Ecological evolution simulation method based on rainfall runoff assimilation

Abstract

The invention discloses an ecological evolution simulation method based on rainfall runoff assimilation. The method comprises the following steps: estimating environmental effects such as terrain erosion, water resource distribution change and photo-thermal resource change by using a rainfall runoff assimilation result output by a rainfall runoff assimilation model; on the basis, the resource supply condition and the species distribution condition of each layer of the food chain are calculated, and the change of a target environment ecosphere is simulated. The rainfall runoff assimilation model is used for simulating the rainfall runoff process, the short-term influence of the rainfall runoff is added in ecological environment evolution simulation, the granularity of the ecological evolution simulation method is improved, and the rainfall runoff assimilation for promoting ecological coupling is more accurate.

Description

Technical Field

[0001] The present invention provides an ecological evolution simulation method based on rainfall runoff assimilation, which relates to the fields of deep learning technology, rainfall runoff assimilation technology and ecological environment evolution simulation technology. Background Art

[0002] Traditional ecological environment evolution simulation technology mainly considers the impact of the atmosphere on ecological environment changes from the perspective of climate. When simulating the ecological environment, the granularity is low, which is suitable for simulating long-term ecological environment changes. The rainfall runoff process has highly nonlinear and non-stationary characteristics, which can change the light and heat resources, water resources and terrain of the region in a short period of time, and has a great impact on the changes of the ecological environment. This impact sometimes shapes a small climate, causing the ecological environment to change. Therefore, the present invention integrates the rainfall runoff assimilation model and the ecological environment evolution, and improves the granularity of the ecological environment evolution simulation method by simulating the rainfall runoff process, and also promotes the ecological coupling rainfall runoff assimilation to be more accurate.

[0003] Traditional methods mainly rely on process-driven hydrological models, set some parameters with physical significance according to the hydrological process, establish complex algorithms to link the parameters based on existing knowledge, and construct hydrological models. When applied, the model parameters are manually adjusted according to the actual observation data to make them fit the actual situation. However, due to the high complexity and many parameters of the model, the process-driven hydrological model has some shortcomings in practical applications.

[0004] Data-driven methods, such as regression analysis and time series analysis, predict rainfall and runoff by capturing numerical trends. Although they can reveal the laws of historical data to a certain extent, their prediction accuracy and stability are limited for complex hydro-meteorological systems. In recent years, with the advancement of machine learning technology, emerging methods such as traditional machine learning and deep learning have been applied to rainfall and runoff assimilation. These methods can better capture the nonlinear characteristics of hydro-meteorological systems, thereby improving the accuracy and stability of forecasts. However, for complex runoff predictions, the prediction accuracy of these technologies has declined seriously. This patent introduces an input transposition mechanism to enable the model to focus more on the sequence characteristics of the variables themselves, give full play to the Transformer's ability to model nonlinear runoff sequences, and improve the model's learning ability and prediction accuracy for complex data patterns of runoff sequences. Summary of the invention

[0005] The purpose of the present invention is to provide an ecological evolution simulation method based on rainfall-runoff assimilation, which simulates the rainfall-runoff process by integrating the rainfall-runoff assimilation model, improves the granularity of the ecological simulation method, and fills the gap in the prior art.

[0006] Specifically, the present invention provides an ecological evolution simulation method based on rainfall runoff assimilation, comprising the following steps:

[0007] Step 1, input the basic geographical data of the simulation area; give a K-layer food chain with the first layer being plants; for each layer of the food chain, there will be a corresponding resource supply map (RSG) and species density map (SDM);

[0008] Step 2, input the basin-wide node hydrological dataset of the simulated regional rivers with a time scale of days and a time span of no less than 1 year, and use the neural network model for rainfall-runoff assimilation training;

[0009] Step 3: Determine the climate characteristics of the target environment and estimate its light and heat resources by combining the latitude and longitude, altitude, sea level extreme temperature and rainfall runoff assimilation model;

[0010] Step 4, combining the digital elevation model, terrain directional map, and mineral map, adding the terrain, runoff, and watershed of the target environment, and initializing the target environment;

[0011] Step 5: Construct the first layer of RSG based on the initial topography, runoff and watershed, and light and heat resources of the target environment 0 , and construct subsequent layers in turn to obtain the initial RSG; for subsequent layers RSG 1 ~RSG K-1 , iteratively using RSG i-1 Building RSG i ;

[0012] Step 6: Continue simulation iteration on an annual scale, including topography, SDM, and RSG; finally output the latest topography, SDM, and RSG.

[0013] Furthermore, the basic geographic data in step 1 include a digital elevation model, a terrain direction map, a mineral map, longitude and latitude, altitude and extreme sea level temperature of the target environment.

[0014] Furthermore, the specific steps of step 2 are as follows:

[0015] Step 2.1, input the basin-wide node hydrological dataset of the simulated regional rivers with a time scale of days and a time span of no less than 1 year. The node data consists of rainfall and runoff;

[0016] Step 2.2, construct a Transformer model with input transposition as a rainfall-runoff assimilation model; the rainfall-runoff assimilation model is constructed by inputting and predict and Where P is rainfall, Q is runoff, i represents the current date to be predicted, j represents the current node to be predicted, and j-1 represents the upstream node of j; the rainfall-runoff assimilation model realizes rainfall-runoff assimilation by using the method of reusing predicted values;

[0017] Step 2.3, setting the loss function and optimizer, selecting appropriate training hyperparameters, and training the rainfall-runoff assimilation model.

[0018] Furthermore, the Transformer model with input transposition in step 2.2 combines the values ​​of the same variable at different time steps into a vector to perform an embedding operation in the embedding layer. Specifically, for an input variable X containing N time steps, each time step records M features, and its shape is N×M, the calculation formula for transposed embedding is in Represents a fully connected layer with input dimension N and output dimension D.

[0019] Furthermore, the hidden layer of the input transposed Transformer model described in step 2.2 is implemented by stacking the multi-head attention encoders of the Transformer model, and the forward propagation calculation steps of each encoder are as follows:

[0020] Step T1, for input X, calculate the multi-head attention module output X attn :

[0021] X attn =MultiHead(Q,K,V)+X=Concat(head l , ..., head h )W O +X,

[0022] in

[0023] Step T2, for the multi-head attention module output X attn , calculate the feedforward module output X FFN :X FFN =max(0, X attn W1+b2)W2+b2+X attn , X FFN That is the result of the encoder forward propagation;

[0024] In the formula, Concat(*, **) represents tensor concatenation operation, softmax(*) represents softmax function, and max(*, **) represents maximum value function.

[0025] Furthermore, the specific steps of step 3 are as follows:

[0026] Step 3.1, obtaining runoff observation data with a time span of one year from the rainfall-runoff assimilation model in step 2, accumulating the rainfall at each node on the same date, and statistically analyzing annual rainfall characteristics such as the total annual rainfall, annual rainfall days, maximum monthly rainfall, and minimum monthly rainfall;

[0027] Step 3.2, based on the existing geographical knowledge, combined with the annual rainfall characteristics, longitude and latitude, altitude, and extreme sea level temperature, determine the climate characteristics of the target environment and estimate the total annual solar thermal resources (referred to as solar thermal resources).

[0028] Furthermore, the specific steps of step 5 are as follows:

[0029] Step 5.1: Construct the first layer of RSG based on the topography, runoff and watershed, and light and heat resources of the initial target environment 0 ;RSG 0 Represent each cell on the map as a node on the graph;

[0030] Step 5.2, according to RSG i-1 Building RSG i Until RSG is completed 1 ~RSG K-1 Build, and execute steps 5.3.1 to 5.3.4 each time;

[0031] Step 5.3.1, for the resources of each species at the i-th level of the food chain, encode their location on the terrain and their accessibility;

[0032] Step 5.3.2, connect multiple nodes into local areas, calculate the fitness score of species at the i-th level of the food chain in each local area, and simulate the competition between species at the i-th level based on the fitness score;

[0033] Step 5.3.3, according to the simulated competition results, set the number of population units for each species in each region to obtain the i-th layer SDM i ;

[0034] Step 5.3.4, from RSG i-1 and SDM i The remaining graph for calculating unused resources in And create RSG i .

[0035] 6. Further, the specific steps of step 6 are as follows:

[0036] Step 6.1, obtaining one year's rainfall runoff assimilation data from the rainfall runoff assimilation model, performing hydrodynamic analysis on the runoff of each node during the flood period in combination with the terrain, calculating possible river diversions, and updating the terrain;

[0037] Step 6.2: Update RSG according to the rainfall at each node 0 The water resource node in RSG is completed 0 Updates;

[0038] Step 6.3, according to RSG i-1 Update RSG i Until RSG is fully updated 1 ~RSG K-1 , the steps are the same as those in step 5.2;

[0039] Step 6.4, output the final terrain, SDM and RSG as results.

[0040] The parts not involved in the present invention are the same as the prior art or can be implemented by using the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The present invention is a flowchart of the method.

[0042] Figure 2 It is the rainfall-runoff assimilation result of the rainfall-runoff assimilation model described in the present invention.

[0043] Figure 3 This is the RSG described in the present invention.

[0044] Figure 4 This is the relationship between RSG and SDM described in the present invention. DETAILED DESCRIPTION

[0045] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0046] 1. Embedding mechanism of input transposition

[0047] For an input X containing N time steps, each time step records M features, and the shape is N×M, the time dimension and feature dimension are exchanged through the matrix transpose operation to obtain X T , with a shape of M×N, performs embedding. Input transposition enables the model to apply self-attention in the time dimension and directly learn the temporal correlation of the input data.

[0048] 2. Self-Attention Mechanism

[0049] For each input sequence X, multiply it by the weight matrix W Q , W K and W V Get the query matrix Q, key matrix K and value matrix V, calculate QK T Get the attention score, divided by To scale it to avoid numerical overflow, the scaled attention score is converted into a weight through the softmax function, and the weight is multiplied by V to get the output of self-attention.

[0050] 3. RSG composition

[0051] RSG is a set of K graphs, corresponding to each layer of the food chain. Figure 3 Given a three-layer food chain RSG, the subscripts from small to large are the resource supply status of plants, herbivores, and carnivores. Taking RSG1 as an example, the area delineated by black lines is the main activity area of ​​antelopes, the black arrows are the daily activity tracks of antelopes, and the gray arrows indicate the connectivity of the same resources. RSG1 shows that antelopes have four main activity areas. Due to the terrain, they only move between two of them, forming three relatively closed activity areas; the largest activity area has the largest number and the most complete types of resources.

Claims

1. An ecological evolution simulation method based on rainfall runoff assimilation, belonging to the category of the combination of computer graphics and artificial intelligence, characterized in that: The steps include: Step 1, input the basic geographical data of the simulated area; give a K-layer food chain with the first layer being plants; for each layer of the food chain, there will be a corresponding resource supply map (RSG) and species density map (SDM); Step 2, input the basin-wide node hydrological dataset of the simulated regional rivers with a time scale of daily and a time span of no less than 1 year, and use the neural network model for rainfall-runoff assimilation training; Step 3: Determine the climate characteristics of the target environment and estimate its light and heat resources by combining latitude and longitude, altitude, extreme sea level temperature, and rainfall-runoff assimilation model; Step 4, combining the digital elevation model, terrain directional map, and mineral map, adding the terrain, runoff, and watershed of the target environment, and initializing the target environment; Step 5: Construct the first layer of RSG based on the initial topography, runoff and watershed, and light and heat resources of the target environment 0 , and construct subsequent layers in turn to obtain the initial RSG; for subsequent layers RSG 1 ~RSG K-1 , iteratively using RSG i-1 Building RSG i ; Step 6: Continue simulation iterations on an annual scale, including topography, SDM, and RSG; The output is the latest terrain, SDM and RSG.

2. The ecological evolution simulation method based on rainfall runoff assimilation according to claim 1, characterized in that: The basic geographic data in step 1 include a digital elevation model, a terrain direction map, a mineral map, longitude and latitude, altitude and extreme sea level temperature of the target environment.

3. The ecological evolution simulation method based on rainfall runoff assimilation according to claim 1, characterized in that: In step 2, the Transformer model with input transposition is used as the neural network model for rainfall runoff assimilation training. In the embedding layer, the values ​​of the same variable at different time steps are combined into vectors to perform embedding operations. The calculation formula is: Where N is the number of time steps, M is the number of features, Represents a fully connected layer with input dimension N and output dimension D.

4. The ecological evolution simulation method based on rainfall runoff assimilation according to claim 1, characterized in that: In step 5, the adaptation scores of species at the i-th level of the food chain are calculated in the local area, and the competition between species in each local area is simulated to generate the i-th level SDM of the SDM. i ; The adaptability score calculation formula is as follows: Among them, s represents the species, R represents the necessary resource set, r represents a necessary resource, and r min 、r max are the minimum and maximum values ​​of resource r in a year respectively, and F(s, r) represents the demand value function of species s for resource r.

Images