Flood level prediction model management system

Abstract

A flood level prediction model management system according to one embodiment of the present invention may include a model input unit to which an LSTM model or a GRU model is input, a data input unit to which a meteorological dataset or a water level dataset is input, and a performance comparison unit that compares the performance of the model input to the model input unit according to the type of the dataset input to the data input unit.

Description

[Technical Field]

[0001] The present invention relates to a flood water level prediction model management system that can compare the performance of water level prediction models based on input data to derive a model that predicts water level, which is a core parameter of floods, and determine the optimal flood water level prediction model. [Background technology]

[0002] Global warming has led to abnormal weather, which is increasing the damage caused by natural disasters worldwide, and the frequency and intensity of natural disasters are expected to continue to increase. In particular, damage caused by floods is increasing, and if floods can be predicted, the economic and human losses caused by floods can be reduced.

[0003] In order to reduce damage caused by floods, it is necessary to accurately predict floods and evacuate people and property in flood-stricken areas at the appropriate time. However, flood forecasting poses many challenges because there are many variables to consider and each factor has spatial and temporal correlations.

[0004] Methods for flood forecasting mainly include hydrological models and data-based intelligent models.

[0005] Hydrological models are models that physically explain runoff confluences by analyzing hydrological characteristics. This method is based on fluid dynamics theory and combines physical laws such as the conservation of mass, momentum, and energy to derive confluence equations. However, these models require researchers' in-depth hydrological knowledge and are subject to variability over time due to factors such as topographical erosion, making them difficult to use over the long term. Furthermore, constructing models using large amounts of input data is difficult, and the nonlinear nature of the many variables that must be considered results in low prediction accuracy. Furthermore, because the characteristics of each river basin differ, individual models must be created for each river, resulting in a lack of versatility.

[0006] On the other hand, data-based intelligent models predict water levels and runoff volume through data analysis based on observed data. One data-based intelligent model applied to the field of hydrology for flood forecasting and management is the Artificial Neural Network (ANN) model. The ANN model uses water levels and meteorological data as input data and predicts water flow using Harmony Search (HS) and Differential Evolution (DE). Using HS and DE to update architecture parameters and select important features to prevent overfitting, the model has been shown to perform better than Radial Basis Function Neural Network (BRFNN) and Multilayer Perceptron (MLP) models, proving that ANN models can be used to predict water flow. Furthermore, numerous studies have been conducted to predict core flood elements, such as ANN-based water level and runoff volume forecasting models. However, most related research uses hydrological data and meteorological data, which are in the category of time series data, as input data, and in the case of ANN models, there is a problem of insufficient memory when calculating sequential data and time series data, and there is a problem of difficulty in finding optimal parameters during the learning process.

[0007] The present applicant has proposed the present invention to solve the above-mentioned problems.

[0008] Related prior art includes Korean Patent Publication No. 10-2020-0087347 (Name of invention: Method and device for recognizing rainfall amount from road surveillance camera images using TSN, Publication date: 2020.07.21.), Korean Registered Patent No. 10-2403270 (Name of invention: Water reservoir water level control system based on artificial intelligence, Registration date: 2022.05.24.), Korean Registered Patent No. 10-2409155 (Name of invention: (Name of invention: Groundwater level prediction system based on LSTM, Registration date: 2022.06.10.), Korean Registered Patent No. 10-2308526 (Name of invention: Flood prediction system, Registration date: 2021.09.28.), Korean Registered Patent No. 10-2159620 (Name of invention: Smart flood management hydrological control automation system based on AI and deep learning, Registration date: 2020.09.18.). [Prior art documents] [Patent documents]

[0009] [Patent Document 1] Korean Patent Publication No. 10-2020-0087347 [Patent Document 2] Korean Patent Registration No. 10-2403270 [Patent Document 3] Korean Patent Registration No. 10-2409155 [Patent Document 4] Korean Patent Registration No. 10-2308526 [Patent Document 5] Korean Patent Registration No. 10-2159620 Summary of the Invention [Problem to be solved by the invention]

[0010] The present invention has been proposed to solve the above-mentioned problems, and provides a flood water level prediction model management system that uses water level data and meteorological data as input data and can derive a water level prediction model based on LSTM-GRU with the best performance by comparing the performance of input models using input data. [Means for solving the problem]

[0011] To achieve the above-mentioned object, a flood level prediction model management system according to one embodiment of the present invention may include a model input unit to which an LSTM model or a GRU model is input, a data input unit to which a meteorological dataset or a water level dataset is input, and a performance comparison unit that compares the performance of the model input to the model input unit according to the type of the dataset input to the data input unit.

[0012] A multi-LSTM model consisting of two layers of LSTM, a multi-GRU model consisting of two layers of GRU, and an LSTM-GRU model composed of an LSTM and a GRU may be input to the model input unit.

[0013] The data input unit may receive input of a data set including water level data, a data set including water level data and AWS weather data, and a data set including water level data and ASOS weather data.

[0014] The performance comparison unit can conduct experiments or performance comparisons on models and input data consisting of a total of nine pieces of input data, which are a combination of three models input into the model input unit and three data sets input into the data input unit.

[0015] The performance comparison unit uses MSE as a basic loss function for learning the model input to the model input unit, and can compare observed values ​​with predicted values ​​using NSE and MAE indicators as auxiliary indicators for actual test data.

[0016] The performance comparison unit uses MSE, NSE, and MAE as performance comparison indices of the water level prediction model input to the model input unit, and can proceed with performance evaluation based on the water level prediction model input to the model input unit and meteorological data through the three indices.

[0017] The performance comparison unit can compare the MSE, NSE, and MAE indices with the error of the peak water level prediction to determine the performance of the model input to the model input unit.

[0018] The system includes a model determination unit that determines and presents the model with the best performance from the Multi LSTM model, the Multi GRU model, and the LSTM-GRU model according to the performance comparison result of the performance comparison unit, and the model determination unit can determine the LSTM-GRU model that uses ASOS weather data and water level data as learning data as the water level prediction model. [Effects of the Invention]

[0019] The flood water level prediction model management system of the present invention can more easily predict water levels using only meteorological data sets and water level data sets.

[0020] The flood water level prediction model management system according to the present invention can use actual observation data to present a model for data-based water level prediction.

[0021] The flood level prediction model management system of the present invention can check the performance differences between models based on input data and the performance differences based on the configuration of the model used to predict time series data, and can present the most suitable water level prediction model with the best performance. [Brief explanation of the drawings]

[0022] [Figure 1] 1 is a diagram illustrating the configuration of a flood water level prediction model management system according to an embodiment of the present invention. [Figure 2] FIG. 2 is a diagram for explaining a method for managing a flood water level prediction model by the system shown in FIG. 1. [Figure 3] 2 is a graph showing the amount of rainfall damage and the population of an exemplary area to which the system of FIG. 1 is applied. [Figure 4] 2 is a graph showing the amount of rainfall damage and the population of an exemplary area to which the system of FIG. 1 is applied. [Figure 5] 2 is a graph visualizing the entire data set upstream and downstream of the testbed area where the system according to FIG. 1 is applied. [Figure 6] 2 is a graph showing quality information of AWS and ASOS data input into the system according to FIG. 1; [Figure 7] FIG. 2 is a diagram showing the locations of water level measuring stations and meteorological data measuring stations in a testbed area where the system according to FIG. 1 is applied. [Figure 8] FIG. 2 illustrates the structure of LSTM and GRU models applied in the system according to FIG. 1. [Figure 9] FIG. 2 illustrates the structure of an LSTM-GRU model applied to the system according to FIG. 1. [Figure 10] 2 is a graph showing loss values ​​according to the number of training iterations (epochs) when training each model in the system shown in FIG. 1. [Figure 11] 2 is a graph showing the difference between the observed value and the predicted value according to the learning data for each model in the system according to FIG. 1 . [Figure 12] 2 is a graph showing the difference between the observed value and the predicted value according to the learning data for each model in the system according to FIG. 1 . [Figure 13] 2 is a graph showing the difference between the observed value and the predicted value according to the learning data for each model in the system according to FIG. 1 . DETAILED DESCRIPTION OF THE INVENTION

[0023] The advantages and / or features of the present invention and methods for achieving them will become apparent from the following detailed description of the embodiments in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, and may be embodied in various different forms. The present embodiments are provided solely to ensure that the disclosure of the present invention is complete and to fully convey the scope of the invention to those skilled in the art to which the present invention pertains. The present invention is defined solely by the scope of the claims. The same reference numerals refer to the same elements throughout the specification.

[0024] Furthermore, in order to efficiently describe the technical components constituting the present invention, the following preferred embodiments of the present invention will omit as much as possible functional components that are already included in the functional components of each system or that are commonly included in the technical field to which the present invention pertains, and will mainly describe functional components that should be added for the present invention. A person skilled in the art to which the present invention pertains should be able to easily understand the functions of components that have already been used in the past among the functional components omitted below, and should also be able to clearly understand the relationship between the omitted components and the components added for the present invention.

[0025] FIG. 1 is a diagram illustrating the configuration of a flood level prediction model management system according to one embodiment of the present invention. FIG. 2 is a diagram illustrating a method for managing a flood level prediction model using the system of FIG. 1. FIGS. 3 and 4 are graphs showing the amount of rainfall damage and the population in an exemplary area to which the system of FIG. 1 is applied. FIG. 5 is a graph visualizing the entire data set upstream and downstream of the testbed area to which the system of FIG. 1 is applied. FIG. 6 is a graph showing quality information of AWS and ASOS data input to the system of FIG. 1. FIG. 7 is a diagram showing the locations of water level measurement stations and meteorological data measurement stations in the testbed area to which the system of FIG. 1 is applied. FIG. 8 is a diagram showing the structure of LSTM and GRU models applied to the system of FIG. 1. FIG. 9 is a diagram showing the structure of an LSTM-GRU model applied to the system of FIG. 1. FIG. 10 is a graph showing loss values ​​according to Epoches when training each model in the system of FIG. 1. FIGS. 11 to 13 are graphs showing the difference between observed values ​​and predicted values ​​according to the learning data for each model in the system of FIG. 1.

[0026] Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

[0027] 1, a flood water level prediction model management system 100 (hereinafter referred to as the "model management system") according to one embodiment of the present invention may include a model input unit 120, a data input unit 140, a performance comparison unit 160, a model determination unit 180, and a database 190. Here, the model input unit 120, the data input unit 140, the performance comparison unit 160, and the model determination unit 180 may constitute a water level prediction model unit 110. That is, the model management system 100 according to one embodiment of the present invention may include the water level prediction model unit 110 and the database 190.

[0028] The model management system 100 according to an embodiment of the present invention can present the most suitable model for predicting water levels when a flood occurs.

[0029] One of the main causes of floods, to which the model management system 100 according to an embodiment of the present invention is applied, is torrential rain, which involves a large amount of rainfall in a short period of time. In the present invention, to select testbeds where significant damage was caused by rainfall, data on the current state of heavy rain damage by region published by the Ministry of the Interior and Safety of the Republic of Korea was used. The amount of rainfall damage and the frequency of rainfall damage occurrence by city in the Republic of Korea can be seen in Figure 3.

[0030] As shown in Figure 3, Gyeonggi Province experienced the second highest amount of rainfall damage among all cities, and the highest frequency of rainfall among all cities. Furthermore, due to its geographical location, Gyeonggi Province is located in the heart of South Korea's capital city, which has the highest population density, and a large population is distributed within the province. Figure 4 shows the population density and population of each city as of 2020.

[0031] As shown in Figures 3 and 4, the model management system 100 according to one embodiment of the present invention selects an area (e.g., Gyeonggi Province) that is expected to suffer the most damage when heavy rain damage and floods occur as a test bed area, and selects Yeoju Weir in Yeoju City, Gyeonggi Province, which measures upstream and downstream data within the Gyeonggi Province area, as an example of the test bed.

[0032] A meteorological data set and a water level data set may be input to the data input unit 140 of the model management system 100 according to an embodiment of the present invention.

[0033] The water level dataset was based on the Korea Water Resources Management Integrated Information System, and data from the water level gauging stations upstream and downstream of the Yeoju Weir in Yeoju City, Gyeonggi Province was used. Due to the nature of rivers, water flows from upstream to downstream, and excessive runoff caused by rainfall can lead to flooding. A water level dataset and a meteorological dataset were used as input data for the data input unit 140 of a model management system according to an embodiment of the present invention. Data was collected every hour from October 2, 2013, to November 12, 2020. The total number of data rows is 71,136 rows for both the upstream and downstream sections. The entire upstream and downstream datasets can be visualized and confirmed using Figure 5. Figure 5(a) shows the dataset upstream of the Yeoju Weir, and Figure 5(b) shows the dataset downstream of the Yeoju Weir.

[0034] The upstream and downstream data sets have the same characteristic of maintaining appropriate water levels except during the summer when there is heavy precipitation, and rapidly increasing water levels during the summer when there is heavy precipitation. As shown in Figure 5, in the case of the upstream of the Yeoju Weir, the increase in water level when rainfall occurs is significantly greater than that downstream compared to when there is not much rainfall.

[0035] In a model management system 100 according to an embodiment of the present invention, two sets of weather data provided by the Korea Meteorological Administration (KMA) were used as training data for the water level prediction model. These data were obtained from the Automatic Weather System (AWS) and the Automated Synoptic Observing System (ASOS), which observe ground-based weather. The AWS is a piece of equipment designed to automatically observe what humans previously observed. It automatically processes all processes, including real-time measurement, calculation, storage, and display, and observes data such as atmospheric pressure, temperature, humidity, wind direction, wind speed, and precipitation in real time. Both the ASOS and AWS datasets use temperature, humidity, and precipitation parameters as input data. Similar to the water level dataset, data was measured every hour.

[0036] Figure 6 is a graph showing the quality information of AWS and ASOS data provided by the Korea Meteorological Administration. Korea has a hot and humid climate in the summer, with heavy rainfall and typhoons. Figure 6 shows that the accuracy of the unmanned AWS data during that period is inferior to other periods and ASOS.

[0037] Meanwhile, the location of the nearest AWS observation station is about 8 km away from the actual testbed, Yeoju Weir, and the location of the nearest ASOS observation station is about 20 km away from Yeoju Weir. Figure 7 shows the locations of the water level measurement stations and meteorological data measurement stations in the testbed area. Figure 7(a) shows the locations of the water level measurement stations and meteorological data measurement stations at Yeoju Weir, and Figure 7(b) shows the location of Gyeonggi Province where Yeoju Weir is located.

[0038] In the case of summer torrential rains, the rainfall area is narrow and heavy rainfall continues. This characteristic can affect the performance of the water level prediction model depending on the distance between the actual test bed and the meteorological data observation point. Therefore, the model management system 100 according to an embodiment of the present invention conducts a comparative experiment of the performance of models when two actual data are used as input data, and derives an optimal water level prediction model. Table 1 shows information on the hydrological and meteorological observation stations used and the data sets used.

[0039] [Table 1]

[0040] Referring to Table 1, "Yeojubo upstream" and "Yeojubo downstream" refer to the upstream and downstream of Yeoju and are hydrological observation stations, while "ASS Yeoju" and "AOSO Icheon" refer to meteorological observation stations. "Latitude" and "Longitude" refer to the longitude and latitude of the hydrological and meteorological observation stations, respectively. Hydrological observation stations measure water levels, while meteorological observation stations measure temperature, humidity, and precipitation.

[0041] The data set shown in Table 1 may be input to the data input unit 140 of the model management system 100 according to an embodiment of the present invention.

[0042] The model management system 100 according to one embodiment of the present invention inputs various types of data, such as a water level data set and a meteorological data set, to derive a water level prediction model. It is possible to derive an optimal water level prediction model not only by comparing the performance of one model for various input data, but also by comparing the performance of various models for various types of input data.

[0043] Therefore, various neural network models for processing time-series data may be selected and input to the model input unit 120 of the model management system 100 according to one embodiment of the present invention.

[0044] First, a Long Short-Term Memory (LSTM) model may be input to the model input unit 120. The LSTM model is one of the recurrent neural network (RNN) architectures used for time-series data processing. RNNs are mainly used for data with temporal correlation, and have a signal circulation structure to predict future data by considering the correlation between previous data and current data. RNNs predict future data based on past data, but have the problem of not being able to store past data for a long time. LSTM is an architecture that emerged to address this problem. It has a total of six parameters and a structure consisting of four gates, which can address not only short-term memory but also long-term memory. The LSTM structure is as shown in FIG. 8(a).

[0045] LSTM networks have a chain structure similar to RNNs, and the recursive modules of RNNs communicate with each other through four layers, rather than just one tanh layer. The state within an LSTM cell is broadly divided into two vectors, where ht represents the short-term state and ct represents the long-term state. Data can be added or removed from the cell state through sigmoid gates, each of which resembles a layer with different individual weights or a series of matrix operations. Since the gates can also retain information from long-ago data, they are designed to solve long-term dependency problems.

[0046] The first step in an LSTM network is to identify and determine unnecessary information to omit from a cell. The relevant cell is a forget gate, and the process involves determining the output of the last LSTM cell (ht-1) at time t-1 and the current input (xt) at current time t using a sigmoid function. The value output by the sigmoid function ranges from 0 to 1, with larger values ​​indicating complete memory retention of previous state information and smaller values ​​indicating more forgetfulness of previous state information, determining the portion of the previous output to omit.

[0047] After passing through the forget gate, the process of selecting the information to be stored begins. Through the forget gate, the previous memory cell (ct-1) is forgotten, new information is added, and the value of each element as newly added information is determined. This time, new information is not accepted unconditionally, but is appropriately selected. The gate that performs this role is called the input gate. A sigmoid function is applied to the last LSTM cell (ht-1) and the current value (xt), and a tanh function, which is an activation function, is added. The value after passing through the sigmoid layer is between 0 and 1, indicating the degree to which the new information is updated. The value after passing through the tanh function is between -1 and 1, indicating the importance of the weighting. The input gate finally performs a Hadamard product operation on the two values, adding the corresponding new memory to the previous cell state (ct-1) to become ct.

[0048] After determining the value of new information through the input gate and selecting the information to store, the next step is to select the output information. The corresponding gate is set as the output gate, and a sigmoid function is taken through the current value (xt) and the value of the last LSTM cell (ht-1), and the Hadamard product is calculated with the current cell state (ct), resulting in a filtered value, which becomes the hidden state.

[0049]

number

[0050] In this case, σ represents the sigmoid function, W represents the weight matrix, and b represents the bias. ct represents the cell state at the current time, and ct-1 represents the cell state at the previous time. And represents the harmonized product operation. Modification (1) represents the process of going through the forget gate, and the process of going through the input gate involves updating the cell state through modifications (2) and (3) and (4). Next, through the output gate represented by modification (5), the state of the final hidden layer is updated through modification (6), and this is how LSTM operates.

[0051] In addition, a gated recurrent unit (GRU; hereinafter, referred to as "GRU") model may be input to the model input unit 120. The GRU is one of the RNN architectures, and is a model that reduces the calculations for updating hidden states while maintaining the solution to the long-term dependency problem of the LSTM, which has improved the problems of the RNN. The LSTM requires more parameters than the existing RNN to solve the long-term dependency problem, and there is a problem of overfitting when there is insufficient data. The GRU can improve these shortcomings by modifying the structure of the LSTM. The structure of the GRU is shown in FIG. 8(b).

[0052] Referring to Figure 8(b), we can see that the structure is definitely simpler than the LSTM structure shown in Figure 8(a). The main difference between GRU and LSTM is that GRU combines the LSTM's forget gate and input gate and replaces them with an update gate. In addition, GRU has a simpler structure than LSTM by combining the cell state and hidden state, and has fewer parameters than LSTM, resulting in lower computational costs.

[0053] The process corresponding to r in Figure 8(b) represents the reset gate, which represents the hidden state of the network through that process. The result of passing through the reset gate is calculated with past hidden layer information to calculate a set of hidden state candidates. At this time, the past hidden state value is multiplied by the value passing through the reset gate to determine the hidden state candidate. Next, the part corresponding to z represents the update gate. This part functions as the LSTM's forget gate and input gate, determining how much of the current information to use. The value calculated through that gate is calculated with the previously calculated hidden state candidate to determine the final hidden state. The following modifier (7) corresponds to the reset gate, modifier (8) corresponds to the update gate, and modifiers (9) and (10) determine the hidden state candidate and final hidden state.

[0054]

number

[0055] Since it is not possible to determine which of GRU and LSTM is better in terms of model performance, the model management system 100 according to one embodiment of the present invention performs experiments based on LSTM and GRU in the performance comparison unit 160 to compare their performance. On the other hand, the performance comparison unit 160 of the model management system 100 according to one embodiment of the present invention uses three performance comparison indices for water level prediction models: MSE (Mean Squared error), NSE (Nash-Sutcliffe coefficient of efficiency), and MAE (Mean Absolute Error).

[0056] MSE is a method used to evaluate the performance of a regression model, and is calculated by squaring the difference between the actual observed value and the predicted value. Since this index is calculated by squaring the difference between the observed value and the predicted value, it is sensitive to abnormal values. In the case of hydrological models, if abnormal values ​​occur compared to the prediction, it may cause loss of life, so the MSE index was selected.

[0057] In the case of NSE, it is an index often used to evaluate the performance of hydrological models and has a value of (-∞ to 1). The closer the value is to 1, the better the model performance.

[0058] In the case of MAE, it means the average of all absolute errors between observed and predicted values, and has the advantage of being able to intuitively check the performance of the model. The equations for the performance comparison index are given as modified (11) to modified (13).

[0059]

number

[0060] The performance comparison unit 160 of the model management system 100 according to an embodiment of the present invention can compare performance by conducting experiments for each model using input data using the LSTM model and the GRU model through the above three indicators.

[0061] The models used in the experiment may include a multi-LSTM model consisting of two layers of LSTM, a multi-GRU model consisting of two layers of GRU, and an LSTM-GRU model consisting of an LSTM and a GRU. Therefore, a multi-LSTM model, a multi-GRU model, and an LSTM-GRU model may be input to the model input unit 120 of the model management system 100 according to one embodiment of the present invention.

[0062] The datasets used for model training input to the model input unit 120 may include a dataset S1 consisting of only water level data, a dataset S2 consisting of water level data and AWS weather data, and a dataset S3 consisting of water level data and ASOS weather data. Such datasets S1, S2, and S3 may be input to the data input unit 140.

[0063] The performance comparison unit 160 of the model management system 100 according to an embodiment of the present invention performs a model-by-model comparison using three data sets, resulting in a performance comparison for a total of nine models. Information about each experiment can be found in Table 2.

[0064] [Table 2]

[0065] In Table 2, S1 indicates a scenario using the water level dataset as training data, S2 indicates a scenario using the water level dataset and the AWS weather dataset as training data, and finally, S3 indicates a scenario using the water level dataset and the ASOS dataset as training data. Therefore, the performance comparison unit 160 conducts a performance comparison experiment using nine scenarios.

[0066] There are three water level prediction model structures used in the model management system 100 according to one embodiment of the present invention, and the overall model configuration is as shown in FIG.

[0067] Referring to Figure 9, all three models use the past 20 hours of data as input data. When learning using learning data corresponding to scenario S1, the input data format is [None, 2], and for S2 and S3, it is [None, 5]. Finally, after passing through a dense layer, the 21st hour data is predicted using the past 20 hours of data.

[0068] In the case of the Multi LSTM model, both hidden layers are composed of LSTM layers, and in the case of the Multi GRU model, both layers are composed of GRUs. In the case of the LSTM-GRU model, the first hidden layer is composed of an LSTM layer and the second hidden layer is composed of a GRU layer.

[0069] In one embodiment of the model management system 100, a water level dataset and a meteorological dataset are input to the data input unit 140. The performance comparison unit 160 conducts experiments using nine sets of input data, combining three models and three data scenarios (S1, S2, and S3). The entire training and test data collection period was measured every hour from October 2, 2013 to November 12, 2021. Of the 71,136 rows of data, 56,908 rows (80%) were used as training data and validation data, and the remaining 14,208 rows were used as testing data. For flood forecasting models, predicting sudden increases and high water levels is important. The performance comparison unit 160 can verify the validity of the model by determining whether the maximum value of all measured water levels is present in the test data.

[0070] The performance comparison unit 160 uses MSE as the basic loss function for model training, and NSE and MAE indicators are used as auxiliary indicators for comparing observed and predicted values ​​for actual test data. The optimization function also uses adam, and the number of units for the LSTM model and GRU model is 256, with the number of training iterations (epochs) being 200, all under the same conditions. For the learning model, a total of nine cases are used, as summarized in [Table 2], and information on the training and validation of each model is shown in [Table 3].

[0071] [Table 3]

[0072] Figure 10 shows the loss graphs according to the number of training iterations while learning each model with each scenario S1, S2, and S3.

[0073] 10, we can see how the loss value changes depending on the number of training iterations (epochs) during model training. In FIG. 10, the horizontal axis of each graph represents the number of training iterations (epochs), and the vertical axis represents the loss value (Loss).

[0074] In addition, blue indicates training data and orange indicates test data in the graphs (a) and (b) of Figure 10. Blue indicates training data and orange indicates validation data in the graphs (c) to (i) of Figure 10.

[0075] According to the performance comparison unit 160, it was confirmed that the Multi-LSTM model had the largest number of hyperparameters and the Multi-GRU model had the smallest number of hyperparameters, depending on the characteristics of the LSTM model. Regarding the learning time, the Multi-GRU model required the least amount of time, but it was also confirmed that the learning time was not proportional to the hyperparameters. Furthermore, referring to FIG. 10, in all cases, the loss value converged toward 0 as the number of training iterations (epochs) increased, indicating that learning was performed well. In the case of the S3_LSTM_GRU model, shown in FIG. 10(h), the loss value increased sharply in the middle, which was determined to be due to overfitting. Therefore, to ensure fairness with other models, the number of training iterations (epochs) was maintained at 200 during the experiment.

[0076] The performance comparison unit 160 found differences in the MSE values ​​used as the loss function depending on the input data characteristics for each model. However, the Multi-LSTM model did not show significant differences depending on the input data. It was also confirmed that the Multi-GRU and LSTM-GRU models exhibited performance differences depending on the input data. For the Multi-GRU model, performance was found to be lower in the S2 and S3 cases, which included weather data in the training data, compared to the S1 case, demonstrating poor performance among the nine cases. This is believed to be due to the lightweight LSTM, which was unable to effectively learn high-dimensional data compared to the LSTM model. However, for the LSTM-GRU model, it was confirmed that the S1 model, which had a small training data dimension, exhibited the lowest verification MSE. Finally, the best-performing case was the S3_LSTM_GRU, which achieved the best results of 0.15 during training and 0.20 during verification.

[0077] The performance comparison unit 160 determines the performance of the model by comparing the highest water level data from the entire data collection period in the test data set with the three indices MSE, NSE, and MAE and the maximum water level prediction error. At this time, the highest water level among the observed data is 3552 cm. The test results can be confirmed through [Table 4], and as shown in Figures 11 to 13, the difference between the observed value and the predicted value according to the learning data for each model can be confirmed through a graph.

[0078] Figure 11 shows a graph comparing the observed values ​​(orange) and predicted values ​​(blue) for the Multi LSTM model (a), the Multi GRU model (b), and the LSTM-GRU model (c) in scenario S1.

[0079] Figure 12 shows a graph comparing the observed values ​​(orange) and predicted values ​​(blue) for the Multi LSTM model (a), the Multi GRU model (b), and the LSTM-GRU model (c) in scenario S2.

[0080] Figure 13 shows a graph comparing the observed and predicted values ​​for the Multi LSTM model (a), the Multi GRU model (b), and the LSTM-GRU model (c) in scenario S3.

[0081] [Table 4]

[0082] The test results of the performance comparison unit 160 were similar to those of the model training and validation. No significant differences were found depending on the input data for the Multi LSTM model, and the NSE value was best when data corresponding to case S1 of the three data sets was used as input data. The maximum water level prediction error for the validation data was also best in case S1, at 81.77 cm. When case S2 data was used for the LSTM model, the NSE was 0.802, slightly lower than in cases S1 and S3, and the maximum water level prediction error was also the lowest at 106.22 cm.

[0083] The Multi-GRU model showed the poorest performance compared to other models in all cases. Similar to the results during validation, this is believed to be due to the GRU model being unable to learn effectively when the input data has many dimensions. When using data from case S1, the GRU model achieved the best results, but its performance was inferior to that of the other models. For cases S2 and S3, the NSE values ​​were 0.31 and 0.356, respectively, the lowest of the nine cases. The maximum water level prediction errors were also the highest, at 150.14 for S2 and 207.17 for S3.

[0084] The LSTM_GRU model was found to have the best average performance of the three models, but in the S1 case, it showed a slightly lower NSE and slightly higher maximum water level prediction error than the LSTM model. However, in the S2 and S3 cases, the NSE values ​​were 0.935 and 0.42, respectively, and the maximum water level errors were 98.06 and 47.16, showing significantly better performance than the other models. Of these, the S3 case achieved the best results out of the nine models in all evaluation indicators, including MSE, NSE, MAE, and maximum water level prediction error.

[0085] Therefore, using the above experimental results, the model determination unit 180 of the model management system 100 according to one embodiment of the present invention can determine that the most suitable weather data for the test bed, Yeoju Dam, is the ASOS data corresponding to the S3 case in terms of model performance.

[0086] From the experimental results of the S3 case, the model determination unit 180 determines the LSTM-GRU model that uses ASOS weather data and water level data as learning data as the most suitable water level prediction model.

[0087] Meanwhile, all data involved in the performance comparison experiment of the three models for the three scenarios S1, S2, and S3 using the dataset may be stored in the database 190.

[0088] As shown in FIG. 2, a method for managing a flood level prediction model using a flood level prediction model management system 100 according to one embodiment of the present invention may include the steps of inputting a model to be tested into a model input unit 120 (S110), inputting a dataset into a data input unit 140 (S120), conducting experiments on dataset scenarios for each model via a performance comparison unit 160 to compare performance (S130), and determining and presenting a model with the best performance via a model determination unit 180 (S140).

[0089] The method for managing a flood water level prediction model according to an embodiment of the present invention may be performed by the flood water level prediction model management system 100 described above.

[0090] The flood water level prediction model management system 100 according to one embodiment of the present invention as described above can more easily predict water levels using only meteorological data sets and water level data sets.

[0091] The systems (apparatuses) described above may be implemented using hardware components, software components, and / or a combination of hardware and software components. For example, the devices and components described in the embodiments may be implemented using one or more general-purpose or special-purpose computers, such as a processor, controller, arithmetic logic unit (ALU), digital signal processor, microcomputer, field programmable array (FPA), programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. The processing device may also access, store, manipulate, process, and generate data in response to the execution of software. For ease of understanding, the processing device may be described as being a single device; however, those skilled in the art will recognize that the processing device may include multiple processing elements and / or multiple types of processing elements. For example, the processing device may include multiple processors or one processor and one controller. Other processing configurations are also possible, such as parallel processors.

[0092] Software may include computer programs, code, instructions, or a combination of one or more of these, capable of configuring or independently or collectively instructing a processing device to operate in a desired manner. The software and / or data may be permanently or temporarily embodied in some type of machine, component, physical device, virtual device, computer storage medium or device, or transmitted signal wave to be interpreted by or provide instructions or data to a processing device. The software may be distributed across network-coupled computer systems, stored and executed in a distributed manner. The software and data may be stored on one or more computer-readable storage media.

[0093] Methods according to embodiments may be embodied in the form of program instructions that can be executed by various computer means and stored on a computer-readable medium. The computer-readable medium may include, alone or in combination, program instructions, data files, data structures, and the like. The program instructions stored on the medium may be specially designed and constructed for the embodiments, or may be known and available to those skilled in the art of computer software. Examples of computer-readable storage media include magnetic media such as hard disks, floppy disks, and magnetic tape; optical media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include not only machine code, such as produced by a compiler, but also high-level language code that can be executed by a computer using an interpreter, for example. The hardware devices described above may be configured to operate as one or more software modules to perform the operations of the embodiments, or vice versa.

[0094] As described above, the embodiments of the present invention have been described using specific details such as specific components, limited embodiments, and drawings, but these are provided merely to facilitate a more comprehensive understanding of the present invention, and the present invention is not limited to the above embodiments. Those skilled in the art will appreciate that various modifications and variations can be made from such descriptions. Therefore, the scope of the present invention should not be limited to the described embodiments, and all modifications equivalent to or equivalent to the scope of the claims, as well as the following claims, are considered to fall within the scope of the present invention.

Claims

1. a model input unit to which an LSTM model or a GRU model is input; a data input section into which a meteorological data set or a water level data set is input; a performance comparison unit that compares the performance of the model input to the model input unit according to the type of data set input to the data input unit.

2. The model input unit includes:

2. The flood level forecasting model management system according to claim 1, wherein a Multi LSTM model consisting of two layers of LSTM, a Multi GRU model consisting of two layers of GRU, and an LSTM-GRU model consisting of LSTM and GRU are input.

3. The data input unit includes: The flood water level prediction model management system according to claim 2, characterized in that a dataset including water level data, a dataset including water level data and AWS meteorological data, and a dataset including water level data and ASOS meteorological data are input.

4. The performance comparison unit 4. The flood water level prediction model management system according to claim 3, characterized in that experiments or performance comparisons are conducted on the models and input data consisting of a total of nine sets of input data, which are a combination of three models input into the model input section and three data sets input into the data input section.

5. The performance comparison unit 5. The flood level forecasting model management system according to claim 4, wherein MSE is used as a basic loss function for learning the model input to the model input unit, and NSE and MAE indices are used as auxiliary indices for actual test data to compare observed values ​​with predicted values.

6. The performance comparison unit 5. The flood water level prediction model management system according to claim 4, characterized in that MSE, NSE and MAE are used as performance comparison indices for the water level prediction model input to the model input unit, and performance evaluation is carried out using the three indices based on the water level prediction model input to the model input unit and meteorological data.

7. The performance comparison unit 6. The flood water level prediction model management system according to claim 5, characterized in that the performance of the model input to the model input unit is determined by comparing the MSE, NSE and MAE indices with the error of the maximum water level prediction.

8. a model determination unit that determines and presents a model having the best performance among a Multi LSTM model, a Multi GRU model, and an LSTM-GRU model according to a performance comparison result of the performance comparison unit; 8. The flood water level prediction model management system according to claim 7, wherein the model determination unit determines the LSTM-GRU model, which uses ASOS meteorological data and water level data as learning data, as the water level prediction model.

Images