Method for optimizing the placement of a mountain rainfall observation network based on radar rainfall data.

Abstract

The present invention discloses a method for optimizing the arrangement of a mountainous rainfall observation network based on radar rainfall data. The method includes step 1 of collecting rainfall data and meshing a target area, step 2 of identifying anomalous observation points at rainfall observation stations, step 3 of determining candidate locations for additional rainfall observation stations, step 4 of evaluating the accuracy and effectiveness of different rainfall observation network optimization means, and step 5 of determining the optimal rainfall observation network arrangement means. [Effects] The method of the present invention, when combined with radar rainfall data, enables effective planning and optimization of rainfall observation networks in complex topographic conditions of mountainous areas, further improves the accuracy and reliability of precipitation observations in mountainous watersheds, and provides more accurate data support for the management and prevention of flood disasters in mountainous areas.

Description

[Technical Field]

[0001] This invention belongs to the field of hydrological meteorological observation technology, and more particularly to a method for optimizing the arrangement of a mountainous rainfall observation network based on radar rainfall data. [Background technology]

[0002] As the effects of global climate change become increasingly pronounced, extreme weather events are becoming more frequent, and flood disasters, particularly in mountainous river basins, pose a serious threat to the safety of people's lives and property. Accurate and reliable rainfall observation is crucial for early flood warnings. Conventional rainfall observation station systems rely on empirical judgment and limited meteorological observation data, and do not adequately consider the impact of complex mountainous topography on rainfall distribution. Faced with the strong spatial heterogeneity and spatiotemporal variability of rainfall in mountainous areas, blind spots are likely to occur in storm observation, making it difficult to accurately capture localized heavy rainfall and storm events, thereby affecting the accuracy of flood forecasts and early disaster warnings.

[0003] Currently, optimization methods for existing rainfall observation networks often improve network coverage and accuracy by selecting optimal locations based on certain optimization algorithms, such as genetic algorithms or particle swarm optimization (PSO). However, conventional optimization methods often rely on data from ground-based rainfall observation stations and ignore the potential of remote sensing technologies (e.g., radar rainfall). Radar data has relatively high spatiotemporal resolution and can provide more comprehensive precipitation observation information, but conventional methods cannot fully utilize this information. Therefore, while conventional optimization methods can improve the performance of observation networks under certain conditions, they still face challenges in accuracy and efficiency in complex mountainous environments and cannot achieve ideal observation effects. Consequently, how to optimize conventional rainfall observation networks in complex mountainous terrain by combining them with high-precision remote sensing data is a technical problem that urgently needs to be solved. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] (Nothing in particular) [Overview of the project] [Problems that the invention aims to solve]

[0005] The object of the present invention is to provide a method for optimizing the arrangement of a mountainous rainfall observation network based on radar rainfall data in order to solve the above technical problems. [Means for solving the problem]

[0006] To achieve the above objective, the present invention provides the following technical solutions. The present invention discloses a method for optimizing the arrangement of a mountainous rainfall observation network based on radar rainfall data, the method comprising (steps 1 to 5 below, i.e.,) Step 1: Collection of rainfall data and mesh division of the target area: Actual rainfall data P from all M rainfall observation stations within the target area. 雨量観測所 and P 雨量観測所 Time-series corresponding radar rainfall data P レーダー Step 1 involves collecting data, performing the same mesh division on the target area based on the mesh distribution of radar rainfall data, recording a total of J meshes, and assigning a unique identifier to each mesh. Step 2, Identification of Anomaly Observation Points at Rain Gauge Stations: Using a GCN model, which is a graph convolutional network model, we identify anomaly observation points at rain gauge stations. First, we construct node features for the rain gauge stations. If the actual rainfall data for each rain gauge station is for N hours, we train the GCN model using the node features of M rain gauge stations and the corresponding actual rainfall data for N hours as the training dataset. The node features of the rain gauge stations are used as model inputs, and the corresponding actual rainfall data are used as model outputs. Subsequently, we use the trained GCN model to identify anomaly observation points at the rain gauge stations. The identified anomaly observation points are removed, and the number of rain gauge stations operating normally is recorded as K. Step 3: Determination of additional candidate positions for rain gauges: Spatial interpolation is performed on the actual rainfall data of K rain gauges every hour by using the Kriging interpolation method to obtain the rainfall field data P of the target watershed area 観測点補間 After the rainfall field data P is generated, for the J meshes divided in step 1, each mesh center point has two sets of rainfall data, which are respectively radar rainfall data P レーダー and the interpolated rainfall field data P 観測点補間 , wherein j=1, 2, ..., J. Calculate the relative error RE of P レーダー and P 観測点補間 at each mesh center every hour j , count the number N of meshes satisfying RE j > D2 in each time period D2,t , set the time periods satisfying N D2,t > D3 as belonging to two types of time periods with significant difference in rainfall distribution, denote the total number of such time periods as T, wherein both thresholds D2 and D3 are determined based on the natural geographical conditions and data status of the watershed, count the average value of the relative error RE j of each mesh within T time periods, sort the average values in descending order, and use the first γ mesh center points as additional candidate positions, which are numbered 1, 2, ..., γ respectively Step 4: Accuracy effect evaluation of optimization means for different rain gauge networks: Set that 1 to γ rain gauges are to be supplemented, and a total of γ types of adding means are generated, wherein the γ types of adding means are respectively adding at the candidate position numbered 1, adding at the candidate positions numbered 1 and 2, and so on, adding at the candidate positions numbered 1, 2, ..., γ, thus the total number of rain gauges corresponding to the γ types of adding means are K+1, K+2, ..., K+γ respectively Rainfall conditions of different adding means are predicted using a machine learning model. For the position s of each rainfall station, a feature vector X(s) is constructed, which includes the geographical position of the observation point, topographic features, meteorological features, and radar rainfall data corresponding to the mesh where the observation point is located. Then, the machine learning model f is trained by taking the feature vectors and actual rainfall data of all K rainfall stations as the training data set, where the feature vector X of a rainfall station is used as the model input, and the corresponding actual rainfall data is used as the model output. Subsequently, using the trained model, for each candidate adding position s*, its feature vector X(s * *), the rainfall condition at the position [Formula] is predicted as follows: [Formula] for all T time periods, predict the rainfall data of each candidate adding position in γ types of adding means for each time period, perform spatial interpolation on the prediction results of each adding means using kriging interpolation, and generate corresponding rainfall field data; for all T time periods, calculate the number of RE j D2 meshes again for each time period, [Formula] denoted as , that is, the number of RE j D2 meshes in the t-th time period, where t=1, 2, …, T, and then calculate the average percentage reduction per in the number of meshes for all time periods, [Formula] step 4, where a larger per indicates that after adding stations, the watershed rainfall field data is closer to the radar rainfall data; Step 5, Determining the optimal rainfall observation network configuration: Considering the construction costs, maintenance costs, and potential socioeconomic effects of rainfall observation stations, a penalty function ω is constructed to show the unfavorable impact of actual conditions on the construction of rainfall observation stations, where C is the construction cost, W is the maintenance cost, and B is the socioeconomic effect. C, W, and B are all functions of additional measures and change with changes in measures. Thus, the penalty function ω is defined as follows: ω = αC + βW - φB (5) In the formula, α, β, and φ are weighting coefficients. By combining per and ω, we construct an optimization target function φ, make it positively correlated with per and negatively correlated with ω, that is, φ = μ1 × per - μ2 × ω (6) In the formula, μ1 and μ2 are weighting coefficients. Step 5 involves calculating the corresponding φ value for each type of additional means and selecting the means with the highest φ value as the optimal rainfall observation station placement means. Includes.

[0007] Furthermore, performing the same mesh division on the target area based on the mesh distribution of radar rainfall data in Step 1 specifically involves analyzing the radar rainfall data file, extracting its spatial resolution, determining the basic unit of the mesh, and then systematically numbering the mesh using matrix coordinates based on the spatial location of the radar data, with row numbers increasing from north to south and column numbers increasing from west to east.

[0008] Furthermore, in step 2, constructing the node features of rainfall observation stations specifically involves converting the geographical location relationships between rainfall observation stations into a graph structure, where each rainfall observation station is considered a single node in the graph structure, the edges between nodes represent the spatial correlation between rainfall observation stations, and the weights of the edges are set by a distance metric or a similarity metric based on geographical distance. The node features of a rainfall observation station include rainfall time series features and geographic information features. The rainfall time series features include rainfall statistics and temporal feature information for historical time periods, and the geographic information features include the latitude, longitude, elevation, and topographic features of each rainfall observation station.

[0009] Furthermore, in step 2, identifying anomalous observation points at rainfall observation stations using the trained GCN model specifically involves calculating the reconstruction error RE predicted by the trained GCN model for all N time periods for each rainfall observation station.

number

number

[0010] Furthermore, in step 4, the topographic features include elevation, gradient, and gradient direction, and the meteorological features include temperature, humidity, and wind speed. [Effects of the Invention]

[0011] The beneficial effects of the present invention are as follows: The method of the present invention, when combined with radar rainfall data, enables effective planning and optimization of rainfall observation networks in complex topographic conditions in mountainous areas, further improves the accuracy and reliability of precipitation observations in mountainous watersheds, and provides more accurate data support for the management and prevention of flood disasters in mountainous areas.

[0012] The present invention will be described in further detail below with reference to the drawings and specific embodiments. [Brief explanation of the drawing]

[0013] [Figure 1] This is a flowchart of the method described in the present invention. [Modes for carrying out the invention]

[0014] The present invention discloses a method for optimizing the arrangement of a mountainous rainfall observation network based on radar rainfall data, and as shown in Figure 1, the method includes the following steps.

[0015] Step 1: Collection of rainfall data and mesh division of the target area.

[0016] For a target area where the rainfall observation network layout should be optimized, first, we collect actual rainfall data P from all M rainfall observation stations within that area. 雨量観測所 and P 雨量観測所 Time-series corresponding radar rainfall data P レーダー We collect data and prioritize selecting rainfall during flood season. Next, we perform the same mesh division on the target area based on the mesh distribution of radar rainfall data (denoted as J meshes in total), and assign a unique identifier to each mesh. Specifically, we analyze the radar rainfall data file, extract its spatial resolution, determine the basic unit of the mesh, and then, based on the spatial location of the radar data, we systematically number the meshes using a matrix coordinate system, with row numbers increasing from north to south and column numbers increasing from west to east, ensuring that each mesh has a unique identifier, thereby facilitating subsequent data processing and analysis.

[0017] Step 2: Identify abnormal observation points at rainfall observation stations.

[0018] Considering the possibility of operational anomalies at conventional rainfall observation stations, a GCN (Graph Convolutional Network) model is used to identify anomaly observation points and ensure that the conventional observation network consists of observation points that are operating normally. First, node features of rainfall observation stations are constructed. If the rainfall data for each rainfall observation station is for N hours, the node features of M rainfall observation stations and the corresponding actual rainfall data for N hours are used as the training dataset to train the GCN model. The node features of the rainfall observation stations are used as the model input, and the corresponding actual rainfall data is used as the model output. Next, the trained GCN model is used to identify anomaly observation points at the rainfall observation stations. The identified anomaly observation points are removed, and the number of rainfall observation stations that are operating normally is recorded as K. Subsequent steps are performed based only on the K rainfall observation stations, without considering the anomaly observation points.

[0019] The specific process is as follows: 1) Construct node features for rainfall observation stations. Convert the geographical location of rainfall observation stations into a graph structure, where each station is considered a single node in the graph structure. The edges between nodes represent the spatial correlation between rainfall observation stations, and the weights of the edges are set by a distance metric (e.g., Euclidean distance) or a similarity metric based on geographical distance. The node features of rainfall observation stations include rainfall time series features and geographic information features. Rainfall time series features include information such as rainfall statistics for historical time periods (e.g., average and variance of precipitation over the past few hours or days), and time features (year, month, quarter, time of day, etc., to which the current time period is located). Geographic information features include the latitude and longitude, elevation, and topographic features of each rainfall observation station. 2) Initialize the model parameters and train the GCN model using the training dataset. Set the number of layers L in the model, the number of neurons d1 in each layer, and the activation function σ, and the lth (l=1, 2, ..., L) layer of the GCN model can be expressed as follows.

number

number

number

number

number

[0020] Step 3: Determine additional candidate locations for rain gauge stations.

[0021] Spatial interpolation is performed on the actual rainfall data from K rain observation stations every hour using the kriging interpolation method, and the rainfall field data P for the target watershed area is generated. 観測点補間 This generates the following. As a result, for the J meshes divided in step 1, each mesh center point now has two sets of rainfall data, and each has radar rainfall data P. レーダー and the interpolated rainfall field data P 観測点補間 This is the case. P at the center of each mesh every hour. レーダー and P 観測点補間 The relative error RE of the two types of rainfall j Calculate (j=1, 2, ..., J) and RE for each time period. j >Number of meshes in D2: ND2,t We statistically analyze N D2,t >The time period D3 is set to belong to two types of time periods where differences in rainfall distribution are significant, and the total number of such time periods is denoted as T, where the thresholds D2 and D3 are both determined based on the natural geographical conditions and data conditions of the watershed. Next, the relative error RE of each mesh within T time periods is calculated. j The average values ​​are statistically analyzed and arranged in descending order. The first γ mesh center points are designated as candidate positions, numbered 1, 2, ..., γ respectively.

[0022] Step 4: Evaluation of the accuracy and effectiveness of optimization methods for different rainfall observation networks.

[0023] The system is configured to add 1 to γ ​​rain gauge stations, generating a total of γ types of additional means. Each of the γ types of additional means is added at candidate position number 1, at candidate positions number 1 and 2, and so on, at candidate positions number 1 and 2, ..., γ. Thus, the total number of rain gauge stations corresponding to the γ types of additional means is K+1, K+2, ..., K+γ. We predict rainfall conditions for different additional methods using machine learning models (random forest, gradient lift tree, etc.). For each rainfall observation station's location s, we construct a feature vector X(s) that includes the geographical location of the observation point, topographic features (including elevation, gradient, gradient direction, etc.), meteorological features (including temperature, humidity, wind speed, etc.), and radar rainfall data corresponding to the mesh where it is located. Next, we train a machine learning model f using the feature vectors and actual rainfall data of all K rainfall observation stations as the training dataset, with the rainfall observation station feature vector X as the model input and the corresponding actual rainfall data as the model output. Then, using the trained model, we predict the feature vector X(s) for each additional candidate location s*. * Based on the rainfall conditions at that location

number

number

number

number

[0024] Step 5: Determining the optimal configuration of the rainfall observation network.

[0025] By combining this with actual economic considerations, we construct an optimization target function for the observation network layout to determine the optimal means of arranging the rainfall observation network. Considering the construction costs, maintenance costs, and potential socioeconomic effects of rainfall observation stations, we construct a penalty function ω to show the unfavorable impact of actual conditions on the construction of rainfall observation stations, where C is the construction cost, W is the maintenance cost, and B is the socioeconomic effect. C, W, and B are all functions of additional means and change with changes in means. Thus, the penalty function ω is defined as follows: ω = αC + βW - φB (5) In the formula, α, β, and φ are weighting coefficients. By combining per and ω, we construct an optimization target function φ, make it positively correlated with per and negatively correlated with ω, that is, φ = μ1 × per - μ2 × ω (6) In the formula, μ1 and μ2 are weighting coefficients, assigned according to their importance. For each type of additional measure, the corresponding φ value is calculated, and the measure with the highest φ value is selected as the optimal rainfall observation station placement measure. This measure not only maximizes the accuracy of rainfall observation but also ensures the economic rationality of construction and maintenance, achieving overall optimization of technology and economy.

[0026] Finally, it should be noted that the above has been used to describe, but is not limited to, the technical solutions of the present invention, and has been described in detail with reference to preferred arrangements. As those skilled in the art will understand, the technical solutions of the present invention can be modified or replaced with equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for optimizing the placement of a mountainous rainfall observation network based on radar rainfall data, Step 1: Collection of rainfall data and mesh division of the target area: Actual rainfall data P from all M rain observation stations within the target area 雨量観測所 and P 雨量観測所 Time-series corresponding radar rainfall data P レーダー Step 1 involves collecting data, performing the same mesh division on the target area based on the mesh distribution of radar rainfall data, recording a total of J meshes, and assigning a unique identifier to each mesh. Step 2: Identifying abnormal observation points at rainfall observation stations: Step 2 involves using a graph convolutional network (GCN) model to identify anomalous observation points at rainfall observation stations. First, node features of the rainfall observation stations are constructed. If the actual rainfall data for each rainfall observation station is for N hours, the node features of M rainfall observation stations and the corresponding actual rainfall data for N hours are used as the training dataset to train the GCN model. The node features of the rainfall observation stations are used as model inputs, and the corresponding actual rainfall data are used as model outputs. Subsequently, the trained GCN model is used to identify anomalous observation points at the rainfall observation stations. The identified anomalous observation points are removed, and the number of rainfall observation stations operating normally is recorded as K. Step 3: Determining additional candidate locations for rain gauge stations: Performing spatial interpolation on the actual rainfall data of K rain gauge stations every hour by using the Kriging interpolation method to obtain rainfall field data P of the target watershed area 観測点補間 is generated, then for the J meshes divided in step 1, each mesh center point has two sets of rainfall data, which are respectively radar rainfall data P レーダー and the interpolated rainfall field data P 観測点補間 , calculating the relative error RE of P レーダー and P 観測点補間 at each mesh center every hour, where j = 1, 2, ..., J, counting the number of meshes N ｊ where RE ｊ > D ２ for each time period, setting time periods satisfying N Ｄ２，ｔ > D Ｄ２，ｔ belong to two types of time periods with significant differences in rainfall distribution, recording the total number of such time periods as T, wherein the threshold D ３ and D ２ are both determined based on the physiogeographical conditions and data status of the watershed, counting the average value of the relative error RE ３ of each mesh within the T time periods, arranging the average values in descending order, and taking the first γ mesh center points as additional candidate positions, with numbers 1, 2, ..., γ respectively: step 3, ｊ ​ Step 4: Evaluation of the accuracy and effectiveness of optimization methods for different rainfall observation networks: The system is configured to add 1 to γ ​​rain gauge stations, generating a total of γ types of additional means. Each of these γ types of additional means is added at candidate position number 1, at candidate positions number 1 and 2, and so on, at candidate positions number 1 and 2, ..., γ. Thus, the total number of rain gauge stations corresponding to the γ types of additional means is K+1, K+2, ..., K+γ, respectively. Using a machine learning model, rainfall conditions for different additional methods are predicted. A feature vector X(s) is constructed for each rainfall observation station's location s, including the geographical location, topographic features, meteorological features, and radar rainfall data corresponding to the mesh where the observation point is located. Subsequently, the feature vectors and actual rainfall data for all K rainfall observation stations are used as the training dataset to train the machine learning model f. The feature vectors X of the rainfall observation stations are used as the model input, and the corresponding actual rainfall data is used as the model output. Subsequently, using the trained model, the feature vector X(s) for each additional candidate location s* is generated. ＊ Based on the rainfall conditions at that location, [Math 1] We predict the following: [Math 2] For all T time zones, rainfall data for each candidate additional position in γ types of additional means is predicted for each time zone, and spatial interpolation is performed on the prediction results for each additional means using the kriging interpolation method to generate corresponding rainfall field data. For T time zones, RE again for each time zone. ｊ >D ２ Calculate the number of meshes, [Math 3] It is written as, that is, the RE of the tth time period ｊ >D ２ Let the number of meshes be , t = 1, 2, ..., T, and then calculate the average percentage decrease in the number of meshes per for all time periods. [Math 4] Step 4 indicates that the larger the per, the closer the rainfall field data and radar rainfall data in the basin are after the addition. Step 5: Determining the optimal rainfall observation network configuration: Considering the construction costs, maintenance costs, and potential socioeconomic effects of rainfall observation stations, a penalty function ω is constructed to show the unfavorable impact of actual conditions on the construction of rainfall observation stations. Let C be the construction cost, W be the maintenance cost, and B be the socioeconomic effect. C, W, and B are all functions of additional measures and change with changes in the measures. Thus, the penalty function ω is defined as follows: ω=αC+βW−φB (5) In the formula, α, β, and φ are weighting coefficients. By combining per and ω, we construct an optimization target function φ, make it positively correlated with per and negatively correlated with ω, that is, f=m １ ×per-m ２ ×ω (6) In the formula: μ １ , μ ２ This is the weighting coefficient, Step 5 involves calculating the corresponding φ value for each type of additional means and selecting the means with the highest φ value as the optimal rainfall observation station arrangement means. A method for optimizing the arrangement of a mountainous rainfall observation network based on radar rainfall data, characterized by including the following:

2. Step 1 involves performing the same mesh division on the target area based on the mesh distribution of radar rainfall data, which specifically involves analyzing the radar rainfall data file, extracting its spatial resolution, determining the basic unit of the mesh, and then systematically numbering the mesh using a matrix coordinate method based on the spatial position of the radar data, with row numbers increasing from north to south and column numbers increasing from west to east, as described in claim 1, for optimizing the arrangement of a mountainous rainfall observation network based on radar rainfall data.

3. The method for optimizing the arrangement of a mountainous rainfall observation network based on radar rainfall data, characterized in that, in step 2, the node features of the rainfall observation stations are constructed by converting the geographical positional relationships between rainfall observation stations into a graph structure, each rainfall observation station is considered as one node in the graph structure, the edges between nodes indicate the spatial correlation between rainfall observation stations, the weights of the edges are set by a distance metric or a similarity metric based on geographical distance, the node features of the rainfall observation stations include rainfall time series features and geographic information features, the rainfall time series features include rainfall statistical features and temporal feature information for historical time periods, and the geographic information features include the latitude and longitude, elevation and topographic features of each rainfall observation station.

4. In step 2, identifying anomalous observation points at rainfall observation stations using a trained GCN model specifically involves calculating the reconstruction error RE predicted by the trained GCN model for all N time periods for each rainfall observation station. [Math 5] During the ceremony: [Math 6] These are the actual measured values ​​and predicted values ​​based on the GCN model at time n of the rainfall observation station, respectively. RE>D １ The rainfall observation station is marked as an anomaly observation point, and threshold D is set here. １ The method for optimizing the arrangement of a mountainous rainfall observation network based on radar rainfall data according to claim 1, characterized in that the location is determined in combination with local weather conditions and natural geographical features.

5. The method for optimizing the arrangement of a mountainous rainfall observation network based on radar rainfall data according to claim 1, characterized in that, in step 4, the topographic features include elevation, gradient, and gradient direction, and the meteorological features include temperature, humidity, and wind speed.

Images