Search

CN-121980961-A - Complex river network flow prediction research of diffusion space-time diagram neural network based on physical information guidance

CN121980961ACN 121980961 ACN121980961 ACN 121980961ACN-121980961-A

Abstract

The invention discloses a complex river network flow prediction method based on a physical information-guided diffusion space-time diagram neural network, and relates to the technical field of intelligent water conservancy and hydrologic time sequence prediction. Aiming at the problem of lacking physical consistency and insufficient long-distance space-time dependency capturing capability of the existing data driving model, the method constructs a diffusion space-time diagram neural network (PI-DiffusionSTG) guided by physical information. The model designs a bidirectional graph diffusion convolution module, respectively simulates forward hydraulic conduction and reverse jacking effects in a river network through forward diffusion and backward diffusion, effectively extracts bidirectional space-time characteristics, adopts a sequence-to-sequence architecture to conduct multi-step flow prediction, simultaneously takes a san-View southern range group describing fluid mechanics as a physical constraint regularization term depth blending loss function, and utilizes a mixing strategy of combining automatic differential with finite difference on a graph to calculate physical residual errors. The method can remarkably improve the prediction precision of the complex river network under extreme conditions, ensure that the prediction result accords with the law of conservation of mass and momentum, and has stronger robustness and physical interpretability.

Inventors

  • LIU JIABAO
  • LONG YIN

Assignees

  • 西南科技大学

Dates

Publication Date
20260505
Application Date
20260302

Claims (10)

  1. 1. A river network flood prediction method based on a physical information-guided diffusion space-time diagram neural network is characterized by comprising the following steps: Firstly, constructing a river network diagram structure with hydrologic stations as nodes and river channel communication relations as directed edges, and acquiring historical flow, water level, corresponding rainfall and other input characteristics of each node; Step two, performing two-way graph diffusion convolution on the river network graph at each time step, wherein the step two comprises the following steps: 1, constructing a forward transfer matrix and a backward transfer matrix according to the upstream and downstream directions of river channel topology; 2, carrying out K-step forward flow diffusion convolution on the node characteristics based on the forward transfer matrix so as to simulate hydraulic conduction of upstream incoming water; 3, carrying out K-step countercurrent diffusion convolution on node characteristics based on the backward transfer matrix so as to simulate the jacking effect of a high water level at the downstream to the upstream; 4, fusing the forward flow characteristic and the backward flow characteristic and executing nonlinear activation to obtain the spatial characteristic embedding of the node; step three, inputting the spatial feature sequence obtained in the step two into a sequence model, and generating river network flow prediction of a plurality of time steps in the future through an encoder-decoder structure; Step four, constructing a physical loss function based on physical constraint, which comprises the following steps: 1 calculating the partial derivative of the predicted flow rate with respect to time by utilizing automatic differentiation; 2, calculating the partial derivative of the predicted flow rate to the space by adopting a central difference or a single-side difference on the river network diagram; 3, calculating a mass conservation residual error and a momentum conservation residual error according to the san View continuity equation and the momentum equation; And fifthly, constructing a composite loss function by jointly restricting data loss, physical loss and mass conservation of a confluence point, and enabling a prediction result to simultaneously meet observation data and a hydrodynamic rule through iterative training.
  2. 2. The river network flood prediction method according to claim 1, wherein the K-step diffusion convolution in the forward direction in the second step is used to capture the long-distance effect of the upstream multi-hop neighborhood on the downstream node.
  3. 3. The river network flood prediction method according to claim 1, wherein the K-step diffusion convolution in the countercurrent direction in the second step is used for simulating backwater and jacking effects caused by the rise of the downstream water level.
  4. 4. The river network flood prediction method of claim 1, wherein the forward diffusion convolution and the backward diffusion convolution respectively employ different learnable weight matrices to distinguish two types of spatial physical processes.
  5. 5. The river network flood prediction method of claim 1, wherein the encoder in step three is configured to extract spatiotemporal features of P historical time steps and generate context vectors, and the decoder generates prediction sequences of F future time steps in an autoregressive manner based on the context vectors.
  6. 6. The river network flood prediction method according to claim 1, wherein the mass conservation residual in the fourth step is calculated by the san winan continuity equation ∂ a/∂ t+ ∂ Q/∂ x=q l .
  7. 7. The river network flood prediction method according to claim 1, wherein the momentum conservation residual in the fourth step is obtained by calculating a san france momentum equation ∂ Q/∂ t+ ∂ (Q2/a)/∂ x+ga ∂ h/∂ x+ gAQ _f=0.
  8. 8. The river network flood prediction method according to claim 1, wherein the center difference of the spatial derivatives in the fourth step is calculated by ∂ Q/∂ x (Q_downlink-Q_upstream)/Deltax.
  9. 9. The river network flood prediction method of claim 1, wherein the composite loss function is comprised of data loss, physical residual loss, and junction boundary condition loss weighting.
  10. 10. The river network flood prediction method according to claim 1, which is suitable for flood real-time prediction and scheduling assistance of different types of watercourses such as plain, hills and mountains.

Description

Complex river network flow prediction research of diffusion space-time diagram neural network based on physical information guidance Technical Field The invention relates to the field of hydrodynamics simulation and artificial intelligence prediction, in particular to a complex river network flood prediction method based on a diffusion space-time diagram neural network guided by physical information. Background The flood prediction is a key technical link in a flood prevention and disaster reduction system, and the key task is to accurately predict the flow change of a future period in a river network so as to support flood prevention scheduling of a river basin, reservoir operation and emergency response. In the prior art, the flood prediction method mainly comprises two major types, namely a physical driving model and a data driving model. The physical driving model is usually constructed based on hydrodynamic partial differential equations such as a one-dimensional or two-dimensional Save Vietnam equation, a shallow water equation and the like, and has definite physical interpretability on the flood wave propagation process. However, in the solving process of the model, a large number of refined basic parameters such as river section, roughness, topography, boundary conditions and the like are required, and meanwhile, a very small time step and a high-density grid are required, so that the calculation cost is high and the real-time performance is poor. In a complex river network, due to factors such as branch-main flow multipoint confluence, local river reach mutation, reverse water surface gradient occurrence and the like, the model calibration difficulty is high, calculation error accumulation is easy to cause, and therefore the application of the model in a fast prediction and large-scale integrated prediction scene is limited. With the development of deep learning, more and more researches adopt a data driving model to realize flood prediction, including methods such as an artificial neural network (RNN), a cyclic neural network (RNN), a long-term and short-term memory network (LSTM) and the like. The model can mine the nonlinear relation between the flow and the rainfall sequence, has high reasoning speed after training is finished, and is suitable for real-time prediction. However, the method usually regards the river basin as an independent time sequence or a simple lumped unit, ignores the spatial topological relation between the upstream and the downstream of the river channel, is difficult to capture the hydrodynamic process propagated along the river channel, and particularly cannot simulate the jacking or reverse flow effect of the main flow caused by the high water level of the main flow on the tributary. In addition, the pure data driving model lacks physical constraints such as mass conservation, momentum conservation and the like, and the prediction result against the hydrodynamic rule is easy to generate in the extreme flood event, so that the robustness and generalization capability are insufficient. In order to introduce the space structure information of the river network, a space-time diagram convolution network is partially researched. The hydrologic site is modeled as graph nodes, and the river network topology structure is represented by the adjacency matrix. However, the existing graph convolution method generally adopts an isotropic neighborhood aggregation mechanism, the difference between upstream forward flow influence and downstream reverse pressure cannot be distinguished, and the multilayer graph convolution is easy to generate an over-smooth phenomenon, so that node characteristics tend to be homogenized, and the effectiveness of a model in long-distance dependent modeling is reduced. Meanwhile, the method is still in a pure data driving paradigm, lacks physical constraint and still has obvious errors in complex river crossing areas. On the other hand, the complex river network topology results in flood evolution exhibiting obvious nonlinear characteristics. At the junction of the main and branch flows, the downstream main flow flood peak may form a significant jacking phenomenon on the branch flows, which affects the flow discharge capacity of the branch flows and even leads to the backward flow of the water flow in extreme cases. The phenomenon makes node flow no longer be simple linear superposition, and the arrival time and amplitude of flood peak highly depend on the relative time sequence of main and branch flood peak, so that the traditional empirical method and the lumped model are difficult to be applicable. In addition, in high-energy environments such as mountainous areas, flood water presents typical 'swell break' characteristics, the time scale is short and severe, the existing model is insufficient in high-frequency dynamic response capability, and the prediction difficulty is further increased. In summary, the existing flood prediction technology mainly fac