Reservoir density flow prediction method and device, computer equipment and storage medium

Abstract

The invention provides a reservoir density flow prediction method and device, computer equipment and a storage medium, and the method comprises the steps: obtaining the current density flow data of a target reservoir, and enabling the current density flow data to comprise velocity field data and concentration field data corresponding to a plurality of moments in a current time period; the velocity field data at each moment comprises flow velocity information at a plurality of spatial positions of the target reservoir at the moment, and the concentration field data at each moment comprises concentration information at a plurality of spatial positions of the target reservoir at the moment; generating multi-frame multi-channel tensor data based on the current density flow data; and processing the multi-frame multi-channel tensor data by using a target neural network for density flow prediction to generate predicted density flow data of the target reservoir in a future time period. Through the process, the density flow of the reservoir is simulated and predicted by using the neural network, and the prediction efficiency is improved while the prediction precision is guaranteed.

Description

Technical Field

[0001] This application relates to the field of computer technology, and more specifically, to a method, apparatus, computer equipment, and storage medium for predicting density currents in a reservoir. Background Technology

[0002] Simulation and prediction of density currents in reservoirs are among the core technologies for ensuring safe reservoir operation, optimizing water resource allocation, and extending reservoir lifespan. Driven by the density difference between sediment-laden water and clear water, density currents directly determine sediment transport and deposition distribution within the reservoir area. Accurate prediction of density currents provides crucial information for developing scientific sediment discharge scheduling plans, thereby effectively controlling reservoir deposition, ensuring flood control and maximizing reservoir capacity utilization, and possessing significant engineering practical value.

[0003] In related technologies, the simulation and prediction of density flows mainly rely on numerical simulation techniques and are carried out under two types of model frameworks: one is a refined two-fluid or multiphase Euler-Euler–Euler model, which is achieved by explicitly solving the equations of each phase and coupling the interphase forces; the other is a mixture / drift flux model suitable for engineering calculations, which takes the momentum of the mixture as the main body and describes the movement of sediment through relative drift velocity.

[0004] However, traditional computational fluid dynamics (CFD) models face significant bottlenecks when dealing with complex boundary conditions in real-world engineering projects. First, computational costs are high, with single simulations often taking hours or even days, making it difficult to meet the speed requirements of real-time scheduling. Second, real-world boundary conditions exhibit strong unsteadiness and multi-scale characteristics. For example, inflow rate (Q) can surge several times within hours during peak flood periods, reservoir water level (E) varies gradually on a diurnal scale with scheduling commands, and sediment concentration (C) shows abrupt changes and lags due to heavy rainfall, tributary inflows, and particle size variations. This complexity renders traditional numerical models, calibrated with finite parameters, ineffective at generalization and prone to error drift when predicting unseen conditions. Therefore, the simulation and prediction accuracy of density flows using the aforementioned methods is relatively low. Summary of the Invention

[0005] In view of this, this application provides a method, apparatus, computer equipment, and storage medium for predicting density currents in reservoirs.

[0006] Specifically, this application is implemented through the following technical solution: In a first aspect, embodiments of this application provide a method for predicting density currents in a reservoir, including: Acquire the current density current data of the target reservoir, wherein the current density current data includes velocity field data and concentration field data corresponding to multiple times within the current time period, the velocity field data at each time includes the flow velocity information at multiple spatial locations of the target reservoir at that time, and the concentration field data at each time includes the concentration information at multiple spatial locations of the target reservoir at that time. Based on the current heterogeneous flow data, generate multi-frame, multi-channel tensor data; The multi-frame, multi-channel tensor data is processed using a target neural network for density current prediction to generate predicted density current data for the target reservoir in the future time period.

[0007] In one optional implementation, the method further includes: Based on the predicted density current data, the target engineering indicators of the target reservoir are determined. Based on the target engineering indicators of the target reservoir, a scheduling strategy for the target reservoir in the future time period is generated.

[0008] In one optional implementation, the target neural network includes an encoding module, a decoding module, and a long short-term memory network module. The step of using the target neural network for recurrent prediction to process the multi-frame, multi-channel tensor data to generate predicted recurrent data for the target reservoir in the future time period includes: The encoding module is used to extract features from the multi-frame, multi-channel tensor data to generate first feature data; The first feature data is processed using the Long Short-Term Memory network module to generate the second feature data; The decoding module is used to decode the first feature data and the second feature data to generate predicted density current data for the target reservoir in the future time period.

[0009] In one optional implementation, the encoding module includes m first convolutional processing units, the decoding module includes m second convolutional processing units, the first feature data includes intermediate feature data output by the m first convolutional processing units, the feature sizes of the m intermediate feature data are different, and m is a positive integer greater than 1. The step of using the Long Short-Term Memory network module to perform feature processing on the first feature data to generate second feature data includes: The Long Short-Term Memory network module is used to perform feature processing on the intermediate feature data with the smallest feature size in the first feature data to generate the second feature data. The step of using the decoding module to decode the first feature data and the second feature data to generate predicted density current data for the target reservoir in the future time period includes: The first second convolution processing unit is used to process the second feature data and the first intermediate feature data to generate the first third feature data; The i-th second convolution processing unit is used to process the (i-1)-th third feature data and the i-th intermediate feature data in the first feature data to generate the i-th third feature data; where i is a positive integer greater than 1 and less than or equal to m, and the feature size of the i-th intermediate feature data is greater than the feature size of the (i-1)-th intermediate feature data. Based on the m-th third feature data, predictive density current data for the target reservoir in future time periods are generated.

[0010] In one optional implementation, the target neural network includes a modulation module, and the method further includes: Obtain the reservoir boundary condition information of the target reservoir in a future time period, including water level boundary information, flow rate boundary information, and sediment content boundary information; Based on the reservoir boundary condition information, multi-channel state data is generated; The modulation module is used to process the multi-channel state data to generate modulation parameters; After using the Long Short-Term Memory network module to perform feature processing on the intermediate feature data with the smallest feature size in the first feature data to generate the second feature data, the method further includes: The second feature data is modulated using the modulation parameters to generate processed second feature data, wherein the decoding module is used to decode the first feature data and the processed second feature data.

[0011] In one optional implementation, the multi-channel tensor data includes static channels and dynamic channels, wherein the static channels include a computational domain mask channel, a time channel, and a spatial location channel; and the dynamic channels include a flow rate channel, a concentration channel, and a boundary condition channel.

[0012] In one optional implementation, the target neural network is trained according to the following steps: Based on the obtained sample boundary condition information, multiple sample cases are generated, wherein each sample case includes water level sample information, flow rate sample information, and sediment concentration sample information; For each of the aforementioned sample cases, the drift flux model is used to generate sample recurrent data corresponding to the sample case; Based on the sample recurrent data corresponding to the multiple sample cases, multiple sample tensor data are generated; Using the multiple sample tensor data, the constructed neural network for predicting heterogeneous currents is trained in multiple rounds until the training cutoff condition is met, thereby generating the target neural network.

[0013] In one optional implementation, obtaining sample boundary condition information includes: The target's interactive interface displays the parameter setting page; In response to a setting operation triggered by the user on the parameter setting page, obtain the sample boundary condition information set by the user.

[0014] In one optional implementation, after generating the predicted density flow data, the method further includes: Based on the predicted density current data, a density current simulation image corresponding to the target reservoir is generated; The target interactive interface is controlled to display the simulated density flow.

[0015] Secondly, embodiments of this application also provide a reservoir density current prediction device, comprising: The acquisition module is used to acquire the current density current data of the target reservoir. The current density current data includes velocity field data and concentration field data corresponding to multiple moments within the current time period. The velocity field data at each moment includes the flow velocity information at multiple spatial locations of the target reservoir at that moment, and the concentration field data at each moment includes the concentration information at multiple spatial locations of the target reservoir at that moment. The generation module is used to generate multi-frame, multi-channel tensor data based on the current heterogeneous flow data; The inference module is used to process the multi-frame, multi-channel tensor data using a target neural network for density current prediction, and generate predicted density current data for the target reservoir in the future time period.

[0016] This application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting density currents in a reservoir.

[0017] This application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-mentioned method for predicting density currents in a reservoir.

[0018] This application proposes a method, apparatus, computer equipment, and storage medium for predicting density currents in a reservoir. It acquires current density current data of a target reservoir and generates multi-frame, multi-channel tensor data based on this data. This multi-frame, multi-channel tensor data is then input into a target neural network for density current prediction. The target neural network is used to predict and simulate density current data for the target reservoir in future time periods. This application utilizes deep learning technology to predict and simulate density currents in a target reservoir using current density current data, improving prediction efficiency while maintaining accuracy. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a method for predicting density currents in a reservoir, as shown in an exemplary embodiment of this application. Figure 2 This application illustrates a method for predicting density currents in a reservoir using an exemplary embodiment, showing velocity field data and concentration field data at time = 82800s. Figure 3 This is a schematic diagram of the modulation processing of the modulation module in a method for predicting density currents in a reservoir, as illustrated in an exemplary embodiment of this application. Figure 4a This is a structural diagram of the target neural network in a reservoir density current prediction method illustrated in an exemplary embodiment of this application; Figure 4b This is a schematic diagram of a CNN convolutional neural network and an LSTM long short memory network model in a reservoir density current prediction method according to an exemplary embodiment of this application. Figure 5 This is a schematic diagram of the sampling space in a reservoir density current prediction method illustrated in an exemplary embodiment of this application; Figure 6 This is a schematic diagram of the architecture of a density current prediction system in a reservoir density current prediction method according to an exemplary embodiment of this application. Figure 7 This is a schematic diagram of a reservoir density current prediction device shown in an exemplary embodiment of this application; Figure 8 This is a schematic diagram of the structure of a computer device provided in this application. Detailed Implementation

[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0021] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0022] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0023] Traditional computational fluid dynamics (CFD) models face significant bottlenecks when dealing with complex boundary conditions in real-world engineering. First, computational costs are high, with single simulations often taking hours or even days, making it difficult to meet the speed requirements of real-time scheduling. Second, real-world boundary conditions exhibit strong unsteadiness and multi-scale characteristics. For example, inflow rate (Q) can surge several times within hours during flood peaks, reservoir water level (E) varies gradually on a diurnal scale with scheduling commands, and sediment concentration (C) shows abrupt changes and lags due to heavy rainfall, tributary inflows, and particle size variations. This complexity renders traditional numerical models, calibrated with finite parameters, inadequate for generalization, and prone to error drift when predicting unseen conditions. Therefore, the accuracy of simulating and predicting density flows using traditional CFD models is relatively low.

[0024] To accurately simulate and predict density currents in reservoirs, this application proposes a method for predicting density currents. This method acquires current density current data for the target reservoir and generates multi-frame, multi-channel tensor data based on this data. This multi-frame, multi-channel tensor data is then input into a target neural network used for density current prediction. The target neural network then predicts and simulates the density current data for the target reservoir in future time periods. This application utilizes deep learning technology to predict and simulate density currents in the target reservoir using current density current data, improving prediction efficiency while maintaining accuracy.

[0025] To facilitate understanding of this embodiment, a detailed description of the reservoir density current prediction method disclosed in this disclosure is provided first. The execution entity of the reservoir density current prediction method provided in this disclosure is generally a computer device with certain computing capabilities. This computer device may include, for example, a terminal device, a server, or other processing devices. The terminal device may be a user equipment (UE), a mobile device, a user terminal, a personal digital assistant (PDA), a handheld device, a computing device, etc. In some possible implementations, the reservoir density current prediction method can be implemented by a processor calling computer-readable instructions stored in memory.

[0026] See Figure 1 The diagram shows a flowchart of a reservoir density current prediction method provided in this embodiment of the present disclosure. The method includes steps S101 to S103, wherein: S101. Obtain the current density current data of the target reservoir, wherein the current density current data includes velocity field data and concentration field data corresponding to multiple times within the current time period. The velocity field data at each time moment includes the flow velocity information at multiple spatial locations of the target reservoir at that time moment, and the concentration field data at each time moment includes the concentration information at multiple spatial locations of the target reservoir at that time moment.

[0027] The target reservoir can be any reservoir. The current density current data for the target reservoir is obtained. This data includes velocity field and concentration field data for multiple moments within the current time period. See [link to relevant documentation]. Figure 2 The data shown includes velocity and concentration field data at a specific moment (e.g., Time=82800s). The velocity field data includes flow velocity information at multiple spatial locations within the target reservoir at that moment, and the concentration field data includes concentration information at multiple spatial locations within the target reservoir at that moment. Spatial locations can be represented with water depth as the vertical axis and distance from a preset location within the target reservoir as the horizontal axis. The concentration information can be the sediment concentration at that spatial location.

[0028] S102. Based on the current heterogeneous flow data, generate multi-frame multi-channel tensor data.

[0029] Based on the current density current data, multi-frame, multi-channel tensor data is constructed. The velocity field data and concentration field data at each time step can be used to construct one frame of multi-channel tensor data. The number of channels and the data type within each channel can be set according to actual needs.

[0030] Optionally, the multi-channel tensor data includes static channels and dynamic channels, wherein the static channels include computational domain mask channels, time channels, and spatial location channels; and the dynamic channels include flow rate channels, concentration channels, and boundary condition channels.

[0031] Static channels represent data whose content does not change over time, while dynamic channels represent data whose content changes over time. Static channels can include, but are not limited to, computational domain mask channels, time-domain channels, and spatial location channels. The computational domain mask channel represents the location of wet regions in the velocity and concentration field data; that is, the mask value for a wet region is 1, and the mask value for a non-wet region is 0. This allows subsequent processing to focus only on features at the wet region location, improving the accuracy of simulation predictions. The time-domain channel represents the time information corresponding to the velocity and concentration field data. The spatial location channel represents the spatial location of each element at its position.

[0032] Dynamic channels may include, but are not limited to, velocity channels, concentration channels, and boundary condition channels. In the velocity channel, the element information at each element location can be the velocity information at the corresponding spatial location; similarly, in the concentration channel, the element information at each element location can be the concentration information at the corresponding spatial location. The boundary condition channel represents the boundary conditions of the target reservoir within the current time period. These boundary conditions include, but are not limited to, water level conditions, flow rate conditions, and sediment concentration conditions. The boundary conditions of the target reservoir within the current time period can be obtained in real time.

[0033] Setting up multiple channels here allows for richer and more comprehensive multi-channel tensor data input to the target neural network, enabling more accurate prediction and simulation of heterogeneous flows using the multi-channel tensor data.

[0034] S103. Using a target neural network for predicting density currents, the multi-frame, multi-channel tensor data is processed to generate predicted density current data for the target reservoir in the future time period.

[0035] After generating multi-frame, multi-channel tensor data, the multi-frame, multi-channel tensor data can be input into the trained target neural network for density current prediction, so that the target neural network can output the predicted density current data of the target reservoir in the future time period based on the multi-frame, multi-channel tensor data.

[0036] The network structure of the target neural network can be set according to actual needs. For example, convolutional neural networks and long short-term memory neural networks can be used to construct the target neural network.

[0037] In one optional embodiment, the target neural network includes an encoding module, a decoding module, and a long short-term memory network module. The step of using the target neural network for recurrent prediction to process the multi-frame, multi-channel tensor data to generate predicted recurrent data for the target reservoir in the future time period includes: Step a1: Use the encoding module to extract features from the multi-frame, multi-channel tensor data to generate first feature data.

[0038] Step a2: Use the Long Short-Term Memory Network module to perform feature processing on the first feature data to generate the second feature data.

[0039] Step a3: Use the decoding module to decode the first feature data and the second feature data to generate the predicted density current data of the target reservoir in the future time period.

[0040] The target neural network includes an encoding module, a decoding module, and a long short-term memory network module. It uses the encoding module to extract and compress spatial features, then uses the long short-term memory network module to extract temporal features, and finally uses the decoding module to upsample the learned high-dimensional features to reconstruct the original physical space and generate predicted density current data for the target reservoir in the future time period.

[0041] In practice, an encoding module extracts features from multi-frame, multi-channel tensor data to generate first feature data. This first feature data is then input into a Long Short-Term Memory (LSTM) network module for feature processing, outputting second feature data. The second feature data is then input into a decoding module, which fuses the first and second feature data and performs decoding to generate predicted density current data for the target reservoir over a future time period. The first feature data reflects the spatial characteristics of the multi-channel tensor data, while the second feature data reflects its temporal characteristics. Fusing and decoding the first and second feature data yields feature-rich data, enabling more accurate generation of predicted density current data for the target reservoir over a future time period.

[0042] In one optional embodiment, the encoding module includes m first convolutional processing units, the decoding module includes m second convolutional processing units, the first feature data includes intermediate feature data output by the m first convolutional processing units, the feature sizes of the m intermediate feature data are different, and m is a positive integer greater than 1.

[0043] Assuming m is 5, the first second convolutional processing unit can extract features from multi-frame, multi-channel tensor data, outputting an intermediate feature data (e.g., the 5th intermediate feature data). This 5th intermediate feature data is then input into the second second convolutional processing unit, which extracts features from it to generate the 4th intermediate feature data. The feature size of the 5th intermediate feature data is larger than that of the 4th. This process continues until the 5th second convolutional processing unit extracts features from the 2nd intermediate feature data to generate the 1st intermediate feature data. These five intermediate feature data constitute the first feature data.

[0044] The step of using the Long Short-Term Memory (LSTM) network module to perform feature processing on the first feature data to generate the second feature data includes: using the LSM network module to perform feature processing on the intermediate feature data with the smallest feature size in the first feature data to generate the second feature data.

[0045] During implementation, the Long Short-Term Memory (LSTM) network module performs feature processing on the intermediate feature data with the smallest feature size in the first feature data (such as the first intermediate feature data) to generate the second feature data, and then inputs the second feature data into the decoding module for processing.

[0046] The step of using the decoding module to decode the first feature data and the second feature data to generate predicted density current data for the target reservoir in the future time period includes: using the first second convolution processing unit to process the second feature data and the first intermediate feature data to generate the first third feature data; using the i-th second convolution processing unit to process the (i-1)-th third feature data and the i-th intermediate feature data in the first feature data to generate the i-th third feature data; where i is a positive integer greater than 1 and less than or equal to m, and the feature size of the i-th intermediate feature data is greater than the feature size of the (i-1)-th intermediate feature data; and generating predicted density current data for the target reservoir in the future time period based on the m-th third feature data.

[0047] In implementation, the decoding module includes m second convolutional processing units. The first second convolutional processing unit upsamples the second feature data to generate processed second feature data. The feature size of the processed second feature data is the same as the feature size of the first intermediate feature data. The processed second feature data is then fused with the first intermediate feature data (e.g., through concatenation, or concatenation followed by feature extraction) to generate the first third feature data. The first third feature data and the second intermediate feature data are input to the second second convolutional processing unit. The second second convolutional processing unit upsamples the first third feature data to generate processed first third feature data with the same feature size as the second intermediate feature data. This processed first third feature data is then fused with the second intermediate feature data to generate the second third feature data. This process continues until the m-th second convolutional processing unit generates the m-th third feature data. Furthermore, based on the m-th third feature data, predicted density current data for the target reservoir in the future time period can be generated. For example, an activation function can be used to process the m-th third feature data to generate predicted density current data for the target reservoir in the future time period.

[0048] The target neural network in this application is a U-net structure, which can extract rich feature data and enable the decoding module to fuse spatial and temporal features to achieve more accurate prediction of heterogeneous flow data.

[0049] Optionally, the target neural network includes a modulation module, and the method further includes: acquiring reservoir boundary condition information of the target reservoir in a future time period, the reservoir boundary condition information including water level boundary information, flow boundary information, and sediment content boundary information; generating multi-channel state data based on the reservoir boundary condition information; and processing the multi-channel state data using the modulation module to generate modulation parameters.

[0050] After using the Long Short-Term Memory Network module to perform feature processing on the intermediate feature data with the smallest feature size in the first feature data to generate the second feature data, the method further includes: using the modulation parameters to perform modulation processing on the second feature data to generate processed second feature data, wherein the decoding module is used to decode the first feature data and the processed second feature data.

[0051] In this application, a FiLM feature linear modulation mechanism can also be set in the target neural network so that the target neural network is not a static input-output mapping, but becomes a conditional prediction model that can dynamically respond and strictly follow external boundary conditions. This can avoid the errors caused by traditional interpolation methods and improve the physical consistency and generalization ability of the neural network under unseen conditions.

[0052] During implementation, reservoir boundary condition information for the target reservoir over a future time period can be obtained. This boundary condition information includes, but is not limited to, water level boundary information, flow rate boundary information, and sediment concentration boundary information. Based on this boundary condition information, derived physical field information can be determined. For example, pressure field information can be determined based on water level boundary information, and inflow velocity can be generated based on flow rate boundary information. Finally, multi-channel state data can be generated based on the reservoir boundary condition information and the derived physical field information. This multi-channel state data may include, but is not limited to, water level boundary channels, flow rate boundary channels, sediment concentration channels, pressure field channels, and inflow velocity channels. The future time period boundary condition information for the target reservoir can be set according to the actual operating conditions of the target reservoir.

[0053] A modulation module is used to process the multi-channel state data to generate modulation parameters. For example, the modulation module may include a fully connected layer and a multilayer perceptron. The fully connected layer processes the multi-channel state data and inputs the generated feature data into the multilayer perceptron for further processing to generate modulation parameters such as (γ, β).

[0054] After generating the second feature data, modulation parameters can be used to modulate the second feature data to generate processed second feature data, such as... ,in This is the second feature data. This is the processed second feature data.

[0055] The processed second feature data is input to the decoding module for further processing. That is, the decoding module is used to decode the first feature data and the processed second feature data. The decoding process can be referred to the above description and will not be repeated here.

[0056] See Figure 3 As shown, it can generate two historical frames of data input channels (14 channels), i.e., multi-channel tensor data, and one future frame of state channels (11 channels), i.e., multi-channel state data. The multi-channel tensor data is input to the encoder (i.e., the encoding module) and the ConvLSTM (i.e., the Long Short-Term Memory network module) for processing to extract abstract spatiotemporal features h (i.e., generating second feature data h). Simultaneously, the multi-channel state data can be input to the FiLM generator (i.e., the modulation module) for processing to generate modulation parameters (γ, β). The modulation module includes a global pooling layer and a multilayer perceptron (MLP). The modulation parameters (γ, β) are then used to modulate the second feature data h, generating processed second feature data. This processed second feature data is then input to the decoding module for processing to output the prediction result, i.e., the predicted heterogeneous flow data.

[0057] The following combination Figure 4a The diagram showing the structure of the target neural network illustrates the processing procedure. In implementation, after generating multi-channel tensor data, the multi-channel tensor data is input into the target neural network. The encoding module of the target neural network extracts features from the multi-channel tensor data to generate first feature data. This first feature data includes multiple intermediate feature data with different feature sizes, such as the first intermediate feature data, the second intermediate feature data, ..., the fifth intermediate feature data. The first intermediate feature data is input into the Long Short-Term Memory (LSTM) network module for processing to generate second feature data. The second feature data is then processed using modulation parameters generated by the modulation module to generate processed second feature data, which is then transmitted to the decoding module for further processing. In the decoding module, the first second convolutional processing unit processes the processed second feature data and the first intermediate feature data with the smallest feature size to generate the first third feature data; the second second convolutional processing unit processes the first third feature data and the second intermediate feature data to generate the second third feature data, and so on, to obtain the fifth third feature data. Finally, based on the fifth third feature data, the predicted density current data corresponding to the target reservoir is generated.

[0058] The first convolutional processing unit in the encoding module and the second convolutional processing unit in the decoding module can use, for example... Figure 4b The CNN convolutional neural network shown is used to extract spatial features. The structure of the long short-term memory network module can be as follows: Figure 4b The LSTM (Long Short-Term Memory) network model shown is used to extract temporal features.

[0059] In one optional embodiment, the method further includes: determining target engineering indicators for the target reservoir based on the predicted density current data; and generating a scheduling strategy for the target reservoir in the future time period based on the target engineering indicators of the target reservoir.

[0060] During implementation, after generating the predicted density current data, the target engineering indicators for the target reservoir can be determined based on this data. These indicators include, but are not limited to, key engineering parameters such as the location of the density current intrusion point, the time of arrival at the dam, and the thickness of the turbid water layer. Based on these target engineering indicators, a scheduling strategy for the target reservoir in the future time period is generated. This strategy can be set according to the actual situation and is not specifically limited here.

[0061] By predicting the density current data of the target reservoir in the future, the target engineering indicators can be determined more accurately based on the predicted density current data, thereby enabling more accurate scheduling of the target reservoir to ensure its operational status.

[0062] The process of training the target neural network is explained below.

[0063] In one alternative implementation, the target neural network is trained according to the following steps: Step b1: Based on the obtained sample boundary condition information, generate multiple sample cases, wherein each sample case includes water level sample information, flow rate sample information, and sediment concentration sample information; Step b2: For each of the sample cases, generate sample recurrent data corresponding to the sample case using the drift flux model; Step b3: Generate multiple sample tensor data based on the sample recurrent data corresponding to the multiple sample cases respectively; Step b4: Using the multiple sample tensor data, train the constructed neural network to be trained for different density current prediction in multiple rounds until the training cutoff condition is met, and generate the target neural network.

[0064] In step b1, during implementation, sample boundary condition information can be obtained, and multiple sample cases can be generated based on this information. The sample boundary condition information can be pre-determined boundary condition information, or it can be boundary condition information set by the user in real-time on the target interactive interface.

[0065] Optionally, obtaining sample boundary condition information includes: controlling the target interactive interface to display a parameter setting page; and in response to a setting operation triggered by the user on the parameter setting page, obtaining the sample boundary condition information set by the user.

[0066] During implementation, a target interactive interface can be set up, which can include a parameter setting page. Users can input boundary condition information on this page, and the system will then retrieve the user-defined sample boundary condition information in response to user actions. This information can then be used to generate multiple sample cases. This process allows for flexible, convenient, and efficient setting of sample boundary conditions.

[0067] Based on the acquired sample boundary condition information, multiple sample cases are generated. For example, Latin hypercube sampling can be used to generate random combinations of inflow rate, water level, and sediment concentration. Each random combination constitutes a sample case, ensuring that the training data covers the entire parameter space. Each sample case includes water level sample information, flow rate sample information, and sediment concentration sample information. See [link to relevant documentation]. Figure 5 As shown in the figure, each black dot represents a sample case, and multiple sample cases can be obtained within this sampling space.

[0068] In step b2, after obtaining multiple sample cases, the drift flux model can be used to generate sample recurrent data corresponding to each of the multiple sample cases.

[0069] For example, the drift flux model can solve the mixture continuity equation, mixture momentum equation, and sediment transport equation to obtain highly guaranteed sample data, namely the sample density flow data (such as sample velocity field data and sample concentration field data) corresponding to each sample case.

[0070] The continuity equation for the mixture is shown in the following formula (1): (1) in, The density of the mixture represents the total mass of the aqueous and sediment phases per unit volume. Let be the velocity of the mixture, representing the overall volume-average velocity vector of the water-sand mixture; For time.

[0071] The momentum equation for the mixture is shown in formula (2) below: (2) in, Pressure of the mixture; The molecular kinematic viscosity of the mixture; Turbulent viscosity; This represents the volume fraction of the k-th phase (such as sediment phase); Let be the density of the k-th phase; Let be the drift velocity of the k-th phase relative to the velocity of the mixture; It describes the movement of the sediment phase relative to the overall mixture; Other interphase forces, such as drag force and lift force.

[0072] The sediment transport equation is shown in formula (3) below: (3) in, This represents the volume fraction of the k-th phase (sediment phase); The absolute velocity of the kth phase ; Let be the drift velocity of the k-th phase; This is the diffusion coefficient of sediment, which includes the effects of molecular diffusion and turbulent diffusion; The spatial gradient represents the volume fraction of sediment. The terms on the right-hand side of the equation represent the diffusion effect caused by the concentration gradient.

[0073] The above process yields sample density flow data for each sample case, namely sample concentration field data and sample velocity field data. The sample concentration field data reflects the changes in concentration over time at multiple spatial locations, and the sample concentration field data reflects the changes in velocity over time at multiple spatial locations.

[0074] In step b3, during implementation, for each sample case, multiple sample tensor data are generated based on the sample heterogeneous flow data corresponding to that sample case. The sample tensor data is multi-channel data, and the multiple channels include, but are not limited to, computational domain mask channel, time channel, spatial location channel, flow velocity channel, concentration channel, boundary condition channel, etc.

[0075] In this application, all physical quantities are normalized channel by channel based on the statistical characteristics of the training set, and the statistical quantities are kept consistent during the inference stage to ensure the standardization of data processing.

[0076] This application incorporates computational domain mask channels and boundary condition channels, employing mask processing and boundary alignment mechanisms to ensure physical plausibility. In practical implementation, the computational domain mask can be dynamically generated based on real-time water levels, ensuring the accuracy of wet and dry area delineation and limiting loss calculations during training and prediction to the effective water area, thus avoiding optimization of ineffective regions.

[0077] In step b4, the neural network to be trained for heterogeneous flow prediction is trained multiple times using multiple sample tensor data until training cutoff conditions are met. Training cutoff conditions may include, for example, the loss function being less than a set loss value, model convergence, and the number of training iterations being greater than a preset number. The neural network with the best performance during the training process is selected as the target neural network.

[0078] For example, sample tensor data can be input into the neural network to be trained. The neural network processes the sample tensor data, outputs prediction results, and determines the loss function value based on the prediction results and the true results. The network parameters of the neural network to be trained are then adjusted using the loss function value, and the training of the neural network to be trained is repeated multiple times. This application can also set evaluation metrics, including mean squared error, mean absolute percentage error, etc.

[0079] For example, the loss function can be the joint mean squared error loss function. In the loss function formula, let the time series length be T, and the effective (wetland) raster set at each time step be... mask That is, the mask value for wet areas is 1, and the mask value for non-wet areas is 0; the prediction result of the deep learning model is denoted as The CFD results are used as a reference and are denoted as follows: Channel weight .

[0080] The loss function is shown in formula (4) below: (4) Where T is the total length of the predicted time series; t is the index of the current time step; p is the index of the spatial grid point; and c is the index of the physical quantity channel, specifically referring to the three core physical quantities: (Sediment volume fraction) (Momentum flux in the x-direction) (Momentum flux in the y direction). It is a binary mask, with a value of 1 in wet areas (i.e., effective areas occupied by water bodies) and 0 in non-wet areas (dry areas such as land or ineffective areas). The binary mask changes dynamically with the reservoir water level (E), ensuring that loss calculations only occur within physically reasonable fluid areas, thus avoiding optimization of ineffective areas. Channel weights are coefficients greater than 0 used to balance the contributions of different physical quantities in the loss function. Since different physical quantities (such as concentration and velocity) have different numerical magnitudes and importance, different channel weights can be set for different physical quantities. As a normalization factor, its function is to calculate the average error within the effective region at each time step t, rather than the sum of errors. This makes the value of the loss function independent of the size of the effective region at the current time step, thus ensuring the stability of the training process.

[0081] This application combines a loss function with a regularization strategy, driven by data. The base loss function employs mask-weighted mean squared error, assigning appropriate weights to different physical quantities. Based on this, boundary impermeability constraints and spatiotemporal smoothing regularization terms are introduced. A rolling consistency loss is designed for multi-step predictions to effectively suppress error accumulation in long-term predictions. Modern deep learning techniques are used to optimize model training and ensure training effectiveness. Furthermore, this application employs an optimization algorithm combining weight decay and a dynamic learning rate scheduling strategy to implement hybrid precision training and gradient pruning techniques. During training, the teacher-forced ratio is gradually reduced to facilitate a smooth transition of the model to autoregressive prediction mode, and an early stopping mechanism preserves the optimal weights.

[0082] This section describes the construction of sample data by building multiple sample cases and their corresponding sample recurrent data. The use of the drift flux model to construct the sample data ensures its accuracy. Subsequently, the sample recurrent data can be used to train the neural network to be trained, generating a target neural network with higher accuracy.

[0083] For example, this application utilizes a validated drift flux numerical model to perform batch calculations within a parameter space designed using Latin hypercube sampling, generating high-fidelity sample datasets of velocity and sediment concentration fields. Furthermore, a neural network model composed of a convolutional network and a long short-term memory network is constructed. By controlling the neural network model to perform supervised learning on the sample dataset, a precise mapping from boundary conditions to the spatiotemporal evolution of the flow field is established. The trained deep learning model (i.e., the target neural network) can, upon receiving new operating parameters, achieve rapid inference and prediction on the order of seconds to minutes, while ensuring the accuracy of simulation predictions, providing immediate and reliable decision support for reservoir sediment scheduling and management.

[0084] This application proposes a collaborative computing framework for neural network architecture, constructing an "encoding-temporal-conditional modulation-decoding" system. The encoder (encoding module) uses convolutional layers to extract and compress spatial features from the input historical velocity and concentration fields. The core temporal module employs ConvLSTM units (long short-term memory network modules) specifically designed for two-dimensional fields to accurately capture the dynamic spatiotemporal dependencies in processes such as the evolution of the heterogeneous flow front and energy transfer. The decoder (decoding module) is responsible for upsampling and reconstructing the learned high-dimensional features back to the original physical space. This architecture surpasses traditional single-model convolutional neural networks (CNNs) or long short-term memory networks (LSTMs), achieving end-to-end learning support for spatiotemporal sequence data.

[0085] The network architecture design of this application adopts an encoder-temporal-decoder structure that integrates spatiotemporal features. The encoder extracts spatial features through multi-scale convolutional layers; the temporal modeling module introduces convolutional long short-term memory units at feature bottlenecks to capture temporal evolution patterns; the feature modulation layer incorporates boundary conditions as global variables into the network computation; and the decoder gradually restores spatial resolution through upsampling operations. The output layer uses dedicated activation functions for different physical quantities to ensure that the output results conform to physical laws.

[0086] Secondly, the neural network architecture of this application also introduces a FiLM feature linear modulation mechanism. In this application, the boundary conditions at future time steps (water level E, flow rate Q, sediment concentration C) and their derived physical fields (such as the pressure field determined by water level E, the inflow velocity determined by flow rate Q, etc.) are treated as an independent multi-channel future state tensor input. A pair of modulation parameters (γ, β) is generated through a fully connected layer and a lightweight multilayer perceptron in the modulation module. At the feature map of the deepest layer of the network, an affine transformation of the core features is achieved through the simple and effective operation of "γ×h + β". This design of this application transforms the neural network model from a static input-output mapping into a "conditional prediction system" that can dynamically respond and strictly adhere to external boundary conditions. This fundamentally avoids the errors caused by traditional interpolation methods and significantly improves the physical consistency and generalization ability of the model under unseen conditions.

[0087] This application specifically enhances the model training and inference mechanisms. Specifically, during the training phase, it employs a mask-weighted multi-channel mean squared error loss (MSE) and a multi-step autoregressive rollout strategy to ensure the model's stability in long-term predictions. During the inference phase, the target neural network model uses the prediction results of the previous step as partial input for the next step, performing closed-loop autoregressive inference. This enables it to autonomously complete the prediction of density current evolution over several days, giving it practical value as a replacement for traditional CFD simulations for long-term operational scenario extrapolation.

[0088] Optionally, after generating the predicted density current data, the method further includes: generating a density current simulation screen corresponding to the target reservoir based on the predicted density current data; and controlling the target interactive interface to display the density current simulation screen.

[0089] During implementation, after generating the predicted density current data, a density current simulation screen corresponding to the target reservoir can be generated based on the predicted density current data; and the target interactive interface can be controlled to display the density current simulation screen so that users can intuitively view the density current simulation screen through the target interactive interface, thereby improving the simulation effect of density current.

[0090] See Figure 6 The diagram shows the architecture of the density current prediction system, combined with... Figure 6 The simulation and prediction process of heterogeneous flows is illustrated by way of example. This application adopts a modular design, with each module connected through a standard data interface to form a complete pipeline of "data generation - model training - intelligent inference - analysis and application". See also... Figure 6 As shown, the core modules of this application include: a parameter sampling module, a numerical simulation module, a data preprocessing module, a deep learning training module, an inference module, a result analysis module, and a user interface module.

[0091] During implementation, the parameter sampling module can generate multiple sample cases based on the acquired sample boundary condition information (such as the user-defined range of inflow rate, water level, and sediment concentration parameters), thus generating a combination of boundary conditions covering the expected operating conditions. The numerical simulation module uses the drift flux model to solve the mixture continuity equation, momentum equation, and sediment transport equation for each sample case, generating sample density flow data corresponding to each sample case (i.e., generating high-fidelity time-series data containing velocity and concentration fields). The sample density flow data corresponding to multiple sample cases constitute the training dataset. The data preprocessing module processes the training samples in the training dataset (such as normalizing, serializing, and encapsulating the training samples into tensors), generating sample tensor data. This converts the raw field data into a standard tensor format suitable for deep learning model training, and then inputs the sample tensor data into the deep learning training module.

[0092] The deep learning training module can train the constructed neural network to obtain the target neural network using sample tensor data. The target neural network can include Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs), and can also incorporate a FiLM modulation mechanism. Through supervised learning, the target neural network model accurately grasps the spatiotemporal evolution of density currents. During the runtime phase, the inference module loads the trained model (i.e., the target neural network), accepts new operating parameters, and outputs the corresponding velocity and sediment concentration field prediction results in real time. In other words, after generating multi-frame, multi-channel tensor data based on the current density current data of the target reservoir, the target neural network deployed in the inference module processes this multi-frame, multi-channel tensor data to generate predicted density current data for the target reservoir in future time periods.

[0093] After obtaining the predicted density current data, the results analysis module can be used to determine the target engineering indicators of the target reservoir based on the predicted density current data, such as extracting key indicators such as the location of the density current intrusion point, the time of arrival in front of the dam, and the thickness of the turbid water layer; and can also generate scheduling strategies for future time periods.

[0094] The user interface module provides an interactive interface for parameter configuration, task control, and result visualization. For example, users can input sample boundary condition information on the target interactive interface. Furthermore, after obtaining the predicted density current data, the user interface module can control the target interactive interface to display the density current simulation screen for user viewing. Multiple modules are interconnected through standardized interfaces, collectively forming a comprehensive simulation system architecture that integrates physical simulation, deep learning, and engineering applications.

[0095] To improve system availability, this application includes a unified configuration and logging system built into each module. The parameter sampling and numerical simulation phases support breakpoint resume and parallel task orchestration; the training phase fully records loss decomposition, learning rate changes, gradient norm, and validation metrics, and automatically saves the optimal model checkpoint; the inference and result analysis phases generate traceable technical reports, including version number, data source, and statistical interval descriptions.

[0096] In terms of data structure design, this application adopts a professional-grade data organization scheme to ensure data processing efficiency and traceability: (1) The sample dataset is stored in HDF5 format and follows a multi-dimensional array layout of “sample × time × channel × height × width”. The channel dimension is composed of static channels and dynamic channels. The static channel contains information such as computational domain mask and spatial coordinates, while the dynamic channel contains historical physical field data and boundary condition encoding. (2) Scalar boundary conditions are stored in the form of independent datasets or attributes to ensure data integrity and access efficiency. (3) All data samples are accompanied by structured metadata, including boundary condition parameters, grid scale, timestamp and source working condition identifier, which supports rapid data location and experimental reproduction. (4) The core data structure includes: parameter combination array (i.e. sample cases), numerical result field data, training sample tensor, model weight file and index output file, which are stored in Npz, JSON, CSV and other formats respectively to meet the data processing needs of different stages.

[0097] In terms of algorithm design, the algorithm process of this application covers the entire process from data generation to prediction application: (1) The parameter sampling algorithm achieves uniform exploration of the multidimensional parameter space through Latin hypercube sampling, ensuring that the training data fully covers possible working conditions. (2) The numerical solution algorithm is based on the finite volume method discrete control equations, and adopts the implicit time integral format to ensure computational stability and accurately simulate the evolution process of the heterogeneous flow. (3) The network architecture algorithm adopts an encoder-temporal-decoder structure. The encoder extracts spatial features through multi-layer convolution, and the intermediate layer embeds ConvLSTM units to capture time dependence. The decoder restores spatial resolution through upsampling. The FiLM conditional modulation mechanism is introduced to realize the flexible injection of boundary conditions. (4) The training and inference algorithm adopts mask-weighted multi-channel mean square error loss and uses an adaptive optimization algorithm for model training. The prediction stage adopts an autoregressive strategy and realizes long-term dynamic forecasting through iterative calculation.

[0098] In terms of user interface design, this application provides an intuitive interactive interface that supports multiple operation modes: (1) Users can set boundary condition parameters through controls such as sliders and input boxes in the graphical interface, and start tasks such as parameter sampling, numerical simulation, model training and rapid prediction with one click. (2) The interface displays the animation of the density current propagation process in real time, dynamically updates the changing trends of key indicators such as the front position and deposition thickness, and supports comparative analysis of results under multiple working conditions. (3) All visualization results, including field data animations and indicator charts, can be exported to standard formats such as PNG with one click to meet the needs of professional analysis and academic exchange.

[0099] In terms of performance optimization, this application ensures efficient operation through a multi-level optimization strategy: (1) adopting hybrid precision training and inference technology to significantly improve computational efficiency and reduce GPU memory usage. (2) implementing gradient checkpointing and distributed data parallelism schemes to optimize the computational performance of large-scale training tasks. (3) optimizing the data pipeline through memory mapping and fixed memory technology to reduce I / O operation overhead. (4) supporting dynamic pruning of the computational domain and block inference mechanism to adapt to the simulation requirements of ultra-long river sections. (5) applying model distillation and weight quantization techniques to reduce resource requirements in the deployment phase. At the same time, the system fully adopts GPU-accelerated computing, implements parallel computing optimization in the numerical simulation phase, introduces a caching mechanism in the data storage phase, applies tensor quantization and model pruning in the inference phase, and achieves reasonable utilization of computing resources through asynchronous task scheduling.

[0100] This application establishes a standardized operating procedure based on the predictive inference process. During implementation, users define the prediction time range and boundary condition sequence through a configuration file, and the system automatically loads the corresponding initial field and mask file. The predictive executor performs autoregressive inference in hourly increments, generating continuous velocity and concentration field evolution results. All output data is timestamped and automatically archived, supporting result traceability and analysis. The user interface in this application provides complete visualization support. The interface displays the boundary condition process line and density current propagation animation in real time, dynamically updating key indicators such as the front position, intrusion point depth, and turbidity layer thickness. The system supports result data export to meet the needs of different application scenarios. This technical implementation scheme, through meticulous modular design, achieves efficient calculation and accurate prediction of reservoir density current simulation while maintaining physical consistency.

[0101] This application proposes a hybrid paradigm of "numerical model-generated dataset + deep learning to replace computation." The core of this paradigm lies in the following: First, a validated drift flux numerical model is used to perform large-scale offline computation within a parameter space (i.e., boundary conditions) covering possible operating conditions, generating a high-fidelity training dataset. Then, a CNN-LSTM hybrid neural network is constructed to learn the complex mapping relationship from boundary conditions to the spatiotemporal evolution of the flow field. Finally, the trained target neural network can compress simulation inference latency from hours to seconds to minutes while maintaining key physical consistency, achieving rapid and high-precision simulation and prediction of reservoir density flows. This provides a feasible technical path for real-time simulation and precise control of reservoir density flows.

[0102] This application possesses the following core innovations: First, in terms of model construction, it creatively proposes a hybrid modeling framework that combines a drift flux model with a clear physical foundation with a CNN-LSTM deep learning model with powerful spatiotemporal feature learning capabilities. This combination strategy enables the model to maintain the consistency of the physical process while achieving a significant leap in computational efficiency.

[0103] Secondly, regarding network structure design, a UNet-ConvLSTM-FiLM hybrid neural network architecture was designed to address the characteristics of density currents in reservoirs. This architecture, through a feature linear modulation layer, achieves continuous dynamic response to boundary conditions such as flow rate and water level, effectively avoiding errors caused by traditional discrete interpolation. Simultaneously, the network explicitly encodes physical prior knowledge such as wetland masks and boundary constraints, significantly reducing the model's requirement for a large number of training samples without increasing data annotation costs.

[0104] Third, regarding training optimization strategies, multi-step prediction consistency constraints and boundary physical regularization methods are introduced. Their synergistic effect effectively suppresses error accumulation in long-term predictions, enabling the model to maintain rolling prediction stability for up to 7 days. At the engineering application level, the training dataset is scientifically constructed using Latin hypercube sampling technology, ensuring the model's generalization ability across a wide range of operating conditions. Furthermore, the entire system adopts a modular design, providing a complete training-inference pipeline and a visual indicator dashboard. It not only supports real-time and near-real-time predictions, providing immediate decision support for reservoir scheduling and ecological management, but also possesses excellent portability, allowing for easy migration to other reservoirs or similar density current scenarios in hydrodynamic problems.

[0105] Compared with traditional hydro-sediment numerical simulation methods, this application demonstrates significant advantages in computational efficiency, prediction accuracy, and engineering applicability. Firstly, in terms of computational efficiency, traditional CFD simulations typically require dozens to hundreds of CPU cores and several hours or even days to complete a simulation of a typical flood event. In contrast, this application, after offline training, can complete dynamic rolling predictions of density flows lasting 1 to 7 days within tens of seconds to minutes using only a single GPU, achieving an overall speedup of 10-10. 4 -10 5 The scale has been expanded from "hours" to "seconds", enabling real-time scheduling.

[0106] Secondly, regarding prediction accuracy, the model demonstrates excellent generalization ability under various previously unseen operating conditions. The relative error in predicting key indicators such as the location of the density current front, the thickness of the main flow band, and the volume of the high-turbidity zone is generally controlled within 20%. Even in extreme conditions with drastic changes in the inflow process, the model maintains stable prediction performance by introducing boundary process priors and adaptive correction mechanisms. Furthermore, in terms of engineering applications, this application can quickly output sediment concentration fields, velocity fields, and derived indicators (such as sedimentation thickness and intrusion point location), directly supporting reservoir sediment discharge scheduling, sedimentation assessment, and ecological management decisions, demonstrating significant practicality and potential for wider application.

[0107] The following explanation uses the Xiaolangdi Reservoir on the Yellow River as an example. First, based on actual operational data from the Xiaolangdi Reservoir, reasonable ranges for key parameters are determined. For example, the inflow rate is set at 2000-4000 m³ / s, the water level fluctuation range is 220-230 meters, and the sediment concentration variation range is 20-200 kg / m³. Based on this parameter space, 100 representative operating condition combinations are generated through a parameter sampling module to ensure that the training data fully covers all possible scenarios in the actual operation of the reservoir.

[0108] Secondly, the numerical simulation module is activated, and a validated drift flux numerical model is used to perform detailed calculations for each set of parameters. Each working condition simulates the complete hydro-sediment evolution process, outputting time-series data including velocity distribution and sediment concentration field, forming a high-quality training dataset.

[0109] After obtaining the raw simulation data, the data preprocessing stage begins. This stage involves rigorous quality control checks, standardization, and serialization of the numerical results. The flow field and concentration field of ten consecutive time steps are constructed into an input sequence, and subsequent time steps are used as prediction targets to form standardized training samples that meet the requirements of deep learning.

[0110] After data preparation, a CNN-LSTM-FiLM hybrid network with multi-scale features is constructed, and a phased training strategy is adopted: in the early stage, the reconstruction accuracy is the main optimization goal, and in the later stage of training, the learning rate is dynamically adjusted, the model's performance on the validation set is closely monitored, and the model weights with the best performance are automatically saved.

[0111] After the model is trained and deployed, the parameters of the sediment discharge scheduling scheme to be evaluated are input into the inference module, such as the inflow rate of 3000 m³ / s, the water level of 225 meters, and the changes in sediment concentration gradient. The inference module performs rapid extrapolation based on the trained model, generating a prediction of the entire process of the density current development over the next 7 days.

[0112] Finally, the results analysis module automatically extracts key characteristic indicators such as the advance rate of the density current front, the extent of the turbid water layer expansion, and the distribution of major sedimentation areas, providing a quantitative decision-making basis for the selection of sediment discharge timing and the dynamic adjustment of scheduling schemes.

[0113] Through the above process, the prediction accuracy and computational efficiency of the deep learning alternative model in this application were evaluated in a real reservoir environment, and its practical value in optimizing sediment discharge scheduling schemes was verified. Comparison with high-fidelity numerical simulation results shows that this application exhibits excellent prediction performance. In terms of computational efficiency, it achieves an order-of-magnitude improvement. Traditional CFD numerical simulation requires tens of hours of computation time to complete a 7-day evolution prediction, while this application can complete it in only 2-3 minutes under the same hardware configuration, improving computational efficiency by more than 1000 times, fully demonstrating the significant advantages of the deep learning alternative model. Furthermore, it is particularly noteworthy that the model demonstrates good adaptability and robustness to complex operating conditions. Whether it is a slow change in reservoir water level or a sharp fluctuation in inflow and sediment concentration, the model's prediction results accurately reflect the response law of density current to changes in boundary conditions, providing reliable technical support for sediment discharge scheduling of the Xiaolangdi Reservoir.

[0114] Therefore, the method provided in this application utilizes numerical simulation of drift flux to generate training data, and learns the spatiotemporal evolution of density flows through a CNN-LSTM-FiLM neural network, achieving rapid prediction of density flows in complex reservoirs. Furthermore, the system structure provided in this application is reasonable and the modules are clearly divided, which not only significantly improves computational efficiency and reduces hardware requirements, but also maintains high prediction accuracy. Through parameter sampling design, the model possesses strong generalization ability and can be applied to reservoirs with different inflow conditions and geometries. It also has broad application prospects and significant practical value in reservoir scheduling, sediment management, and ecological protection.

[0115] Corresponding to the aforementioned embodiments of the reservoir density current prediction method, this application also provides embodiments of a reservoir density current prediction device. Figure 7A schematic diagram of the reservoir density current prediction device provided in this application specifically includes: The acquisition module 701 is used to acquire the current density current data of the target reservoir, wherein the current density current data includes velocity field data and concentration field data corresponding to multiple times within the current time period. The velocity field data at each time includes the flow velocity information at multiple spatial locations of the target reservoir at that time, and the concentration field data at each time includes the concentration information at multiple spatial locations of the target reservoir at that time. The generation module 702 is used to generate multi-frame multi-channel tensor data based on the current heterogeneous flow data; The inference module 703 is used to process the multi-frame, multi-channel tensor data using a target neural network for predicting density currents, and generate predicted density current data for the target reservoir in the future time period.

[0116] In an optional embodiment, the apparatus further includes: a result analysis module 704, used for: Based on the predicted density current data, the target engineering indicators of the target reservoir are determined. Based on the target engineering indicators of the target reservoir, a scheduling strategy for the target reservoir in the future time period is generated.

[0117] In one optional implementation, the target neural network includes an encoding module, a decoding module, and a long short-term memory network module. The inference module 703, when using the target neural network for predicting density currents to process the multi-frame, multi-channel tensor data and generate predicted density current data for the target reservoir in future time periods, is used for: The encoding module is used to extract features from the multi-frame, multi-channel tensor data to generate first feature data; The first feature data is processed using the Long Short-Term Memory network module to generate the second feature data; The decoding module is used to decode the first feature data and the second feature data to generate predicted density current data for the target reservoir in the future time period.

[0118] In one optional implementation, the encoding module includes m first convolutional processing units, the decoding module includes m second convolutional processing units, the first feature data includes intermediate feature data output by the m first convolutional processing units, the feature sizes of the m intermediate feature data are different, and m is a positive integer greater than 1. The reasoning module 703, when using the long short-term memory network module to perform feature processing on the first feature data to generate the second feature data, is used to: use the long short-term memory network module to perform feature processing on the intermediate feature data with the smallest feature size in the first feature data to generate the second feature data. The reasoning module 703, when using the decoding module to decode the first feature data and the second feature data to generate the predicted density current data of the target reservoir in the future time period, is used for: The first second convolution processing unit is used to process the second feature data and the first intermediate feature data to generate the first third feature data; The i-th second convolution processing unit is used to process the (i-1)-th third feature data and the i-th intermediate feature data in the first feature data to generate the i-th third feature data; where i is a positive integer greater than 1 and less than or equal to m, and the feature size of the i-th intermediate feature data is greater than the feature size of the (i-1)-th intermediate feature data. Based on the m-th third feature data, predictive density current data for the target reservoir in future time periods are generated.

[0119] In one optional implementation, the target neural network includes a modulation module, the modulation module being further configured to: Obtain the reservoir boundary condition information of the target reservoir in a future time period, including water level boundary information, flow rate boundary information, and sediment content boundary information; Based on the reservoir boundary condition information, multi-channel state data is generated; The modulation module is used to process the multi-channel state data to generate modulation parameters; After using the Long Short-Term Memory network module to perform feature processing on the intermediate feature data with the smallest feature size in the first feature data to generate the second feature data, the inference module 703 is further configured to: The second feature data is modulated using the modulation parameters to generate processed second feature data, wherein the decoding module is used to decode the first feature data and the processed second feature data.

[0120] In one optional implementation, the multi-channel tensor data includes static channels and dynamic channels, wherein the static channels include a computational domain mask channel, a time channel, and a spatial location channel; and the dynamic channels include a flow rate channel, a concentration channel, and a boundary condition channel.

[0121] In an optional embodiment, the apparatus further includes a deep learning training module 705 for training the target neural network according to the following steps: Based on the obtained sample boundary condition information, multiple sample cases are generated, wherein each sample case includes water level sample information, flow rate sample information, and sediment concentration sample information; For each of the aforementioned sample cases, the drift flux model is used to generate sample recurrent data corresponding to the sample case; Based on the sample recurrent data corresponding to the multiple sample cases, multiple sample tensor data are generated; Using the multiple sample tensor data, the constructed neural network for predicting heterogeneous currents is trained in multiple rounds until the training cutoff condition is met, thereby generating the target neural network.

[0122] In an optional embodiment, the device further includes a user interface module 706, which, when acquiring sample boundary condition information, is used to: The target's interactive interface displays the parameter setting page; In response to a setting operation triggered by the user on the parameter setting page, obtain the sample boundary condition information set by the user.

[0123] In one optional implementation, after generating the predicted heterogeneous flow data, the user interface module 706 is further configured to: Based on the predicted density current data, a density current simulation image corresponding to the target reservoir is generated; The target interactive interface is controlled to display the simulated density flow.

[0124] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0125] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0126] This application also provides a computer-readable storage medium storing a computer program that can be used to execute the reservoir density current prediction method described in the above embodiments.

[0127] This application also provides a computer device, see [link to relevant documentation] Figure 8 The diagram shown is a structural schematic of the computer device provided in this application. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the reservoir density current prediction method described in the above embodiments. Of course, in addition to software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0128] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0129] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0130] 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.

[0131] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the functions specified in one or more boxes. In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0132] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0133] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0134] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0135] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0136] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0137] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for predicting density currents in a reservoir, characterized in that, The method includes: Acquire the current density current data of the target reservoir, wherein the current density current data includes velocity field data and concentration field data corresponding to multiple times within the current time period, the velocity field data at each time includes the flow velocity information at multiple spatial locations of the target reservoir at that time, and the concentration field data at each time includes the concentration information at multiple spatial locations of the target reservoir at that time. Based on the current heterogeneous flow data, generate multi-frame, multi-channel tensor data; The multi-frame, multi-channel tensor data is processed using a target neural network for density current prediction to generate predicted density current data for the target reservoir in the future time period.

2. The method according to claim 1, characterized in that, The method further includes: Based on the predicted density current data, the target engineering indicators of the target reservoir are determined. Based on the target engineering indicators of the target reservoir, a scheduling strategy for the target reservoir in the future time period is generated.

3. The method according to claim 1, characterized in that, The target neural network includes an encoding module, a decoding module, and a long short-term memory network module. The process of using this target neural network for recurrent prediction to process the multi-frame, multi-channel tensor data to generate predicted recurrent data for the target reservoir in the future includes: The encoding module is used to extract features from the multi-frame, multi-channel tensor data to generate first feature data; The first feature data is processed using the Long Short-Term Memory network module to generate the second feature data; The decoding module is used to decode the first feature data and the second feature data to generate predicted density current data for the target reservoir in the future time period.

4. The method according to claim 3, characterized in that, The encoding module includes m first convolutional processing units, the decoding module includes m second convolutional processing units, the first feature data includes intermediate feature data output by the m first convolutional processing units, the feature sizes of the m intermediate feature data are different, and m is a positive integer greater than 1; The step of using the Long Short-Term Memory network module to perform feature processing on the first feature data to generate second feature data includes: The Long Short-Term Memory network module is used to perform feature processing on the intermediate feature data with the smallest feature size in the first feature data to generate the second feature data. The step of using the decoding module to decode the first feature data and the second feature data to generate predicted density current data for the target reservoir in the future time period includes: The first second convolution processing unit is used to process the second feature data and the first intermediate feature data to generate the first third feature data; The i-th second convolution processing unit is used to process the (i-1)-th third feature data and the i-th intermediate feature data in the first feature data to generate the i-th third feature data; where i is a positive integer greater than 1 and less than or equal to m, and the feature size of the i-th intermediate feature data is greater than the feature size of the (i-1)-th intermediate feature data. Based on the m-th third feature data, predictive density current data for the target reservoir in future time periods are generated.

5. The method according to claim 4, characterized in that, The target neural network includes a modulation module, and the method further includes: Obtain the reservoir boundary condition information of the target reservoir in a future time period, including water level boundary information, flow rate boundary information, and sediment content boundary information; Based on the reservoir boundary condition information, multi-channel state data is generated; The modulation module is used to process the multi-channel state data to generate modulation parameters; After using the Long Short-Term Memory network module to perform feature processing on the intermediate feature data with the smallest feature size in the first feature data to generate the second feature data, the method further includes: The second feature data is modulated using the modulation parameters to generate processed second feature data, wherein the decoding module is used to decode the first feature data and the processed second feature data.

6. The method according to claim 1, characterized in that, The multi-channel tensor data includes static channels and dynamic channels, wherein the static channels include computational domain mask channels, temporal channels, and spatial location channels; The dynamic channels include: flow rate channel, concentration channel, and boundary condition channel.

7. The method according to any one of claims 1-6, characterized in that, The target neural network is trained according to the following steps: Based on the obtained sample boundary condition information, multiple sample cases are generated, wherein each sample case includes water level sample information, flow rate sample information, and sediment concentration sample information; For each of the aforementioned sample cases, the drift flux model is used to generate sample recurrent data corresponding to the sample case; Based on the sample recurrent data corresponding to the multiple sample cases, multiple sample tensor data are generated; Using the multiple sample tensor data, the constructed neural network for predicting heterogeneous currents is trained in multiple rounds until the training cutoff condition is met, thereby generating the target neural network.

8. The method according to claim 7, characterized in that, Obtain sample boundary condition information, including: The target's interactive interface displays the parameter setting page; In response to a setting operation triggered by the user on the parameter setting page, obtain the sample boundary condition information set by the user.

9. The method according to claim 8, characterized in that, After generating the predicted density flow data, the method further includes: Based on the predicted density current data, a density current simulation image corresponding to the target reservoir is generated; The target interactive interface is controlled to display the simulated density flow.

10. A device for predicting density currents in a reservoir, characterized in that, The device includes: The acquisition module is used to acquire the current density current data of the target reservoir. The current density current data includes velocity field data and concentration field data corresponding to multiple moments within the current time period. The velocity field data at each moment includes the flow velocity information at multiple spatial locations of the target reservoir at that moment, and the concentration field data at each moment includes the concentration information at multiple spatial locations of the target reservoir at that moment. The generation module is used to generate multi-frame, multi-channel tensor data based on the current heterogeneous flow data; The inference module is used to process the multi-frame, multi-channel tensor data using a target neural network for density current prediction, and generate predicted density current data for the target reservoir in the future time period.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method according to any one of claims 1-9.

12. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor performs the steps of the method according to any one of claims 1-9.

Images