, the adjustment target is pushed to the operator's terminal, and control commands are issued after manual confirmation. When the remaining time is touched When the first threshold is reached, a level three risk warning signal is generated, indicating that a large-scale through-draft vortex is about to be fully formed. This triggers an emergency overriding mechanism, bypassing the manual confirmation process and automatically executing the regulating unit's actions.

8. The digital operation and management method for water conservancy projects as described in claim 7, characterized in that, Step S5 includes the following sub-steps: Step S51: Receive risk warning signal, activate digital adaptive anti-vortex control system, and input the current online monitoring data and real-time dynamic coordinates on the vortex core morphology phase diagram as the initial system state into the model predictive control algorithm; Step S52: Within the set future prediction time domain, the model predictive control algorithm constructs a control cost function constrained by hydrodynamic evolution, with the optimization objective of breaking the eddy evolution chain and minimizing the overall head loss cost to the unit. The optimal anti-eddy intervention strategy is obtained through rolling optimization solution. Step S53: Based on the anti-vortex intervention strategy, control commands are sent to the water conservancy project digital execution system to execute the regulating unit actions. The regulating unit actions include fine-tuning the active load of the local or adjacent units and adjusting the asymmetric opening of the intake gate. Step S54: By executing the regulating unit action, the flow field boundary conditions at the inlet are actively changed, the secondary circulation effect is stimulated, and the corresponding coordinates of the inlet flow state on the vortex core morphology phase diagram are forced to degenerate to the flow state safety zone.

9. The digital operation and management method for water conservancy projects as described in claim 8, characterized in that, Step S52 specifically includes: Using the spatiotemporal graph operator manifold mapping network as the state prediction model, a multi-objective constrained control cost function J is constructed within the set future prediction time domain. The control cost function J consists of weighted penalty terms, which include flow state safety penalty terms, power generation economic loss penalty terms, and equipment action smoothing penalty terms. The processing logic for flow state safety penalty items is as follows: Calculate the Euclidean distance between the predicted coordinate point of the inlet flow regime in the vortex core morphology phase diagram and the center coordinate point of the flow regime safety zone within the future prediction time domain. The farther the predicted coordinate point is from the center coordinate point of the flow regime safety zone, the larger the value of the corresponding flow regime safety penalty term. When the predicted coordinate point approaches or enters the danger zone, the flow regime safety penalty term is dynamically adjusted to force the predicted coordinate point to retreat back to the flow regime safety zone. The logic for handling the penalty for economic losses in power generation is as follows: Calculate the difference between the active load adjustment of the unit and the target value of the original grid dispatch command after the implementation of the anti-vortex intervention strategy, as well as the converted value of the local head loss at the inlet caused by the adjustment of the asymmetrical opening of the inlet gate; The processing logic for the equipment motion smoothing penalty is as follows: Calculate the sum of squares of the rate of change of the intake gate opening and the rate of change of the unit guide vane opening between the current control time and the previous control time; The control cost function is obtained by linearly superimposing the flow safety penalty term, the power generation economic loss penalty term, and the equipment action smoothing penalty term after multiplying them by the corresponding adaptive weight coefficients. A sequential quadratic programming optimization solver is used to perform rolling time-domain optimization on the control cost function to solve for the future control sequence. Extract the control increment of the first control cycle in the future control sequence and output it as the current optimal anti-vortex intervention strategy.

10. A digital operation and management system for water conservancy projects, applied in a digital operation and management method for water conservancy projects as described in any one of claims 1-9, characterized in that, It includes a vortex nucleus phase diagram module, an online monitoring module, a flow regime coordinate mapping module, a phase space trajectory prediction module, and a control intervention module; The vortex core phase diagram module is used to perform three-dimensional flow field calculations on the water intake of a water conservancy project, obtain three-dimensional vortex core isosurfaces, construct topological feature vectors, and generate vortex core morphology phase diagrams through state projection. The online monitoring module is used to collect wall pulsating pressure and local cross-sectional flow velocity at the inlet in real time, and to acquire reservoir status data, which are then fused to generate online monitoring data. The fluid coordinate mapping module is used to convert online monitoring data into dynamic spatiotemporal graph data, input the dynamic spatiotemporal graph data into the spatiotemporal graph operator manifold mapping network, and output real-time dynamic coordinates; The phase space trajectory prediction module is used to record the coordinate movement trajectory of real-time dynamic coordinates in the vortex core morphology phase diagram, predict the future movement trajectory using a time-series prediction algorithm, judge the future movement trajectory, and generate the remaining contact time and risk warning signal. The control intervention module is used to receive risk warning signals, use model predictive control algorithms to obtain anti-vortex intervention strategies, and execute the regulating unit actions.

Images