Rainfall station network layout method

Abstract

The invention discloses a rainfall station network layout method, and relates to the technical field of hydrometric station network layout, and the method comprises the steps: obtaining the rainfall data of a historical rainfall station network in a target region, the historical rainfall station network comprising a plurality of rainfall stations; extracting spatio-temporal characteristics of the rainfall data, and obtaining a rainfall prediction value of each spatial unit based on the spatio-temporal characteristics; obtaining prediction precision and a mutual information mean value based on the rainfall prediction value, constructing a fitness function based on the prediction precision and the mutual information mean value, and obtaining an optimal solution based on the fitness function; and obtaining a target rainfall station network layout result based on the optimal solution. The method solves the problems that an existing rainfall station network layout method is insufficient in spatial-temporal feature capture capability, is easy to fall into a local optimal solution due to dependence on a certain heuristic algorithm, and is low in optimization efficiency and cannot effectively balance and collaboratively optimize multiple targets.

Description

Technical Field

[0001] This invention relates to the field of hydrological station network layout technology, specifically, to a method for laying out a rainfall station network. Background Technology

[0002] In the field of hydrology, measured precipitation data is an indispensable foundation for hydrological research, water resources planning, management, and development. As a key infrastructure for acquiring this data, the scientific and rational design of the rain gauge network directly determines the quality and efficiency of data utilization, and is crucial for improving regional hydrological monitoring capabilities.

[0003] Currently, most rain gauge network layouts integrate ground-based measured rainfall data with remote sensing precipitation products to guide the design of the network, aiming to achieve complementary advantages in accuracy and coverage between the two data sources. However, existing fusion methods still have significant limitations: simply using ground station data fails to effectively utilize the continuous spatial information provided by satellite remote sensing, resulting in insufficient ability to capture the spatiotemporal characteristics of precipitation; and focusing only on a single objective, such as monitoring accuracy, leads to poor adaptability of the optimized network layout. Summary of the Invention

[0004] To address the shortcomings of existing rain gauge network layout methods, such as insufficient ability to capture spatiotemporal features and inability to effectively balance and coordinate the optimization of multiple objectives, this invention provides a rain gauge network layout method, which includes:

[0005] Acquire precipitation data from a historical rainfall station network within the target area, wherein the historical rainfall station network includes several rainfall stations;

[0006] Extract the spatiotemporal features of the precipitation data, and obtain the rainfall prediction value for each spatial unit based on the spatiotemporal features;

[0007] Based on the rainfall forecast values, the prediction accuracy and the mean mutual information are obtained. Based on the prediction accuracy and the mean mutual information, a fitness function is constructed. Based on the fitness function, the optimal solution is obtained.

[0008] The target rainfall station network layout result is obtained based on the optimal solution.

[0009] This method captures the spatiotemporal characteristics of precipitation data through a spatiotemporal deep learning model. At the same time, this method establishes a three-dimensional optimization objective function, namely the fitness function, which includes prediction accuracy, mean mutual information, and number of stations. This achieves a dynamic balance among the three objectives, reducing the number of stations, lowering costs, increasing data utilization, reducing the limitations of station network layout, and improving the adaptability of station network layout.

[0010] Furthermore, the specific steps for obtaining the precipitation data include:

[0011] Acquire historical rainfall data and historical remote sensing precipitation products from the historical rainfall station network, and match the historical rainfall data and historical remote sensing precipitation products based on a preset spatiotemporal benchmark;

[0012] Based on a preset scale, obtain the spatiotemporal data blocks of the remote sensing images of the historical remote sensing precipitation products, and obtain a multi-temporal feature matrix based on the spatiotemporal data blocks of the remote sensing images.

[0013] The historical rainfall data is discretized to obtain the first data.

[0014] The precipitation data is obtained based on the multi-time series feature matrix and the first data.

[0015] Furthermore, the spatiotemporal features of the precipitation data are extracted based on a pre-trained spatiotemporal depth model. According to the input order, the spatiotemporal depth model includes a feature extraction layer, a bidirectional network layer, and a fully connected layer.

[0016] The feature extraction layer includes several convolutional modules arranged in parallel, and each convolutional module includes several convolutional kernels and a batch normalization layer.

[0017] Furthermore, the specific steps for obtaining the rainfall forecast value include:

[0018] The feature extraction layer extracts the spatial texture features of the multi-temporal feature matrix and the temporal features of the first data;

[0019] The bidirectional network layer obtains the attention weights of the temporal features, and the spatial texture features and the temporal features are weighted and fused based on the attention weights to obtain the spatiotemporal features;

[0020] The fully connected layer obtains the rainfall prediction value based on the spatiotemporal characteristics.

[0021] Furthermore, the first formula for calculating the fitness function is:

[0022] ;

[0023] in, This represents the function value of the fitness function. , and All represent weighting coefficients. Indicates the accuracy of the prediction. This represents the mean mutual information between any two rainfall stations. Indicates the number of selected sites. This indicates the total number of rain gauge stations in the rain gauge network.

[0024] A three-dimensional optimization objective function is established, which includes prediction accuracy, information redundancy (mean of mutual information) and number of sites. This reduces the number of sites, lowers costs, and increases data utilization, breaking through the limitations of traditional approaches that only pursue fewer sites or only pursue high accuracy. The dynamic balance of the three objectives is achieved through weight adjustment.

[0025] Furthermore, the second formula for obtaining the mean mutual information is:

[0026] ;

[0027] ;

[0028] in, This represents the mutual information between any two rainfall stations. Indicates the first The first rain gauge station and the first The average mutual information between rainfall stations This represents the joint probability distribution of precipitation values ​​between any two rain gauge stations. Represents element Marginal probability distribution, Represents element Marginal probability distribution, Indicates from The number of combinations of choosing 2 elements from 3 distinct elements. and Both represent integers greater than or equal to 1. and Each represents the selected set of rainfall stations. and Each element represents a selected set of rainfall stations.

[0029] Introducing mutual information (MI) to replace traditional spatial distance as a measure of redundancy better aligns with the inherent correlation of meteorological data. This is because the core of redundancy lies in information overlap, not physical distance. Traditional spatial distance methods only consider the proximity of stations, ignoring the potential climate correlations between distant stations (such as the influence of the same front) and the significant environmental differences between nearby stations (such as mountainous areas versus plains). MI, based on information theory, does not rely on distribution assumptions and directly quantifies the degree of information sharing between data from two stations. Furthermore, the correlations between meteorological elements such as temperature and precipitation are often non-linear (e.g., implicit coupling between precipitation and air pressure, and lagged correlations in temperature). Compared to traditional distance methods, MI can simultaneously identify both linear and non-linear correlations, better reflecting the complex characteristics of meteorological data and better adapting to its non-linear relationships.

[0030] Furthermore, the specific steps for obtaining the optimal solution based on the rainfall forecast and the fitness function include:

[0031] A1. The rainfall stations are binary encoded to obtain the first station, and several of the first stations are randomly selected to obtain the initial population.

[0032] A2. Obtain the first fitness of the initial population based on the fitness function;

[0033] A3. Based on the preset tournament scale, randomly select several first sites, and obtain the parent population based on the first fitness and the first site.

[0034] A4. Perform crossover and mutation on the parent population in sequence to obtain a new population;

[0035] A5. Determine whether the number of iterations is greater than the preset number. If yes, the iteration ends and the optimal solution is obtained based on the new population. Otherwise, update the initial population to the new population, delete the new population, and return to A2.

[0036] The selection operation is changed to a tournament selection strategy, and the tournament size is set to enhance population diversity, reduce premature convergence, and improve robustness.

[0037] Furthermore, the specific steps for obtaining the target rainfall station network layout include:

[0038] S1. Based on the optimal solution, obtain an initial particle swarm, initialize the initial parameters of the initial particle swarm, randomly initialize the first position and first velocity of each particle in the initial particle swarm, and obtain the individual historical best position, the group historical best position, the individual historical best fitness value, and the group historical best fitness value based on the first position and the first velocity.

[0039] S2. Determine whether the termination condition is met. If so, obtain the target rainfall station network layout result based on the individual's historical best position and the group's historical best position.

[0040] S3. Otherwise, update the first position and the first velocity of each particle;

[0041] S4. Obtain the second fitness of each particle;

[0042] S5. Update the individual's historical best position and the individual's historical best fitness value based on the second fitness and the individual's historical best fitness value;

[0043] S6. Based on the updated individual historical best fitness value and the updated individual historical best position, update the group's historical best position and the group's historical best fitness value, and return to S2.

[0044] Considering that relying on a single heuristic algorithm can easily lead to local optima and is often accompanied by high computational complexity and low optimization efficiency, this method adopts a hybrid optimization algorithm of genetic optimization algorithm and particle swarm optimization algorithm to jointly optimize the layout of rain gauge network. This can balance global exploration and local development capabilities, reduce the algorithm from getting stuck in local optima, and improve optimization efficiency.

[0045] By employing the particle swarm optimization algorithm, local optima are reduced, dimensionality reduction improves efficiency, and the computation time for large-scale layouts is shortened.

[0046] Furthermore, the third calculation formula for updating the first position is as follows:

[0047] ;

[0048] in, Represents particles In the The speed of each iteration Represents particles In the The speed of each iteration Indicates inertia weight, and All represent learning factors. and Both represent random numbers within the interval [0,1]. Represents particles The individual optimal solution. This represents the optimal solution for the group. Represents particles In the The position of the next iteration. and Both represent integers greater than or equal to 1;

[0049] The fourth calculation formula for the first velocity is updated as follows:

[0050] ;

[0051] in, Represents particles In the The position of the next iteration. This represents a function used to generate binomial random numbers.

[0052] Furthermore, the fifth formula for calculating the attention weights is as follows:

[0053] ;

[0054] in, Indicates the first Attention weights at each time step and They represent and The weight matrix, Indicates a hidden state. Indicates cell state, Indicates the bias term. Indicates the total time step. Represents an integer greater than or equal to 1. This represents the activation function.

[0055] One or more technical solutions provided by this invention have at least the following technical effects or advantages:

[0056] 1. This method integrates remote sensing precipitation products and ground-measured rainfall data, captures spatiotemporal features through a spatiotemporal deep learning model, and employs a hybrid optimization algorithm combining genetic optimization and particle swarm optimization to collaboratively optimize the layout of rain gauge networks. This balances global exploration and local development capabilities, reduces the algorithm's tendency to get stuck in local optima, and improves optimization efficiency. Simultaneously, a three-dimensional optimization objective function, i.e., a fitness function, is established, encompassing prediction accuracy, mean mutual information, and the number of stations. This achieves a dynamic balance among the three objectives, reducing the number of stations, lowering costs, increasing data utilization, mitigating the limitations of the station network layout, and improving its adaptability.

[0057] 2. Establish a three-dimensional optimization objective function that includes forecast accuracy, information redundancy (mean mutual information), and number of stations. This reduces the number of stations, lowers costs, and increases data utilization, breaking through the limitations of traditional approaches that only pursue fewer stations or high accuracy. The dynamic balance of the three objectives is achieved through weight adjustment. Mutual information (MI) is introduced to replace the traditional spatial distance measure of redundancy, which is more in line with the correlation nature of meteorological data. Attached Figure Description

[0058] The accompanying drawings, which are provided to further illustrate embodiments of the invention and constitute a part of this invention, are not intended to limit the scope of the invention.

[0059] Figure 1 This is a flowchart illustrating a rain gauge network layout method according to the present invention;

[0060] Figure 2 This is a schematic diagram of the spatiotemporal depth model;

[0061] Figure 3 This is a schematic diagram of the original rain gauge network.

[0062] Figure 4 This is a schematic diagram of the information entropy value of the target area;

[0063] Figure 5This is a schematic diagram of the optimized rain gauge network distribution. Detailed Implementation

[0064] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other.

[0065] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0066] Example 1

[0067] refer to Figures 1-5 This embodiment provides a method for rain gauge network layout, the method including:

[0068] Acquire precipitation data from a historical rainfall station network within the target area, wherein the historical rainfall station network includes several rainfall stations;

[0069] The specific steps for obtaining the precipitation data include:

[0070] Obtain historical rainfall data and historical remote sensing precipitation products from the historical rainfall station network, and match the historical rainfall data and historical remote sensing precipitation products based on a preset spatiotemporal reference; Figures 3-4 As shown, based on the actual situation and the distribution map of the original rain gauge network, daily remote sensing precipitation products and daily measured rainfall data of the study area are selected in terms of time and space. The satellite remote sensing precipitation products and the measured rainfall data of the ground stations are unified to a consistent spatiotemporal reference to verify the coordinate system matching and ensure accurate spatial correspondence.

[0071] Based on a preset scale, obtain the spatiotemporal data blocks of the remote sensing images of the historical remote sensing precipitation products, and obtain a multi-temporal feature matrix based on the spatiotemporal data blocks of the remote sensing images; for example, taking the ground rain gauge station as the core, extract the spatiotemporal data blocks of the remote sensing images according to a preset scale, and use the edge filling strategy to handle the boundary effect to form a multi-temporal feature matrix.

[0072] The historical rainfall data is discretized to obtain the first data; if the measured precipitation data on the ground is normalized, it is simultaneously discretized to provide input data with uniform scale and suitable form for subsequent mutual information index calculation.

[0073] The precipitation data is obtained based on the multi-time series feature matrix and the first data.

[0074] Extract the spatiotemporal features of the precipitation data, and obtain the rainfall prediction value for each spatial unit based on the spatiotemporal features;

[0075] Spatial units can be obtained by dividing the target area into grid cells based on rainfall stations.

[0076] The spatiotemporal features of the precipitation data are extracted based on a pre-trained spatiotemporal depth model (such as an existing deep learning model). According to the input order, the spatiotemporal depth model includes a feature extraction layer, a bidirectional network layer, and a fully connected layer. The feature extraction layer includes several convolutional modules arranged in parallel, and each convolutional module includes several convolutional kernels and a batch normalization layer.

[0077] For example, the feature extraction layer uses 8 parallel TimeDistributed convolutional modules, each containing 32 3×3 convolutional kernels, and a batch normalization layer to reduce gradient vanishing. Different channels are used to extract multi-scale spatial texture features and the temporal features of precipitation sequences. The bidirectional network layer (such as Bidirectional Long Short-Term Memory Network, BiLSTM) uses an attention mechanism to weight and fuse the temporal features of precipitation sequences. The fully connected layer outputs the precipitation prediction value.

[0078] In this embodiment, during the training of the spatiotemporal depth model, five-fold time series cross-validation (TSCV) can also be used, with the training set and validation set divided in an 8:2 ratio. An early stopping mechanism is set (the training stops when the root mean square error of the validation set increases by more than 0.5% for five consecutive rounds). After the fully connected layer of the spatiotemporal depth model, a loss function and an optimizer can also be included. The loss function uses the mean square error (MSE), and the optimizer uses the AdamW algorithm with weight decay. The convergence stability is improved by using a cosine annealing learning rate scheduling strategy.

[0079] ;

[0080] in, Indicates the first Learning rate of the round, This represents the initial learning rate. Indicates the maximum number of rounds. Represents an integer greater than or equal to 1.

[0081] The specific steps for obtaining the rainfall forecast value include:

[0082] The feature extraction layer extracts the spatial texture features of the multi-temporal feature matrix and the temporal features of the first data;

[0083] The bidirectional network layer obtains the attention weights of the temporal features, and the spatial texture features and the temporal features are weighted and fused based on the attention weights to obtain the spatiotemporal features;

[0084] The fully connected layer obtains the rainfall prediction value based on the spatiotemporal characteristics.

[0085] Based on the rainfall forecast values, the forecast accuracy (such as the mean square error between the forecast value and the actual value) and the mean mutual information are obtained. Based on the forecast accuracy and the mean mutual information, a fitness function is constructed. Based on the fitness function, the optimal solution is obtained.

[0086] Prediction accuracy: The prediction error of site monitoring data (the smaller the better), which reflects the reliability of the site data.

[0087] Information redundancy (mutual information mean): the degree of data duplication between stations (the lower the better), reflecting the information efficiency of the rainfall monitoring network and spatial representativeness.

[0088] Number of sites: The total number of sites to be retained (the fewer the better), reflecting the economics of optimization and controlling the cost of the number of sites.

[0089] The first formula for calculating the fitness function is:

[0090] ;

[0091] in, This represents the function value of the fitness function. , and All represent weighting coefficients. Indicates the accuracy of the prediction. This represents the mean mutual information between any two rainfall stations. Indicates the number of selected sites. This indicates the total number of rain gauge stations in the rain gauge network.

[0092] If the regional budget is tight, it can be increased. (Prioritize reducing the number of sites); if extremely high data accuracy is required, the number of sites can be increased. (Prioritizing low error) to achieve a dynamic balance of the three objectives through weight adjustment.

[0093] The second formula for obtaining the average mutual information is as follows:

[0094] ;

[0095] ;

[0096] in, This represents the mutual information between any two rainfall stations. Indicates the first The first rain gauge station and the first The average mutual information between rainfall stations This represents the joint probability distribution of precipitation values ​​between any two rain gauge stations. Represents element Marginal probability distribution, Represents element Marginal probability distribution, Indicates from The number of combinations of choosing 2 elements from 3 distinct elements. and Both represent integers greater than or equal to 1. and Each represents the selected set of rainfall stations. and Each element represents a selected set of rainfall stations.

[0097] The fifth formula for obtaining the attention weight is:

[0098] ;

[0099] in, Indicates the first Attention weights at each time step and They represent and The weight matrix, Indicates a hidden state. Indicates cell state, Indicates the bias term. Indicates the total time step. Represents an integer greater than or equal to 1. This represents the activation function.

[0100] The specific steps for obtaining the optimal solution based on the rainfall prediction value and the fitness function include:

[0101] A1. The rainfall stations are binary encoded to obtain the first station. Several of the first stations are randomly selected to obtain the initial population. If the individual stations are encoded using binary vectors, the length of the vector is equal to the total number of stations, where 1 indicates that the corresponding station is selected and 0 indicates that it is not selected. The probability of each station being selected is 0.6 during initialization, which can accelerate the convergence process of the algorithm.

[0102] A2. Obtain the first fitness of the initial population based on the fitness function;

[0103] A3. Based on the preset tournament size, randomly select several first stations, and obtain the parent population based on the first fitness and the first station; if the selection operation in the existing genetic algorithm is changed to a tournament selection strategy, and the tournament size is set to 3, the diversity of the population can be enhanced, the problem of premature convergence can be reduced, and the robustness can be improved.

[0104] A4. Perform crossover and mutation on the parent population in sequence to obtain a new population; if a uniform crossover operator (cxUniform) is used, the crossover probability is set to 0.7, and a lower mutation probability is used to balance the global exploration and local development capabilities and reduce the problem of the algorithm getting stuck in local optima.

[0105] A5. Determine whether the number of iterations is greater than the preset number. If yes, the iteration ends and the optimal solution is obtained based on the new population. Otherwise, update the initial population to the new population, delete the new population, and return to A2.

[0106] The target rainfall station network layout result is obtained based on the optimal solution.

[0107] The specific steps for obtaining the target rainfall station network layout include:

[0108] Based on the existing Particle Swarm Optimization (PSO) algorithm:

[0109] S1. Obtain an initial particle swarm based on the optimal solution, initialize the initial parameters of the initial particle swarm, randomly initialize the first position and first velocity of each particle in the initial particle swarm, and obtain the individual's historical best position, the group's historical best position, the individual's historical best fitness value, and the group's historical best fitness value based on the first position and the first velocity; for example, the first 5 elite solutions in the optimal solution output by the genetic algorithm are used as the initial particle swarm of the particle swarm optimization algorithm.

[0110] S2. Determine if the termination condition is met (e.g., the number of iterations has been reached). If so, obtain the target rainfall station network layout result based on the individual's historical best position and the group's historical best position. Figure 5 As shown, the optimized rain gauge network layout scheme performs well in terms of rainfall spatial characteristics and total information entropy. The station distribution is objective and reasonable, effectively ensuring the information content and accuracy of the measured rainfall data obtained by the rain gauge network. It can also effectively retain the ability of the rain gauge network to capture spatial characteristics. Based on the individual historical best position and the group historical best position, the corresponding stations are obtained, the optimized vector data is generated, and the optimized rain gauge network is saved.

[0111] S3. Otherwise, update the first position and the first velocity of each particle;

[0112] S4. Obtain the second fitness of each particle;

[0113] S5. Update the individual's historical best position and the individual's historical best fitness value based on the second fitness and the individual's historical best fitness value;

[0114] S6. Based on the updated individual historical best fitness value and the updated individual historical best position, update the group's historical best position and the group's historical best fitness value, and return to S2.

[0115] The third calculation formula for updating the first position is:

[0116] ;

[0117] in, Represents particles In the The speed of each iteration Represents particles In the The speed of each iteration Indicates inertia weight, and All represent learning factors. and Both represent random numbers within the interval [0,1]. Represents particles The individual optimal solution. This represents the optimal solution for the group. Represents particles In the The position of the next iteration. and Both represent integers greater than or equal to 1;

[0118] The fourth calculation formula for the first velocity is updated as follows:

[0119] ;

[0120] in, Represents particles In the The position of the next iteration. This represents a function used to generate binomial random numbers.

[0121] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0122] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for layouting a rain gauge network, characterized in that, The method includes: Acquire precipitation data from a historical rainfall station network within the target area, wherein the historical rainfall station network includes several rainfall stations; Extract the spatiotemporal features of the precipitation data, and obtain the rainfall prediction value for each spatial unit based on the spatiotemporal features; Based on the rainfall forecast values, the prediction accuracy and the mean mutual information are obtained. Based on the prediction accuracy and the mean mutual information, a fitness function is constructed. Based on the fitness function, the optimal solution is obtained. The target rainfall station network layout result is obtained based on the optimal solution.

2. The rain gauge network layout method according to claim 1, characterized in that, The specific steps for obtaining the precipitation data include: Acquire historical rainfall data and historical remote sensing precipitation products from the historical rainfall station network, and match the historical rainfall data and historical remote sensing precipitation products based on a preset spatiotemporal benchmark; Based on a preset scale, obtain the spatiotemporal data blocks of the remote sensing images of the historical remote sensing precipitation products, and obtain a multi-temporal feature matrix based on the spatiotemporal data blocks of the remote sensing images. The historical rainfall data is discretized to obtain the first data. The precipitation data is obtained based on the multi-time series feature matrix and the first data.

3. The rain gauge network layout method according to claim 2, characterized in that, The spatiotemporal features of the precipitation data are extracted based on a pre-trained spatiotemporal depth model. According to the input order, the spatiotemporal depth model includes a feature extraction layer, a bidirectional network layer, and a fully connected layer. The feature extraction layer includes several convolutional modules arranged in parallel, and each convolutional module includes several convolutional kernels and a batch normalization layer.

4. The rain gauge network layout method according to claim 3, characterized in that, The specific steps for obtaining the rainfall forecast value include: The feature extraction layer extracts the spatial texture features of the multi-temporal feature matrix and the temporal features of the first data; The bidirectional network layer obtains the attention weights of the temporal features, and the spatial texture features and the temporal features are weighted and fused based on the attention weights to obtain the spatiotemporal features; The fully connected layer obtains the rainfall prediction value based on the spatiotemporal characteristics.

5. The rain gauge network layout method according to claim 1, characterized in that, The first formula for calculating the fitness function is: ； in, This represents the function value of the fitness function. , and All represent weighting coefficients. Indicates the accuracy of the prediction. This represents the mean mutual information between any two rainfall stations. Indicates the number of selected sites. This indicates the total number of rain gauge stations in the rain gauge network.

6. The rain gauge network layout method according to claim 5, characterized in that, The second formula for obtaining the mean mutual information is: ； ； in, This represents the mutual information between any two rainfall stations. Indicates the first The first rain gauge station and the first The average mutual information between rainfall stations This represents the joint probability distribution of precipitation values ​​between any two rain gauge stations. Represents element Marginal probability distribution, Represents element Marginal probability distribution, Indicates from The number of combinations of choosing 2 elements from 3 distinct elements. and Both represent integers greater than or equal to 1. and Each represents the selected set of rainfall stations. and Each element represents a selected set of rainfall stations.

7. A method for layout of a rain gauge network according to claim 4, characterized in that, The specific steps for obtaining the optimal solution based on the rainfall forecast and the fitness function include: A1. The rainfall stations are binary encoded to obtain the first station, and several of the first stations are randomly selected to obtain the initial population. A2. Obtain the first fitness of the initial population based on the fitness function; A3. Based on the preset tournament scale, randomly select several first sites, and obtain the parent population based on the first fitness and the first site. A4. Perform crossover and mutation on the parent population in sequence to obtain a new population; A5. Determine whether the number of iterations is greater than the preset number. If yes, the iteration ends and the optimal solution is obtained based on the new population. Otherwise, update the initial population to the new population, delete the new population, and return to A2.

8. A method for layout of a rain gauge network according to claim 4, characterized in that, The specific steps to obtain the target rainfall station network layout results include: S1. Based on the optimal solution, obtain an initial particle swarm, initialize the initial parameters of the initial particle swarm, randomly initialize the first position and first velocity of each particle in the initial particle swarm, and obtain the individual historical best position, the group historical best position, the individual historical best fitness value, and the group historical best fitness value based on the first position and the first velocity. S2. Determine whether the termination condition is met. If so, obtain the target rainfall station network layout result based on the individual's historical best position and the group's historical best position. S3. Otherwise, update the first position and the first velocity of each particle; S4. Obtain the second fitness of each particle; S5. Update the individual's historical best position and the individual's historical best fitness value based on the second fitness and the individual's historical best fitness value; S6. Based on the updated individual historical best fitness value and the updated individual historical best position, update the group's historical best position and the group's historical best fitness value, and return to S2.

9. A method for layout of a rain gauge network according to claim 8, characterized in that, The third calculation formula for the first position is updated as follows: ； in, Represents particles In the The speed of each iteration Represents particles In the The speed of each iteration Indicates inertia weight, and All represent learning factors. and Both represent random numbers within the interval [0,1]. Represents particles The individual optimal solution. This represents the optimal solution for the group. Represents particles In the The position of the next iteration. and Both represent integers greater than or equal to 1; The fourth calculation formula for the first velocity is updated as follows: ； in, Represents particles In the The position of the next iteration. This represents a function used to generate binomial random numbers.

10. A method for layout of a rain gauge network according to claim 4, characterized in that, The fifth formula for obtaining the attention weights is: ； in, Indicates the first Attention weights at each time step and They represent and The weight matrix, Indicates a hidden state. Indicates cell state, Indicates the bias term. Indicates the total time step. Represents an integer greater than or equal to 1. This represents the activation function.

Images